Patent Docs

Andersen v. Stability AI: Defendants' Motion to Dismiss Narrows the Case, But Only Slightly

By Joshua Rich --

In the lawsuit brought against them for using visual artists' work to teach their large language model, and producing near-identical copies in response to prompts, Stability AI, Midjourney, DeviantArt, and Runway AI moved to dismiss almost all of the claims asserted against them.  Those claims include copyright infringement, violations of the Digital Millenium Copyright Act ("DMCA"), unjust enrichment, violation of the Lanham Act for false endorsement and trade dress infringement, and, against DeviantArt alone, breach of contract.  Judge Orrick of the U.S. District Court for the Northern District of California, although seemingly skeptical of the merits of some of the other claims, dismissed only the DMCA and breach of contract claims with prejudice and the unjust enrichment claims without prejudice and with leave to amend.  Thus, the sprawling, heavily litigated case will go forward based on many different theories for recovery (including quite a few novel ones).

The case began when visual artists Sarah Andersen, Kelly McKernan, and Karla Ortiz[1] sued Stability AI, Midjourney, and DeviantArt on various claims arising out of the creation and use of the Stable Diffusion large language model.  Allegedly, Stability AI developed Stable Diffusion and trained it through the use of LAION ("Large-Scale Artificial Intelligence Open Network") datasets that include the plaintiffs' artwork.  Each of the defendants has then allegedly employed Stable Diffusion as the engine for the AI engine for its proprietary artwork-producing product, including in creating images "in the style" of the plaintiffs' art (some of which are nearly identical copies of existing works).  The plaintiffs asserted claims against all three defendants for copyright infringement (both direct and inducement), violations of the DMCA, violation of the right of publicity, unfair competition, and declaratory relief, as well as breach of contract claims against DeviantArt.

All three of the initial defendants separately filed motions to dismiss the original Complaint, but raised similar arguments.  As it turned out, Kelly McKernan and Karla Ortiz had not registered any of their copyrights, so they had to concede that they could not assert such claims.  The court also narrowed Sarah Anderson's copyright claims to those that she had registered, but allowed those to go forward.  However, the court had ruled that the factual pleadings for inducement of copyright infringement were too conclusory to support the claim, so dismissed those with leave to file an amended complaint with more factual basis.  The same was true for the claims based on the DMCA, right to publicity, unfair competition, and breach of contract; Judge Orrick explained at length how the original Complaint had too little factual explanation for a defendant to identify whether there was a plausible cause of action stated in any of the claims, but allowed the plaintiffs to replead their allegations with more factual support so as to identify such claims.

Rather than merely adding more factual averments in support of the claims they already pled, the plaintiffs fundamentally changed the pleadings.  They added seven more named plaintiffs and a new defendant, Runway AI, who they accuse of developing Stable Diffusion with Stability AI.  They also changed the asserted causes of action, dropping some (right of publicity and declaratory relief) and adding others (violation of the Lanham Act).  The ten named plaintiffs then averred additional facts in support of the claims, albeit not in all of the ways Judge Orrick identified the claims to be deficient.

Once again, each of the defendants moved to dismiss the claims asserted against them.[2]  While most of the arguments overlapped, there were some differences based on the claims asserted against each defendant and their factual circumstances.  Ultimately, based on the defendants' motions to dismiss, Judge Orrick dismissed the DMCA and breach of contract claims with prejudice and the unjust enrichment claims without prejudice and with leave to amend.

The first argument raised by Stability AI and DeviantArt was that adding more plaintiffs and new claims was not what Judge Orrick gave leave to do.  Judge Orrick acknowledged that to be true and, while leave to amend is freely given, noted that the proper course of action would have been to request leave to add the new plaintiffs and claims.  But after rapping the plaintiffs' knuckles for not doing so, he found that he would have granted leave for those changes had he been asked and allowed the new plaintiffs and claims to proceed.

Substantively, Midjourney raised an argument that bridged all of the copyright claims -- that certain plaintiffs' copyright registrations were insufficient to support the claims of infringement.  Namely, two plaintiffs (Sarah Anderson and Julia Kaye) registered some their works as part of compilations and plaintiff Gerald Brom registered some of his works as text rather than artwork.  But every plaintiff had at least one visual work (that is, a work of visual art) registered and asserted.  Midjourney also argued that the plaintiffs had not identified all of the copyright registrations that were being asserted.  But the Federal system requires only notice of plausible claims, not identification of all supporting facts.  Therefore, Midjourney's arguments were not a basis for dismissing any claim (although any copyright not properly registered cannot be the basis for a claim of infringement).

Only one party, DeviantArt, argued that the First Amended Complaint completely failed to make out a claim for direct copyright infringement against it.  It did not train the Stable Diffusion model and the plaintiffs had failed to identify facts sufficient to tie it to any infringement in the original Complaint.  However, to do so it required the court to rely on a review of academic articles cited but not incorporated in the First Amended Complaint to determine the plausibility of the plaintiffs' allegations.  That is simply asking too much of the court to demand a deep dive into the technology, far beyond the face of the First Amended Complaint, to determine the veracity of the allegations.  Similarly, DeviantArt asked the court to find that use of the plaintiffs' works in Stable Diffusion was fair use.  Those fact-based arguments are suited for a motion to dismiss, not a motion for summary judgment.

Although it did not contest that the plaintiffs had stated a claim for direct infringement through teaching the Stable Diffusion tool, Runway AI argued that certain theories of direct infringement related to Stable Diffusion 1.5 (namely that the model itself, after training, was an infringing copy of plaintiffs' works or that distributing Stable Diffusion 1.5 violated the plaintiffs' distribution rights) failed to state a claim.  Since there was no dispute that at least one theory of direct infringement (that training Stable Diffusion on the plaintiffs' works infringed their copyrights) stated a claim, the court had no need to address other theories.  However, it noted that proof of the theories was based on what facts could be proven, which is improper to resolve on a motion to dismiss.

All of the defendants argued that the claims for inducing copyright infringement were deficient.  Two theories of induced infringement were advanced.  First, as alleged against Stability AI and Runway AI, the Stable Diffusion models themselves were infringing works and their distribution (such as to Midjourney and DeviantArt) constitutes infringement.  The defendants argued that was just a direct infringement claim repackaged under a different theory.  Judge Orrick found, however, that the plaintiffs were entitled to plead the two forms of infringement in the alternative and to determine how Stable Diffusion works and is implemented by users.  Any potential overlap can be resolved later, after discovery.

Second, the defendants argued that the plaintiffs failed to aver facts that would support a claim that that are encouraging others to use Stable Diffusion to create infringing outputs.  The plaintiffs had identified a statement by Stability AI's CEO indicating that Stable Diffusion could "recreate" any of the images on which it had been taught,[3] as well as articles by academics and others identifying the fact that training images could sometimes be reproduced as outputs.  Those facts took the case out of the VCR paradigm of Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417 (1984), under which the marketing of a product that could be easily used to infringe copyrights -- but also capable of substantial noninfringing use -- would not create a presumed intent to cause infringement.  That is, here, there was actual evidence that the marketing of Stable Diffusion was being done with the knowledge that it would likely facilitate infringement by others.  Thus, Judge Orrick found that the plaintiffs had averred facts sufficient to go forward with a claim for inducement of copyright infringement.

Under the DMCA, the plaintiffs brought two different claims.  Under § 1202(a), they asserted that the defendants were falsely alleging that they owned the copyrights in the plaintiffs' copyrighted works by claiming copyright in the Stability Diffusion large language model.  The defendants countered by arguing that a claim of copyright in the Stable Diffusion large language model was not a claim of copyright "in connection with" any works produced by the large language model and that there was no allegation that they knowingly provided false copyright management information ("CMI") with the intent to induce copyright infringement.  Under § 1202(b)(1), they argued the defendants intentionally removed or altered CMI.  Judge Orrick found neither claim plausible.  He found a viewer would not read the copyright license governing the Stable Diffusion model to necessarily apply to works produced by the model.  And, consistent with earlier precedent in the same district,[4] he found none of the defendants "removed" or "altered" any CMI, they just did not affix CMI to AI-generated works nearly identical to existing works.  Because the newly-generated works are not truly identical to the works used to train the model, he found § 1202(b)(1) did not require ensuring the CMI was on the works.  Thus, Judge Orrick dismissed all of the DMCA claims with prejudice.

The defendants next moved to dismiss the plaintiffs' unjust enrichment because, they argued, the claims are preempted by the Copyright Act.  To survive preemption, a state law cause of action (like unjust enrichment) must have some additional element that makes the protected rights qualitatively different from copyright rights.  As the claims were pled, there was little dispute that the unjust enrichment claims did not include any extra elements beyond copyright infringement.  Instead, the plaintiffs argued in briefing that the defendants were profiting off of the plaintiffs' reputations by mimicking their works based on prompts using the artists' names.  While that theory might not be preempted, because it did not appear in the First Amended Complaint the parties did not have a fair opportunity to address it.  Accordingly, Judge Orrick dismissed the unjust enrichment claim with leave to replead in a Second Amended Complaint, if the plaintiffs chose to do so.

Midjourney argued that the Lanham Act claim asserted by five of the plaintiffs, claiming that Midjourney falsely claimed that the artists had endorsed its product by Midjourney's CEO including them on a list of artists on the Discord platform that its tool could mimic and the company itself included user-created works incorporating the artists' names in its showcase.  Midjourney argued that the plaintiffs failed to show falsity, relying on portions of a Discord thread not relied upon by the plaintiffs and requesting judicial notice of the evidence.  That type of disputable evidence is exactly what judicial notice is not intended to permit.  But more fundamentally, the fact that there is disputed evidence shows that dismissal is not appropriate; that argument is better presented at the summary judgment stage.  Midjourney makes two other arguments that suffer the same deficiencies: that invocation of the plaintiffs' names to identify their styles have no artistic relevance to the underlying works and that it cannot be liable for vicarious trade dress infringement because the plaintiffs have not identified all of the hallmarks of their works.  Again, these are factual arguments not resolvable as a matter of law, the standard at the motion to dismiss phase.

DeviantArt also moved to dismiss the breach of contract claim asserted against it.  Essentially, the plaintiffs argued that the contractual provision between them indicating that it did not claim copyright in any of the artists' works was breached when DeviantArt incorporated Stable Diffusion into its own AI tool.  Just as he had done before, however, Judge Orrick rejected the argument that DeviantArt had breached that provision, even if third parties (whether Stability AI and Runway AI in teaching Stable Diffusion or end users in using DeviantArt's tool) were potentially infringing.  DeviantArt itself had not exceeded the scope of its limited license by them doing so.  The plaintiffs further argued that DeviantArt had breached an implied covenant of good faith and fair dealing, but could not tie it to a specific contractual provision that was frustrated by DeviantArt's conduct.  The breach of contract claims were therefore dismissed with prejudice.

The motions to dismiss were notable for being supported by requests from Runway AI and Midjourney (and DeviantArt, albeit informally) for the court to take judicial notice of briefing from other cases and academic articles mentioned in the First Amended Complaint.  Judicial notice of pleadings is proper only to prove they exist and were filed, not to incorporate the arguments made in them.  But that is exactly what the defendants were trying to do -- supplement their arguments with those made in the briefing and articles.  Judge Orrick therefore refused to take judicial notice of the requested documents.

All in all, the proceedings on the motions to dismiss reveal quite a bit about the case.  First, current statutory law and precedent is poorly fitted for the resolution of disputes over ownership of AI-generated non-textual creative works.  The plaintiffs here have struggled to identify the appropriate claims to assert and the defendants have struggled to find good defenses to undercut them.  Second, both plaintiffs and defendants will be throwing every argument they can at the other side, regardless of the strength of the argument, hoping the court will leverage it in their favor.  Finally, Judge Orrick has made it clear that the critical inflection point in the case will be summary judgment, at which point it should be clear to everyone how he will rule on the merits.

[1] The lawsuit is brought as a putative class action, but there were only three named plaintiffs in the original Complaint.  There are six different classes identified in the First Amended Complaint based on relief sought (injunctive relief or damages) and which dataset class members' works are found in.  Given that class certification requires questions or law or fact common to the class(es) and the representative parties having claims typical of the class(es), the number of different subclasses is a bad omen for class certification.  See Fed. R. Civ. P. 23(a).

[2] Other than DeviantArt, none of the defendants moved to dismiss the direct copyright infringement claims.

[3] Runway AI tried argue that Stability AI’s CEO’s remarks pointed a finger only at Stability AI, but since the two worked together on Stable Diffusion and there was other evidence of intent, Judge Orrick rejected that argument.

[4] Doe 1 v. GitHub, Inc., No. 22-CV-06823-JST, 2024 WL 235217 (N.D. Cal. Jan. 22, 2024) (Tigar, J.).  Judge Orrick noted that other districts have found that large language models have and obligation to maintain the original creators' CMI on nearly identical works.

ABA Issues Formal Ethics Opinion on Use of Generative AI Tools

By Joshua Rich --

Following in the footsteps of the U.S. Patent and Trademark Office[1] and the state bars of California,[2] Florida,[3] New Jersey,[4] New York,[5] and Pennsylvania,[6] the American Bar Association has weighed in on attorney's ethical use of Generative AI (GAI) tools with a formal ethics opinion entitled "Generative Artificial Intelligence Tools."[7]  The ABA opinion highlights many of the same ethical rules as the previous guidance, opinions, and reports, but from a different perspective.  As a result, it identifies issues and proposes ethical requirements slightly different from others.  And while the ABA's suggested steps for discharging ethical obligations are not binding on any attorney, the concerns are universal and the suggested steps likely to be persuasive if complications arise.

Unlike the other guidance, the ABA formal opinion is limited to ethical considerations arising out of generative AI.  The ABA recognizes that lawyers are already using AI in many contexts, ranging from legal research to technology-assisted document review to contract analytics.  There are ethical issues that arise in those other contexts, but they are different from those that relate to GAI.  Further, the opinion recognizes that the guidance would need to be updated as technology develops, "anticipat[ing] that [the ABA] Committee and state and local bar association ethics committees will likely offer updated guidance on professional conduct issues relevant to specific GAI tools as they develop."[8]

The ABA opinion starts where the Model Rules do, with the duty of competence.[9]  There are three ways that the use of GAI implicates the duty of competence:  knowing the GAI tools available to be used, understanding the capabilities and liabilities of any GAI tool the lawyer chooses to use, and ensuring that use of the GAI tool does not return inaccurate information.

On the first issue, knowledge of available GAI tools, the opinion counsels that:

[E]ven in the absence of an expectation for lawyers to use GAI tools as a matter of course, lawyers should become aware of the GAI tools relevant to their work so that they can make an informed decision, as a matter of professional judgment, whether to avail themselves of these tools or to conduct their work by other means. . . .  Ultimately, any informed decision about whether to employ a GAI tool must consider the client's interests and objectives.[10]

That is, lawyers cannot remain competent by simply ignore the possible use of GAI tools; they must learn whether such a tool is reasonably necessary for their client's work.

Once lawyers decide to use a GAI tool, they must understand the tool well enough to be able to explain it to clients, to allow them to make an informed decision whether the tool should be used for their project.

This means that lawyers should either acquire a reasonable understanding of the benefits and risks of the GAI tools that they employ in their practices or draw on the expertise of others who can provide guidance about the relevant GAI tool's capabilities and limitations.  This is not a static undertaking.  Given the fast-paced evolution of GAI tools, technological competence presupposes that lawyers remain vigilant about the tools' benefits and risks.  Although there is no single right way to keep up with GAI developments, lawyers should consider reading about GAI tools targeted at the legal profession, attending relevant continuing legal education programs, and, as noted above, consulting others who are proficient in GAI technology.[11]

For most lawyers, this means they will have to continually ensure they understand the benefits and risks of not only the technology they are currently using, but also updates and new tools.  GAI tools will therefore add to the educational burden borne by lawyers.

Finally, the ABA's opinion highlights one of the most notorious risks of using a GAI tool, providing inaccurate responses such as "hallucinations" that would lead to incorrect legal advice or made up citations submitted to courts.  The formal opinion asserts that lawyers must engage in "an appropriate degree of independent verification or review of [the] output," with the level of review dependent on the tool and task being performed.[12]  For submission to a court or critical advice, careful review of every citation and statement would be in order;[13] for basic letters or other less important work, less effort might be needed.

The opinion next addresses the duty of confidentiality, perhaps the most acute concern for most lawyers in using GAI tools.  All of the previous guidance identifies the risk of submitting a client's confidential information in prompts, which may run afoul of a lawyer's duty to avoid disclosure of such information.  That is, client information included in a GAL tool prompt is put in the hands of the GAI tool model, and may be used to teach the model and get disclosed to others.  But the opinion emphasizes another ethical risk unique to law firms: potential disclosure or use within the firm of one client's information for the benefit of another.  The opinion identifies considerations that lawyers must consider in both situations, as well as how to discharge the related ethical duties.

As a general matter, a lawyer must first determine if client information will be adequately protected from disclosure.  "In considering whether information relating to any representation is adequately protected, lawyers must assess the likelihood of disclosure and unauthorized access, the sensitivity of the information, the difficulty of implementing safeguards, and the extent to which safeguards negatively impact the lawyer's ability to represent the client."[14]  Those considerations intersect with the duty of competence, as a lawyer must understand the GAI tool and associated issues to evaluate those considerations.

The novel concern addressed in the opinion is intra-firm disclosure of client confidences through the use of a GAI tool.  The opinion sees no way to avoid such disclosure (as long as the firm uses the tool on more than one client's projects) and, instead, suggests that lawyers obtain informed consent for such potential disclosure from clients:

[A GAI tool] may disclose information relating to the representation to persons in the firm (1) who either are prohibited from access to said information because of an ethical wall or (2) who could inadvertently use the information from one client to help another client, not understanding that the lawyer is revealing client confidences.  Accordingly, because many of today's self-learning GAI tools are designed so that their output could lead directly or indirectly to the disclosure of information relating to the representation of a client, a client's informed consent is required prior to inputting information relating to the representation into such a GAI tool.[15]

Of course, if client confidences are segregated within a GAI tool, the risk of disclosure dissipates; using the tool in that way, however, severely limits the benefits of the tool.  More likely, the lawyer will have to obtain the client's informed consent to a potential disclosure through use of the GAI tool.  In either circumstance, however, under the duty to reasonably consult with the client, "clients would need to be informed in advance, and to give informed consent, if the lawyer proposes to input information relating to the representation into the GAI tool."[16]

For the consent to be informed, the client must have the lawyer's best judgment about why the GAI tool is being used, the extent of and specific information about the risk, including particulars about the kinds of client information that will be disclosed, the ways in which others might use the information against the client's interests, and a clear explanation of the GAI tool's benefits to the representation.  Part of informed consent requires the lawyer to explain the extent of the risk that later users or beneficiaries of the GAI tool will have access to information relating to the representation.  To obtain informed consent when using a GAI tool, merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient.[17]

In order to provide the fulsome explanation necessary to obtain informed consent, a lawyer will have become educated about the specific GAI tool, at least in terms of the legal obligations related to access to information:

As a baseline, all lawyers should read and understand the Terms of Use, privacy policy, and related contractual terms and policies of any GAI tool they use to learn who has access to the information that the lawyer inputs into the tool or consult with a colleague or external expert who has read and analyzed those terms and policies.  Lawyers may need to consult with IT professionals or cyber security experts to fully understand these terms and policies as well as the manner in which GAI tools utilize information.[18]

This required self-education is not unlike that which a lawyer must undertake in other situations where they entrust data to supervised personnel or third parties.  They must also establish clear policies for permissible use of GAI and take reasonable steps to ensure compliance with those policies (and all professional obligations) by subordinate lawyers, other firm personnel, and third parties.[19]

Finally, the opinion raises potential effects that GAI tools may have on the reasonableness of fees charged.  Lawyers charging an hourly rate must bill only their actual time worked; they cannot "value bill" for the efficiency realized through use of a GAI tool.  Even if the lawyer charges a flat fee, if the use of a GAI tool avoids all or nearly all work, the fee may be unreasonable.  And charging the client for the use of a GAI tool may not be ethical.  "To the extent a particular tool or service functions similarly to equipping and maintaining a legal practice, a lawyer should consider its cost to be overhead and not charge the client for its cost absent a contrary disclosure to the client in advance."[20]  And a lawyer cannot charge for all of the education needed to learn about the GAI tool and other issues necessary to obtain informed consent from the client.

In short, the ABA formal opinion points out the many potential ethical pitfalls that arise out of the use of a GAI tool.  But the opinion also provides some guidance on how to avoid those pitfalls.  As tools develop and become better integrated into law firm practice, the requirements set forth in the opinion should become less burdensome and easier to meet.

[1] "Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the United States Patent and Trademark Office," 89 Fed. Reg. 25,609 (Apr. 11, 2024).  An outstanding discussion of the PTO's Guidance is available at https://www.patentdocs.org/2024/04/the-usptos-guidance-on-use-of-ai-based-tools-in-practice.html.

[2] State Bar of Cal. Standing Comm. On Prof'l Resp. & Conduct, "Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law" (2023), available at https://www.calbar.ca.gov/Portals/0/documents/ethics/Generative-AI-Practical-Guidance.pdf.

[3] Fla. State Bar Ass'n, Prof'l Ethics Comm., Op. 24-1 (Jan. 19, 2024), available at https://www.floridabar.org/etopinions/opinion-24-1/.

[4] NJ S. Ct. Comm. on AI & the Cts., "Preliminary Guidelines on New Jersey Lawyers' Use of Artificial Intelligence" (Jan. 24, 2024), available at https://www.njcourts.gov/sites/default/files/notices/2024/01/n240125a.pdf?cb=aac0e368.

[5] NY State Bar Ass'n Task Force on Artificial Intelligence, "Report and Recommendations of the New York State Bar Association Task Force on Artificial Intelligence" (Apr. 6, 2024), available at https://nysba.org/app/uploads/2022/03/2024-April-Report-and-Recommendations-of-the-Task-Force-on-Artificial-Intelligence.pdf .

[6] Pa. State Bar Ass'n Comm. on Legal Ethics & Prof'l Resp. & Philadelphia Bar Ass'n Prof'l Guidance Comm., Joint Formal Op. 2024-200 "Ethical Issues Regarding the Use of Artificial Intelligence," (May 22, 2024), available at https://www.pabar.org/Members/catalogs/Ethics%20Opinions/Formal/Joint%20Formal%20Opinion%202024-200.pdf.

[7] Am. Bar Ass'n Standing Comm. on Ethics & Prof'l Resp., "Generative Artificial Intelligence Tools" Formal Op. 512 (July 29, 2024), available at https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf.

[8] ABA Formal Op. 512, p. 2.

[9] See Model Rules R. 1.1.

[10] ABA Formal Op. 512, p. 5.

[11] ABA Formal Op. 512, p. 3.

[12] ABA Formal Op. 512, p. 4.

[13] As the opinion points out, submission of information to a tribunal that has been provided by a GAI tool also implicates Model Rules 3.1, 3.3, and 8.4(c).  ABA Model Op. 512, p. 9-10.  The same duties would apply with regard to submissions to the USPTO.

[14] ABA Formal Op. 512, p. 6.

[15] ABA Formal Op. 512, p. 7.

[16] ABA Formal Op 512, p. 8 (citing Model Rules of Prof'l Conduct R. 1.4).  Even if no client information will be inputted, the client must be informed that a GAI tool is used if it asks.

[17] ABA Formal Op. 512, p. 7.

[18] ABA Formal Op. 512, p. 7.

[19] ABA Formal Op. 512, p. 10-11 (citing Model Rules of Prof'l Conduct R. 5.1, 5.3).

[20] ABA Formal Op. 512, p. 13.

USPTO Publishes Updated Subject Matter Eligibility Guidance Focusing on AI

By Michael Borella --

As required by President Biden's Executive Order 14110 ("Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence"), the U.S. Patent and Trademark Office has published an update to its subject matter eligibility guidance for examiners regarding inventions employing artificial intelligence (AI).  As this is the first update to the USPTO's subject matter eligibility guidance in almost five years, there was some hope that it would address stakeholders' long-held concerns.  These include the difficulty of crafting claims and a specification directed to an invention that incorporates AI, especially in view of the dramatic disparity in how examiners apply the law in view of the USPTO's previous guidance.

However, the updated guidance is remarkably non-substantive at best, and at worst appears to take the position that advances in AI and uses of AI are ineligible unless they are implemented in specific hardware circuitry or cause a significant change to the state of an ancillary system.  Moreover, the updated guidance largely reads as general eligibly guidance with few actual AI-specific considerations.

Indeed, the updated guidance makes it painfully clear that it (like the USPTO's previous guidance on the topic) "does not create any right or benefit, substantive or procedural, enforceable by any party against the USPTO."  In other words, examiners are free to ignore the updated guidance as they see fit, which is how examiners have approached the USPTO's previous subject matter eligibility guidance over the last decade.

In any event, it is not until the halfway point of the updated guidance that it addresses the USPTO's currently eligibility test.  In particular, the updated guidance focuses on Step 2A, prong one (determining whether a claim recites a judicial exception) and step 2A, prong two (determining whether the claim integrates the recited judicial exception into a practical application of the exception).  Unfortunately, the USPTO largely employs hypothetical examples in doing so.

Regarding Step 2A, prong one, the USPTO reminds us that "USPTO personnel must draw a distinction between a claim that 'recites' an abstract idea . . . and one that merely involves, or is based on, an abstract idea."  In the former case, the claim requires further eligibility analysis and in the latter case the claim is eligible.

The USPTO sets forth three hypothetical examples of claims that do not recite an abstract idea: (i) custom and specific circuits designed to implement arrays of neurons and synaptic weights, (ii) a processor, livestock herd monitor, and transmitter collectively arranged for monitoring health and activity in a herd of dairy livestock animals, and a "treatment method comprising administering rapamycin to a patient identified as having Nephritic Autoimmune Syndrome Type 3."  Notably, only the first of these examples even relates to AI.

The USPTO goes on to state that examiners should identify specific claim limitations that fall within the three groupings of abstract ideas:  mathematical concepts, certain methods of organizing human activity, and mental processes.  This leads to the rather absurd result that more specific recitations in a claim (which are encouraged by a large number of Federal Circuit eligibility opinions) can ultimately render a claim ineligible, whereas broader recitations of the same invention (which are discouraged by these Federal Circuit eligibility opinions) can result in a claim being eligible because it does not recite an abstract idea.

Regarding math, the USPTO essentially repeats it's thin and unworkable distinction between claims that recite mathematical concepts and claims that are based on or involve a mathematical concept.  In practice, examiners do not understand this distinction and the USPTO's terse discussion in the updated guidance suggests that they are not alone.  To state what should be obvious, all science and engineering advances involve mathematical concepts in one way or another -- a claim that recites "taking the sum of a first output and a second output" should not be subject to more scrutiny than a claim that recites "determining a result based on a first output and a second output."

The certain methods of organizing human activity category has been viewed expansively by the courts and some examiners to include anything remotely related to human activity.[1]  Here, the USPTO relies on three recent Federal Circuit eligibility opinions to flesh out the category.  The problem is that all three examples found claims to be ineligible.  It would be nice if the USPTO could provide at least a hint of what might not fall within this gaping chasm of a category.

On mental processes, the USPTO continues to apply what can only be considered a legal fiction.  In essence, the current practice is that any invention that, even in theory, could be performed by one or more human beings with thought, pencil and paper, or other tools, is a mental process even if it would take millions of years to obtain a useful result.  Indeed, the mental process category encompasses processes that are not performed mentally.

Here, the USPTO at least gives us one example of an invention that (in its view) does not recite a mental process ("a specific, hardware-based RFID serial number data structure [that is] uniquely encoded").  However, when compared to the other examples that do recite a mental process (e.g., "the collection of information from various sources and understanding the meaning of that information"), the USPTO appears to be slicing the bacon transparently thin.[2]

Moving on to Step 2A, prong two, the USPTO reiterates that examiners should determine whether "there are any additional elements recited in the claim beyond the judicial exception(s), and [evaluate] those additional elements individually and in combination to determine whether they integrate the exception into a practical application of that exception."  Despite the USPTO's reminder that claims are to be considered "as a whole" in this inquiry, such language is, in practice, a paper tiger.  Neither the USPTO nor the courts have ever explained what it means to consider claims as a whole, and both have routinely considered claims piecemeal while contending that they are actually considering claims as a whole.

There are three ways to find a claim non-abstract under Step 2A, prong two -- where the claim (i) improves the functioning of a computer, (ii) improves another technology or technical field, or (iii) is meaningfully limited to a particular technological environment.  But none of these have been defined in the case law, much less clearly defined.  The USPTO notes that specific solutions to specific problems are more likely to be eligible, but provides little insight that fleshes out this notion.  Instead, the USPTO again falls back on case law examples.

These examples demonstrate that an eligible invention may look ineligible if you tilt your head three degrees to the left and vice-versa.  For example, a very specifically-worded claim for determining haplotype phase based on pedigree data in human genetic sequences was an improvement over prior attempts to do so, but not a "technical" improvement because the entire claim was putatively a mathematical process that ran on general-purpose computers.

While the USPTO touted the claims of McRO, Inc. v. Bandai Namco Games America Inc. as an example of eligibility (which it is despite a lot of math being calculated by a general-purpose computer driving the thrust of the claimed invention), in practice it is rare that making an analogy to these claims will sway an examiner.  Instead (tilting their heads three degrees to the left), examiners typically conclude that any claim similar to McRO actually recites an ineligible abstract idea executing atop a non-specific chunk of silicon.[3]  In other words, there is little or no appreciable difference between the haplotype claim and those of McRO despite one being eligible and the other not.

Toward the very end of the updated guidance, the USPTO addresses AI-assisted inventions.  It states:

For the subject matter eligibility analysis under 35 U.S.C. 101, whether an invention was created with the assistance of AI is not a consideration in the application of the Alice/Mayo test and USPTO eligibility guidance and should not prevent USPTO personnel from determining that a claim is subject matter eligible.  In other words, how an invention is developed is not relevant to the subject matter eligibility inquiry.

Thus, the words of the claim matter but the use of AI in developing the underlying invention does not.  Aside from this clear statement, the updated guidance is cagey in general and surprisingly thin on AI-specific content.[4]

Practitioners have looked forward to iterations of the USPTO's subject matter eligibly guidance to add clarity and expedite prosecution.  After a decade of such guidance, it is clear that the USPTO's example-based approach has failed.  The chosen examples do not clarify the murky principles and distinctions that the words of the guidance set forth.  Moreover, the USPTO's reasoning regarding these examples is contradictory, muddled, opaque, and ultimately not helpful.  The greatest evidence for this is how different examiners apply the guidance in dramatically different ways.

Again, the fault lies with the Federal Circuit's incoherent parade of patent eligibility horribles masquerading as judicial decisions.  But instead of acknowledging that the emperor is stark raving naked, the USPTO once again tries to pretend otherwise.

It is time to stop.

[1] One of my favorite examples is the examiner I spoke to who insisted that fully-automated navigation of a self-driving vehicle was a method of organizing human activity because the goal of the invention was to deliver people from point A to point B.

[2] Credit for this confusion should be placed at the feet of the Federal Circuit, the decisions of which these examples are based.  

[3] A relevant definition of "analogy" in the American Heritage Dictionary is "[a] form of reasoning based on the assumption that if two things are known to be alike in some respects, then they are probably alike in other respects."  Too many examiners seem to think that, to be analogous, two things must be exactly the same.

[4] The USPTO has released three new sets of example claims to support the guidance.  They were not specifically addressed herein and may be discussed in a subsequent article.  In short, they do not provide enough further insight to render the updated guidance particularly helpful.

How AI Is Changing the Internet

By Michael Borella --

This Internet has gone through many revolutions, technical and otherwise.  Each time it has emerged stronger and more robust than before.  One can trace the origins of the Internet to the connection of four computers in 1969.  The advent of email a few years later, the standardization of TCP/IP as its communication protocol in the mid-1980s, commercial dialup service providers not long after that, the popularization of the World Wide Web in 1993, and the immediately-following rise of ecommerce gave us the main aspects of what we currently think of as the Internet.  Inexpensive broadband, the search engine, online gaming, social media, widespread mobile access, and videoconferencing bring us to today.

The next Internet revolution is being driven by generative AI, and it is happening right now.  While we cannot say what is going to be in place on the other side of this inflection point, we do know that our online experience will be different in a few years.

Consider just the web.  It relies on an economic model that has been largely static for over two decades.  But generative AI is already beginning to disrupt this model.  Today, publishers and search engines have a mostly symbiotic relationship.[1]  They need each other.  But what happens when AI is used to subsume publisher content?

Publishers (including small and independent publishers, blogs, businesses, and so on) seek to be highly ranked in search engines, especially for search terms that are relevant to their content.  For example, an independent newsletter that reviews cars (as just one example) would want to appear as high as possible (preferably in the top ten) search results for queries involving the word "car", "automobile", "motor", and so on.  This will drive traffic to the publisher's newsletter, allowing the publisher to sell ads and/or subscriptions.  With the resulting revenue, the publisher can grow its newsletter by hiring more writers and editors, who in turn produce more content.  This content, if of sufficient quality, makes it more likely that the publisher will be highly ranked in search results -- a virtuous cycle.

Search engines (or more correctly, search engine providers) crawl the web and use sophisticated algorithms to rank publishers in terms of their sites' relevance to search terms.  A successful search engine may receive billions of search requests per day from users who want to be referred to relevant publishers.  Each page of search results may be an ordering of links to publishers and perhaps a short description or summary of the content that can be found by following each link.  Search engines make money by displaying advertisements as sponsored links.  They also allow publishers to bid on search terms in a form of auction.  The publisher who bids the highest for a given search term often ends up at the top of the search result as a sponsored link, assuming that their content is indeed relevant to the search term.

While this symbiosis has not yet been broken, it is getting wobbly.  It is no coincidence that the largest investors in generative AI are the companies that own and operate the largest search engines.  OpenAI, which is nearly half-owned by Microsoft, has admitted to training its GPT series of models on content that it absorbed from publishers on the Internet, often without their permission.  Some would contend that the models that underlie the popular ChatGPT generative AI tool have been trained on the Internet as a whole.[2]  Indeed, The New York Times and other publishers have sued OpenAI and Microsoft for allegedly violating their copyrights and (at least in the case of the Times) allegedly producing near-verbatim copies therefore in response to large language model (LLM) prompts.[3]

Search engine results recently changed from being a series of links to publisher sites (with ad-sponsored results on top and clearly flagged) to leading with a generated AI overview as the leading result.  This overview seeks to answer the searcher's question or address their need without requiring them to click on links and visit publisher web sites.  So far, AI overview sections have also been displaying what the underlying AI model infers to be the most relevant publisher links and/or those used a source material for the overview.

In isolation, AI-generated search results are neither here nor there.  Users might appreciate having them summarize publisher content to provide results in a more convenient and readily-consumable form.  Of course, the generation and providing of these results is still clearly experimental and subject to the same incorrectness, hallucinations, and bias that plague LLMs.  But users may further appreciate not having to navigate to certain publisher sites that display annoying ads or are sophisticated clickbait.

At this point it seems safe to assume that the quality of these AI summaries will continue to improve.  If so, at what point does this impact the publishers?  The news industry -- one of the biggest and most important publishing sectors -- has lost one-third of its newspapers and two-thirds of its news journalists in the last two decades, mostly in the area of local news.[4]  AI is likely to accelerate this decline, with search engines absorbing user traffic rather than directing it to publisher sites.  As a result, publishers are likely to have less ad revenue and fewer subscriptions, putting their businesses at risk.

Given that a sea change toward AI-generated search results is playing out in real time, it is difficult to predict what the economic model the web will be based on five years from now.  However, we can game out a few possible scenarios.

• Search engines use AI to subsume the publishing industry.  Publishers go out of business en masse, leaving only the LLMs that generate news-like search results.  Many view this possible future as dystopian and potentially dangerous for society without the fourth estate, as there would be fewer incentives to subsidize investigative journalism that seeks to root out corporate and governmental corruption.  In a world without news organizations, it would be extraordinarily difficult to determine "the truth" as we would rely mostly on second and third hand accounts and other forms of hearsay.  Ironically, this would also reduce the utility of search engines, likely resulting in a revenue drop for those companies as well.

• In a slightly less dystopian future, publishing does not die but instead gets sorted into two camps.  The first are traditional publishers with large audiences, like the Times, who survive due to their name recognition and quality journalism.  The second consists of small independent publishers with niche audiences that survive on subscriptions.  Both have the ability to be a source of real-time "hot" news that is too fresh for the AI models to ingest.  They also can serve as destinations for users (e.g., the users check their apps or bookmarked links frequently), allowing the publishers to flourish even without referrals from search engines.  This has been referred to as the "barbell" strategy, where the sweet spots for publishers are at either end of the spectrum in terms of size (the very large media companies or the very small independents) with very little in between.

• Another dark scenario is that the search engines buy up most or all major publishers, reorienting the publishing business from serving the public directly toward generating training data for AI models.  While the incentives here are complex, it is likely that the search engine companies would have at least some control over reporting.  This could lead to less reporting on any negative externalities associated with these companies or -- even worse -- different "facts" being used to train different search engines.  If this notion seems far-fetched, keep in mind that the Washington Post is owned by Amazon founder Jeff Bezos[5] and that Meta was considering the purchase of Simon & Schuster to train its models.[6]

• A more likely path forward is that the search engine providers (and AI companies in general) sign licensing deals with major publishers in order to access and use their content.  This is already happening, as the Financial Times, News Corp., Axel Springer, Le Monde, Prisa Media, and The Associated Press have all agreed to license their content to OpenAI.

• Another possibility is that search engine providers are legally prohibited from unauthorized ingestion of publisher content without permission, which makes the licensing discussed above even more appealing.  This might be a consequence of a victory for the Times in its dispute with OpenAI.

Right now, it is too early to tell which of these scenarios may play out, or if a different one arises.  If anything, we can be certain that the business model of the Internet will be different in five years or less.  Search engines will return AI-generated answers with less emphasis of links to source material, publishers will need to engage more directly with their audience and supply a continuous stream of high-quality content, and search-engine optimization may become obsolete.

Like all disruptions, AI search comes with opportunities as well. As the technological and legal landscape evolves we should not be surprised if new models of publishing emerge, adapted to this evolution.

[1] This is an admittedly a simple overview of the relationship.  There is quite a bit more nuance, but this level of detail is sufficient for the discussion that follows.

[2] https://futurism.com/the-byte/ai-training-data-shortage.

[3] https://www.patentdocs.org/2024/02/the-new-york-times-case-against-openai-is-different-heres-why.html.

[4] https://localnewsinitiative.northwestern.edu/posts/2024/05/08/ai-local-news-report/index.html.

[5] https://www.theguardian.com/media/2023/jan/29/tech-moguls-media-jeff-bezos-washington-post.

[6] https://www.theguardian.com/books/2024/apr/09/meta-discussed-buying-publisher-simon-schuster-to-train-ai.

Black Widow Versus OpenAI and What it Means for Singers and Voice Actors

By Michael Borella and Joshua Rich --

It seems like every week brings a new dispute over artificial intelligence (AI), mostly focused around new features and uses of generative AI.  Last week was no exception, with OpenAI unveiling human-like voices soon being available in its GPT-4o multimodal language model.  In what would turn out to be one of the biggest tech executive self-owns this side of Elon Musk, OpenAI CEO Sam Altman tweeted (Xed?) the single word "her" on the same day that the new feature was announced.

Almost immediately, the public put together the dots.  One of the GPT-4o voices, named Sky, was eerily similar to that of actress Scarlett Johansson, and Johansson had supplied the voice for an AI voice assistant in the 2013 Spike Jonze movie, Her.[1]  Johansson released a statement stating that Altman had been in contact with Johansson for months, seeking to use her voice in what would become GPT-4o.  Johansson had repeatedly refused.

OpenAI responded to the controversy by "pausing" the use of Sky with its new model, then later stating that they used a voice actor to develop Sky and did not intentionally mimic Johansson's voice.[2]

If she were to take OpenAI to court, Johansson would have a good case.  Despite copyright law not providing protection for one's voice, the common law may (at least in California).  There is clear precedent arising out of similar attempts to mimic a celebrity's voice when the celebrity chose not to participate.

In 1985, the Ford Motor Company created an ad campaign for one of its cars.  Ford asked singer Bette Midler to sing one of her own songs in the commercials, but Midler's agent categorically refused on her behalf.[3]  Undeterred, Ford hired one of Midler's former backup singers to record Midler's song, Do You Want to Dance, for the commercial.  The backup singer was asked to sound as much as possible like Midler.  When the commercials aired, many people felt the voice in the commercial sounded exactly like Midler's; some believed that it was Midler singing.

Midler filed suit claiming violation of the California common law right of publicity.  The District Court granted summary judgment to Ford, finding no legal principle protecting imitation of a famous voice, as opposed to using the famous person's own voice (which could amount to copyright infringement) or physical image (which was protected by both statute and common law in California).

But Midler won a reversal on appeal to the 9th Circuit, the panel holding that "California will recognize an injury from an appropriation of the attributes of one's identity."[4]  The panel wrote:

A voice is as distinctive and personal as a face.  The human voice is one of the most palpable ways identity is manifested.  We are all aware that a friend is at once known by a few words on the phone . . . .  A fortiori, these observations hold true of singing, especially singing by a singer of renown.  The singer manifests herself in the song.  To impersonate her voice is to pirate her identity.

We need not and do not go so far as to hold that every imitation of a voice to advertise merchandise is actionable.  We hold only that when a distinctive voice of a professional singer is widely known and is deliberately imitated in order to sell a product, the sellers have appropriated what is not theirs and have committed a tort in California.[5]

If the Midler case establishes that it would be illegal to deliberately imitate the distinctive voice of a professional singer who is widely known, a question remained as to who would be considered "widely known."  A few years later, some limits were added to that inquiry.

Singer Tom Waits emerged from folk music in the 1970s with a signature gravelly voice and songs about the darker sides of society.  Though an inductee into the Rock and Roll Hall of Fame, Waits is nowhere near as famous as Midler, and has what could be described as "cult" following despite his songs being widely licensed.[6]

Frito-Lay created a radio commercial for a new variation of its Doritos chips.  It asked Waits to provide his voice to the commercial, but he refused.[7]  So Frito-Lay hired a Waits imitator to provide the soundtrack for the commercial.

Waits sued and was awarded $2.6 million at trial.  On appeal, the 9th Circuit affirmed the decision, stating that the trial court properly instructed the jury that a widely known voice "is known to a large number of people throughout a large geographic area."[8]

Notably, Midler and Waits could not have prevailed on grounds of copyright infringement.  Ford had obtained a copyright license from the writers of Midler's song, and Waits did not own the copyright to his.  But their claims on grounds of right to publicity might not be available to any professional singer, only to those with distinctive and widely known voices.

Although Johansson is a singer,[9] her speaking voice is at issue in her dispute with OpenAI.  Nonetheless, the Midler and Waits cases provide a persuasive foundation for concluding that OpenAI violated her publicity rights.  Like in those cases, she was asked for permission to use her voice, refused, and the other party hired someone with a remarkably similar voice.  Further, Johansson's voice is distinctive and widely known from her three decades as an actress.  And perhaps most importantly, her voice had been "Her"[10] voice -- the paradigm of an AI voice.  Thus, OpenAI would have enjoyed the commercial benefit of Johansson's implicit endorsement of its technology even though it had failed to obtain her approval.

As artists and creatives wrestle with how to deal with generative AI models being trained on their works and potentially spitting out substantially similar versions thereof, the Midler and Waits line of cases provides yet another arrow in the legal quiver to protect their life's work.  The difference is that an AI model could be trained to replicate a famous person's voice rather than a sound-alike human being found.[11]  Nonetheless, it remains unclear what level of celebrity is required to trigger a claim.  Less well known or regional singers or voice actors may not qualify, for example.

Also, right to publicity laws overlap to some extent with emerging anti-deepfake legislation.  The latter may be more readily available to less famous individuals -- for example, the proposed "No Fakes Act of 2023" provides a cause of action and remedies for a deepfake victim so long as they are "a human being, living or dead."[12]

Thus, before long, you may not need to be an Avenger to take on Sam Altman.

[1] Ironically, Her was dystopian and did not have a happy ending; it is not a good comparison for a company trying to sell the benefits of AI to the world.

[2] A comparison of the two voices can be found here.  Sky does sound robotic, but the resemblance is clear.

[3] According to the court records, "[t[he conversation went as follows:  'Hello, I am Craig Hazen from Young and Rubicam.  I am calling you to find out if Bette Midler would be interested in doing . . . ?' [Midler’s agent]:  'Is it a commercial?'  'Yes.'  'We are not interested.'"

[4] Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988).

[5] Id., 849 F.2d at 463.

[6] Ironically, Waits had an on-and-off romantic relationship with Midler.

[7] Waits was very outspokenly against using his voice or music in commercials.

[8] Waits v. Frito-Lay, Inc., 978 F.2d 1093, 1102 (9th Cir. 1992).

[9] More irony:  Johansson is a fan of Waits and has covered his songs.

[10] In the movie, the name behind the voice was Samantha.

[11] There was just such a dispute one week earlier, when Drake used an AI replica of the voice of the deceased (or allegedly so) rapper Tupac Shakur, as well as an AI replica of the voice of the living rapper Snoop Dogg, in his diss track "Taylor Made Freestyle."  The Shakur estate threatened to sue Drake and he took the track down from social media immediately.

[12] Tennessee has enacted the Ensuring Likeness, Voice, and Image Security (ELVIS) Act to protect the voices of singers and other musicians, both living and dead.

State Regulations of AI in Elections

By Michael Borella and Yuri Levin-Schwartz* --

The recent rise in generative artificial intelligence (AI) models has granted powerful tools to the public that enable the creation of realistic, yet fake, images, sounds, and videos.  Though more than six months remain until the U.S. general election in November, these tools have already been used to create fake calls from President Biden and fake images of President Trump.

Motivated by the impact that such tools could have on an election, at least ten states have enacted statutes governing the use of AI models in electioneering.  These states are California, Indiana, Michigan, Minnesota, New Mexico, Oregon, Texas, Utah, Washington, and Wisconsin (some of these statutes do not specifically mention or imply AI but are written broadly enough to cover use of AI).  Moreover, four states, Michigan (bill 1, bill 2), Minnesota, New York, and Washington, recently enacted further statutes governing the use of synthetic media in elections.**

The statutes have a variety of proscriptions, ranging from requiring the disclosure of the use of AI models to creating the possibility of injunctions and criminal charges for the deceitful use of such models.  While potentially powerful, many of these statutes require a showing for deceitful intent.  Such a showing may prove difficult to establish and it may hinder effective use of these statutes to regulate AI-generated content especially if such content is released shortly before an election.

Generally, state statutes do not call for an outright ban on the use of AI models to generate images, sounds, or videos relating to candidates or political parties.  However, nearly all statues require some form of disclosure to indicate that the content was generated using AI.  An example is California Election Code § 20010, which prohibits the distribution of "materially deceptive audio or visual media . . . of [a] candidate with the intent to injure the candidate's reputation or to deceive a voter into voting for or against the candidate."  This prohibition does not apply if the audio or visual media includes a disclosure, which states: "This [Image/Video/Audio] has been manipulated."  Such a disclosure must "appear in a size that is easily readable by the average viewer" or "clearly spoken manner and in a pitch that can be easily heard by the average listener."  Such disclosures are crucial, especially as AI-generated content becomes increasingly common and humans are less able to differentiate real images from those generated using AI.

Though each state's statute is different, a common trend across these statutes is the creation of disincentives for the misuse of AI-generated content in elections.  Three states, Michigan, Minnesota, and Texas, establish criminal penalties for the misuse of AI-generated content in the context of an election.  One such example is Texas Election Code Ann. § 255.004, which makes it a Class A misdemeanor to create and cause to be published or distributed "a deep fake video," defined as "a video, created with the intent to deceive, that appears to depict a real person performing an action that did not occur in reality."  The punishment for committing a Class A misdemeanor is a fine not exceeding $4,000 and/or a jail term not exceeding a year.

As noted, many statutes addressing the use of AI in elections has some form of intent requirement, generally an intent to deceive or affect an election.  Such a requirement has the advantage of limiting liability in cases where the use of AI was not intended to harm an individual or affect an election.  However, such a threshold can hinder effective use of injunctive relief, especially when content is created shortly before an election.  Interestingly, neither New Mexico's statue (N.M. Stat. Ann. § 1-19-26.4) nor Indiana's statue (Ind. Code § 3-9-8-6) require intent for liability.  Thus, both statutes regard misuse of AI during an election as a strict liability offense.  But like other states, both allow the publication of election content generated using AI if the content has a disclosure acknowledging the use of AI in the production of such content.

So far, a relatively small number of states have taken steps to address potential misuse of AI during the upcoming 2024 election.  These statutes could help curb misuse of AI in the generation of misleading video, pictorial, and audio content -- but only within those states.  Further, only two of these states (Michigan and Wisconsin) are considered "swing states" that could end up deciding the election.  Notably, none of these statutes would punish the developers of AI models, thus implicitly acknowledging the dual use nature of these models.

* Yuri Levin-Schwartz, Ph.D., is a law clerk at MBHB.

** For more information on state statutes governing the use of AI in elections, see "Artificial Intelligence (AI) in Elections and Campaigns".

Palworld: Friend or Foe?

By Andrew Velzen --

On January 19, 2024, Palworld launched into early access on Xbox and Windows.  To say it was an overnight success is an understatement.  By only a month in, Palworld had been played by over 25 million players.[1]  To put this in perspective, monumental success stories like The Legend of Zelda: Tears of the Kingdom and Hogwarts Legacy (both released in 2023) have sold in the ballpark of 20-25 million copies.[2]  It is not surprising that Palworld's meteoric rise has generated a lot of buzz in the industry.

What is Palworld

Palworld is a survival game developed and published by Pocket Pair that is "set in an open world populated with animal-like creatures called 'Pals,' which players can battle and capture to use for base building, traversal, and combat.  Players may also assign the 'Pals' to bases where they will automatically complete tasks for the player.  Palworld can be played either solo or online with up to 32 players on one server."[3]  Most saliently for this discussion, though, is the fact that the "Pals" bear a not-so-subtle resemblance to Pokémon (e.g., in name and/or likeness).[4]  As just one of countless examples, compare the following image of the Pal Grizzbolt (left) with the Pokémon Electabuzz (right):[5]

In fact, given the visual overlap, gamers have started colloquially referring to Palworld as "Pokémon with Guns."

Accusations of Infringement

Given the striking resemblance, it isn't surprising that many in the industry have wondered whether any of Game Freak Inc. / The Pokémon Company / Nintendo Co., Ltd. (collectively referred to herein as the "Pokémon Owners") would initiate action for copyright infringement against Pocket Pair, especially given Nintendo's notoriously litigious reputation.[6],[7]  Interestingly, Pocket Pair has a history of making games that appear, at least facially, to be clones of other games.[8]  As far as I'm aware, none of the Pokémon Owners have yet to take any legal action against Palworld.  That isn't to say that the Pokémon Owners aren't seriously considering such action, though.  In fact, regarding Palworld, The Pokémon Company issued the following statement:[9]

We have received many inquiries regarding another company's game released in January 2024.  We have not granted any permission for the use of Pokémon intellectual property or assets in that game.  We intend to investigate and take appropriate measures to address any acts that infringe on intellectual property rights related to the Pokémon.  We will continue to cherish and nurture each and every Pokémon and its world, and work to bring the world together through Pokémon in the future.

I have heard 3 main allegations levied regarding the creation of the Pals vis-a-vis Pokémon.  Namely, that Pocket Pair: (a) generally copied many concepts of Pokémon in a creatively bankrupt fashion, (b) obtained three-dimensional (3D) models of Pokémon and only made slight modifications to those models before pushing them out the door, and/or (c) trained a generative artificial intelligence (AI) using 3D models of Pokémon and then had that AI produce 3D models for Pals.[10]  To my knowledge, there has been no publicly accessible discovery thus far, so we can't, with absolute confidence, ascertain which of these three allegations (if any) are true.  Still, I plan to discuss the potential merits of a copyright infringement suit in the United States against Pocket Pair under each of these hypothetical copying regimes.[11]

What Would the Complaint Allege?

A successful copyright infringement suit requires that plaintiff has a registered copyright on an underlying copyrightable work (or an exclusive license thereto), that defendant infringed on an exclusive right conferred by the copyright (without a license), and that no defenses apply (e.g., there is no fair use defense).  The primary questions in a hypothetical Pokémon Owners v. Pocket Pair case are: (i) which aspects of Pokémon are covered by the Pokémon Owners' copyrights (i.e., which aspects are copyrightable) and (ii) would Pocket Pair have a reasonable fair use defense.  I will analyze both of these questions for each of the three sets of potential facts.

First – What if Pocket Pair simply copied some or most of the general concepts that make Pokémon successful?

As a first consideration, does the fact that Palworld is a game that involves catching monsters and using them in combat mean that Pocket Pair has infringed a copyright?  I think most people would inherently understand that this fact, taken alone, would be insufficient to constitute copyright infringement, but let's briefly discuss why.

As a general rule, "ideas" aren't copyrightable (rather, only expressions are).  This is, unsurprisingly, typically referred to as the "idea-expression dichotomy," and is codified in 17 U.S.C. § 102(b).[12]  One tenant of the idea-expression dichotomy is that certain concepts are so common / ubiquitous to a given medium or genre that such concepts are not copyrightable (these concepts are sometimes referred to as scènes-à-faire).  In the video game world, this could include things like user interface menus or combat or character creation, but can extend even further to the entire underlying genre of a game (e.g., the general concept of a "fighting game" is not protectable by copyright).  I won't get into it here, but for a good illustration of where the line is between what is a scène-à-faire and what is protectable by copyright, contrasting Atari v. Amusement World with Tetris v. Xio is instructive.[13]

Here, the idea of monsters fighting with one another or against a player is probably not something that could be covered by a copyright.  For example, the concept of monsters that take forms that are subtly or directly inspired by various animals, people, or forces of nature (as many Pokémon are), is something that has been present in the culture of Japan for thousands of years in the form of the Yōkai.[14]  As a specific example in the context of Pokémon / Palworld, the general concept of a fox (kitsune) with some magical powers is well-known, so it cannot be the subject of a copyright.  In fact, both games include multiple fox creatures (one of the Pals even being named "kitsun"):[15]

Given all this, it is unsurprising that, historically, the Pokémon Owners have not pursued litigation against games that are far more similar to Pokémon (at least in terms of gameplay) than Palworld is (e.g., Digimon World, Temtem, etc.).[16]

Second – What if Pocket Pair copied underlying Pokémon models?

So, while the general concepts employed in Pokémon aren't likely protectable by copyrights, what about the 3D character models, themselves?  While it is only conjecture at this point whether actual Pokémon models were copied in the production of Palworld, there are some in the industry that have promulgated what they believe to be strong evidence of such copying.[17]  This narrative has unsurprisingly been opposed by Pocket Pair.[18]  Regardless, for the sake of argument, let's assume that Pocket Pair did in fact have access to the underlying 3D Pokémon models and only made minor tweaks (e.g., slight modifications to limb-shape, face-shape, hair-shape, or hair color; remeshing; reskinning; modifying relative proportions; etc.) to arrive at the Palworld models.  Would this be copyright infringement?

3D models are protectable by copyright, so long as the 3D model isn't merely an exact one-to-one digital rendering of something that exists in the physical world (i.e., so long as some originality / creativity was present in the development of the model).[19]  Further, to my knowledge, no Pokémon are exact replicas of any naturally occurring creature (i.e., while there may be cow Pokémon, a 3D scan of an actual bovine was not merely replicated to generate Miltank).[20]  Given this, the 3D models of Pokémon could be subject to copyright protection.  Thus, if the 3D models of Pokémon were directly imported into Palworld, the Pokémon Owners would have a strong case for infringement.  In fact, this isn't so far off from something that almost happened (not as a result of the actions of Pocket Pair, but as a result of the actions of a Palworld modder).  The modder uploaded a YouTube video and a tweet that previewed a mod that imported Pokémon directly into Palworld.[21]  Nintendo, unsurprisingly, swiftly issued a Digital Millennium Copyright Act (DMCA) strike against both the video and the tweet.  Even further, in an abundance of caution, the popular modding site Nexus Mods preemptively banned the uploading of any Palworld mods related to Pokémon.[22]

So, we know that it would not be okay for Pocket Pair to directly import / use Pokémon models in Palworld, but, assuming Pocket Pair had lawful access to the Pokémon models (e.g., could reverse engineer them in a lawful way), would it be okay if Pocket Pair modified the models and put them into Palworld.  This is where the typical lawyer answer of "it depends" comes into play.

One of the rights that a copyright holder has (e.g., that the Pokémon Owners have here in relation to the 3D Pokémon models) is the right to prepare derivative works.[23]  The U.S. Copyright Office lists the right to prepare derivative works as including "[a] revision of a website" and "[a] new version of an existing computer program."[24]  As such, the Pokémon Owners own the right to prepare a "new version" of the original 3D Pokémon models.

This right only protects the Pokémon Owners to the extent that a resulting 3D model is derivative of the original work, though.  In some cases, the resulting work is so transformative that the producer of the resulting work has a fair use defense.  Hence, the question here becomes whether the modifications to the underlying Pokémon models is sufficiently transformative (and, therefore, subject to the fair use defense).

In order to determine what is and is not transformative, we can look to the 4 fair use factors: (i) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (ii) the nature of the copyrighted work; (iii) the amount and substantiality of the portion used in relation to the copyright work as a whole; and (iv) the effect of the use upon the potential market for or value of the copyrighted work.

The above factors are obviously debatable and fact-specific.  However, if the Palworld models are, in fact, only minor modifications to the Pokémon models, I believe at least 3 of the 4 factors would weigh against a finding of fair use.  Under these assumed facts, the following considerations are likely relevant to the factors recited above:

i.  The Palworld models are being used commercially, in bad faith (e.g., without permission from the Pokémon Owners and without public admission), and for essentially the same purpose as the Pokémon models (weighing against a finding of fair use);

ii.  The creation of the original Pokémon models was a highly creative endeavor (weighing against a finding of fair use);

iii.  Only a few dozen or so of the plethora of Pokémon models have been replicated in Palworld (weighing in favor of a finding of fair use); and

iv.  Palworld and Pokémon are both video games competing in the same market for gamers to purchase their games and, even if they have yet to do so, the Pokémon Owners could make a similar game to Palworld (hence, it was a market that the Pokémon Owners would've been able to exploit) (weighing against a finding of fair use).

One final note regarding fair use.  One of the stronger ways to demonstrate fair use is to show that a work is a parody.  One could clearly argue that "Pokémon with Guns" is a parody of Pokémon.  Palworld includes many aspects that are not only opposite of Pokémon, but clearly meant to poke fun at Pokémon.  For example, rather than petting a Pokémon, you can butcher a Pal or rather than an animal/trainer relationship between players and Pokémon, there can be a master/slave relationship between players and Pals.[25]  Thus, there is at least a colorable argument that Palworld exists to ridicule some of the underlying tenants of Pokémon.[26]

Still, in my view, the primary purpose of Palworld is not to make fun of Pokémon, but to be a commercial success in its own right by exploiting the exact same market that the Pokémon Owners could, themselves, exploit.  Hence, in my view, it would be difficult for Pocket Pair to argue that Palworld is a parody (especially in light of all the fair use factors considered above).

Third – What if they used the models to train an AI, though?

Under the third set of potential facts, Pocket Pair is assumed to have taken the Pokémon models, trained a generative AI (genAI) using the Pokémon models, and then used the genAI to produce the Pal models (e.g., by employing a generative adversarial network (GAN)).  This third set of circumstances is perhaps the most interesting (as the first set of circumstances likely does not constitute copyright infringement and the second set of circumstances likely does constitute copyright infringement).

The foundational question here is clearly "is it copyright infringement to use copyrighted material to train a generative AI?" Fortunately, to answer this question, we can look to a growing number of cases concerning this exact issue.  For instance, many developers of large language models (LLMs) are being sued for using copyrighted material to train their LLMs.[27]  Unfortunately, though, none of these cases have yet been decided, so we aren't sure how U.S. courts feel about this (which is exacerbated by the fact that none of the cases have percolated above the district court level either, and this issue will likely be an issue that appeals courts, or even the U.S. Supreme Court, will need to weigh-in on).

Still, we can compare / contrast the Palworld saga with one of the AI cases; let's use the one with perhaps the highest notoriety -- New York Times v. Microsoft and OpenAI.  In that case, the New York Times has alleged that OpenAI / Microsoft's genAIs "rely on large-language models that were built by copying and using millions of The Times's copyrighted news articles, in-depth investigations, opinion pieces, reviews, how-to guides, and more" and that "the outputs of the defendants' GenAI models compete with and closely mimic the inputs used to train them, copying Times works for that purpose is not fair use."[28]  Whether such use does, in fact, constitute fair use (contrary to the New York Times's assertion) is a judgment call.  From one point of view, the LLM is ingesting a bunch of data (with the Times's articles being just a portion of it) and then generating something wholly new based on a future prompt.  This is arguably not unlike the way in which a musician, after having listened to copious amounts of other songs from other artists, then takes inspiration from those songs and writes their own new song. Almost always, this new song is validly copyright-eligible.

The odd case where the new song is not copyright-eligible (and may, in fact, constitute copyright infringement) is when the new song is too similar to an older song.  For example, Pharrell Williams, Robin Thicke, and Clifford Joseph Harris (T.I.) lost a case against Bridgeport Music regarding the song "Blurred Lines" in view of the previous Marvin Gaye song "Got to Give It Up" since, in the eyes (ears?) of the jury, there was sufficient overlap between the i) signature phrase ii) hooks iii) bass lines iv) harmonic structures, and v) keyboard chords of the two songs to constitute copyright infringement.[29]

If I'm prognosticating by analogizing to the case of LLMs trained on copyrighted material, I'm inclined to guess that courts will think that it's more of a question of the output, rather than the inputs (i.e., more a result of what the models spit out, rather than how they're trained).  Said another way, you as an individual have a right to read an article you paid for.  You can also save a PDF copy of the article onto your computer for your own future reading.  Even further, you could perform processing (e.g., optical character recognition (OCR)) on that article if you like.  However, you cannot forward that article on to 10 other people.  Likewise, it stands to reason that you can use an article you purchased to train your own LLM, so long as the LLM doesn't output copyrighted material directly to third parties.  Hence, if an LLM trained on a number of New York Times's articles responds to a query with an answer based on information it has gathered from those articles, that's probably fine.  For example, if an LLM is prompted with "provide me with a brief summary of the Red Sox / Yankees game from October 20, 2004," and, based on information ingested from the New York Times's articles, the LLM spits out "the Red Sox hit 4 home runs to beat the Yankees 10-3 in game 7, capping off a monumental comeback after being down 3 games to 0 entering game 4 of the ALCS," I don't think there would be a problem (as the information from the article would be sufficiently "transformed" so as to constitute fair use).[30]  If, however, when prompted with "provide me with a copy of the first article on the front page of New York Times on October 21, 2004," the LLM spits out a verbatim copy of the article, there is a much more colorable argument that copyright infringement has occurred (especially if the New York Times's articles constituted a large plurality, much less a majority, of the training data).

Assuming many of these cases turn on what the LLM will and will not output (and that's a big assumption, as judges can be about as predictable as a random number generator), we can return to Palworld.  Assuming the Pal models were generated by an AI trained on Pokémon models, and we were to compare the complete set of Pal models to the complete set of Pokémon models, there isn't altogether that much overlap.  Only some of the Pal models show striking overlap with Pokémon models (and certainly fewer than 50% of them) and most Pokémon models have no clear Palworld analog whatsoever.  Further, even for the models that do overlap, the genAI has clearly at least tweaked some combination of color, meshing, size, shape, etc.  Given this, I would think this situation is more akin to someone listening to all of Taylor Swift's albums and writing a song in Taylor's pop style that has lyrics about the same general topics as Taylor's songs (i.e., likely fair use), rather than someone listening to The Tortured Poet's Department and releasing an album where half the songs are direct copies of those in Taylor's album (i.e., likely not fair use).

Given the above, I find it unlikely that, if a genAI were trained (at least partially) on the Pokémon models and used to create each of the Palworld models, copyright infringement has occurred.

Don't forget about trade secrets?

As just one final thought (and something I haven't seen discussed much in the Palworld discourse), regardless of whether Palworld constitutes copyright infringement, if the Palworld 3D models were based on actual Pokémon 3D models (e.g., regardless of whether they were subsequently used to train a genAI or tweaked by a human), the Pokémon Owners may have a cause of action based on trade secret protections.  For example, if Pocket Pair illegally accessed a Pokémon Owner's server to obtain the underlying 3D model and/or if a former or current employee of the Pokémon Owners improperly provided the 3D model to Pocket Pair, such actions could constitute a misappropriation of trade secrets.  Hence, Pocket Pair profiting from them would be improper (and Pocket Pair could be liable for damages).[31]  That is, of course, assuming there was no reasonable way to reverse engineer the Pokémon 3D models without improper means (e.g., without violating Pokémon terms of service).  Given the above, in addition to or instead of a copyright infringement action, the Pokémon Owners could succeed in a trade secrets suit against Palworld if Pocket Pair stole or otherwise improperly obtained the Pokémon models.

Conclusion

It's obviously debatable whether the Pokemon Owners would succeed in a suit against Pocket Pair over Palworld.  It would come down to the facts regarding whether and how the models for the Pals were generated (and, if a lack of direct evidence were present, what a jury might think of circumstantial evidence of copying).  It becomes even more convoluted if each of the three factual situations posited above occurred in conjunction with one another (rather than only one of the three occurring alone). Further, given the success of Palworld, I doubt that Pocket Pair would have the stomach for full on litigation against the Pokémon Owners (i.e., would settle before anything progressed very far).

This is 100% conjecture, but if I were a psychic Pokémon casting Future Sight, my guess would be that the Pokémon Owners will let sleeping dogs lie and decline to pursue this in any way.  The only exception, in my view, would be if the Pokémon Owners have some smoking gun regarding illicit access to their 3D models.  In such a case, I would assume the Pokémon Owners would send a cease and desist to Pocket Pair threatening litigation and that Pocket Pair would respond by removing (or heavily modifying the character models of) the Pals that were created as a direct result of the improperly obtained Pokémon models.

As Professor Oak says, this "tale of grand adventure is about to unfold."[32]

[1] https://twitter.com/Palworld_EN/status/1760682105171062876

[2] https://www.forbes.com/sites/paultassi/2024/02/06/hogwarts-legacy-barely-holds-off-tears-of-the-kingdom-for-most-sales-of-2023/?sh=46beea2f128a

[3] https://en.wikipedia.org/wiki/Palworld

[4] On that note, if you're looking for some late '90s nostalgia and some music to read by, might I recommend – https://www.youtube.com/watch?v=rg6CiPI6h2g

[5] https://gamerant.com/palworld-electric-element-pals-pokemon-counterparts-similarity-explained-grizzbolt-electabuzz/. Plenty of other articles have been written enumerating the Pals that resemble Pokémon counterparts, as well. See also, https://www.ign.com/articles/palworld-vs-pokmon-comparison-just-how-similar-are-the-designs, https://clutchpoints.com/all-the-pokemon-look-alikes-in-palworld, https://www.sportskeeda.com/mmo/5-palworld-pal-designs-similar-popular-pokemon, https://gamerant.com/palworld-pokemon-controversy-stolen-designs-every-pal-similarity-explained/.

[6] https://www.ign.com/articles/as-palworld-blows-up-could-nintendo-successfully-sue-the-controversial-pokemon-with-guns-game 

[7] https://www.thegamer.com/a-snapshot-of-nintendos-convoluted-legal-history/  

[8] In 2020, Pocket Pair released a survival game called Craftopia that had a very similar aesthetic to Nintendo's The Legend of Zelda: The Breath of the Wild (https://www.gameinformer.com/2024/02/22/what-is-palworld-the-new-pokemon-with-guns-survival-game-everyones-playing). Interestingly, Pocket Pair's own webpage for Craftopia includes the Pokemon trademarked slogan of "Gotta Catch 'Em All!" - https://www.pocketpair.jp/craftopia.

[9] https://www.ign.com/articles/the-pokemon-company-makes-an-official-statement-on-palworld-we-intend-to-investigate  

[10] https://youtu.be/b5W1OFnnj4w?si=DwxJRNFo_N8l7Q4h&t=237  

[11] If you would like to see an EU / UK lawyer's brief thoughts on the matter, see https://www.rockpapershotgun.com/does-palworld-break-pokemons-copyright-we-asked-a-lawyer   

[12] https://www.law.cornell.edu/uscode/text/17/102

[13] Atari, Inc. v. Amusement World, Inc., 547 F. Supp. 222 (D. Md. 1981); Tetris Holding, LLC v. Xio Interactive, Inc., 863 F. Supp. 2d 394 (D.N.J. 2012); see also Applying Copyright Law to Videogames: Litigation Strategies for Lawyers, John Quagliariello, Harvard Journal of Sports & Entertainment Law, Vol. 10

[14] https://mluce.ro/articles/the-yokai-roots-of-pokemon/

[15] https://bulbapedia.bulbagarden.net/wiki/Vulpix_(Pok%C3%A9mon); https://bulbapedia.bulbagarden.net/wiki/Ninetales_(Pok%C3%A9mon); https://palworld.fandom.com/wiki/Foxparks; https://palworld.fandom.com/wiki/Foxcicle; https://palworld.fandom.com/wiki/Kitsun

[16] https://en.wikipedia.org/wiki/Digimon_World; https://en.wikipedia.org/wiki/Temtem        

[17] https://www.videogameschronicle.com/news/palworld-pokemon-plagiarism-accusations-pile-up-as-ceo-responds/   

[18] https://www.pcgamer.com/palworld-accused-of-copying-pokemon-models-developer-pushes-back-says-artists-are-being-slandered/    

[19] See, e.g., Meshwerks v. Toyota Motor Sales U.S.A., Inc., 528 F.3d 1258 (10^th Circuit, 2008)   

[20] https://bulbapedia.bulbagarden.net/wiki/Miltank_(Pok%C3%A9mon)      

[21] https://www.polygon.com/24047136/palworld-pokemon-mod-release-toastedshoes   

[22] https://www.destructoid.com/nexus-mods-bans-pokemon-based-content-for-palworld/     

[23] 17 U.S.C. § 106(2)      

[24] https://www.copyright.gov/circs/circ14.pdf      

[25] https://www.rockpapershotgun.com/palworlds-skin-deep-pokemon-parody-is-part-of-a-bigger-horror-story       

[26] For an interesting rebuttal to the rockpapershotgun parody article reference above, see https://www.themarysue.com/palworld-why-a-pokemon-parody-may-be-considered-fair-use-but-satire-isnt/

[27] See, e.g., Leovy v. Google (N.D. Cal.); Nazemian v. NVIDIA Corporation (N.D. Cal.); Tremblay v. OpenAI (N.D. Cal.); Silverman v. OpenAI (N.D. Cal.); Chabon v. OpenAI (N.D. Cal.); New York Times v. Microsoft and OpenAI (S.D.N.Y.)

[28] https://www.thefashionlaw.com/from-chatgpt-to-deepfake-creating-apps-a-running-list-of-key-ai-lawsuits/

[29] Williams v. Gaye, 895 F.3d 1106 (9th Cir. 2018)

[30] This situation can be analogized to Authors Guild, Inc. v. Google, Inc. (2nd Cir. 2015), in which Google digitized copyrighted works and displayed snippets of them in response to a search, and the court found that such a use was transformative enough to constitute a fair use.

[31] 18 U.S.C. § 1836; 18 U.S.C. § 1839(5)

[32] https://gamerant.com/pokemon-professor-oak-quotes-ranked/

The USPTO's Guidance on Use of AI-based Tools in Practice

By Joshua Rich and Michael Borella --

Since President Biden issued his Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, the U.S. Patent and Trademark Office has been investigating the potential pitfalls of practitioners' use of AI in patent and trademark practice.  On April 11, the Office issued its "Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the United States Patent and Trademark Office."[1]  And while the Guidance does not include or propose any new rules, it provides useful reminders of how to ensure compliance with the rules (along with a dash of advocacy for the USPTO's positions on prosecution strategy).

The starting point for the USPTO's analysis of the effects of AI is the potential efficiencies and cost savings that use of AI-based tools can provide.  After all, the USPTO is already using such tools itself:

For example, patent examiners are performing AI-enabled prior art searches using features like More Like This Document (MLTD) and Similarity Search in the Office's Patents End-to-End (PE2E) Search tool.

The Office recognizes that practitioners are already (and increasingly) using tools to locate prior art, review patent applications, and monitor examiner behavior.  But along with the benefits of using AI-tools come risks:  hallucination, disclosure of confidential client information, and violation of export control laws and rules.  The Guidance attempts to show practitioners how they should -- and (in some cases) must -- mitigate those risks based on the existing rules and regulations.

The USPTO's guidelines rely on six different sets of rules and policies as sources for the obligations for practitioners.

• The duty of candor: "Each individual associated with the filing and prosecution of a patent application has a duty of candor and good faith in dealing with the Office, which includes a duty to disclose to the Office all information known to that individual to be material to patentability as defined in this section."[2]

• The signature requirement for nearly all submissions to the USPTO, which carries with it the implied certification that the person signing or submitting the paper was the one actually signing it and vouching under penalty of perjury that ""[a]ll statements made therein of the party's own knowledge are true, all statements made therein on information and belief are believed to be true"[3] and an obligation to make a reasonable inquiry to ensure the accuracy of those statements.[4]

• The obligation of confidentiality: "A practitioner shall not reveal information relating to the representation of a client unless the client gives informed consent, the disclosure is impliedly authorized in order to carry out the representation, the disclosure is permitted by paragraph (b) of this section, or the disclosure is required by paragraph (c) of this section."[5]

• The foreign filing license requirement and regulations governing export of technology, including ITAR (International Traffic in Arms Regulations), EAR (Export Administration Regulations), and AFAEAR (Assistance to Foreign Atomic Energy Activities Regulations).  Notably, the USPTO has previously said that "[a] foreign filing license from the USPTO does not authorize the exporting of subject matter abroad for the preparation of patent applications to be filed in the United States."[6]  Further, ITAR, EAR, and AFAEAR prohibit not only exporting certain technical data in the traditional sense -- that is, sending the information overseas -- they also prohibit allowing nationals of certain foreign countries to have access to that information, even within the United States.

• Policies regarding access to the USPTO's electronic systems, including Patent Center, P-TACTS, ESTTA, and USPTO.gov. Among other things, only individuals may have USPTO.gov accounts and must not share the account with others (including AI-based tools).

• Professional responsibility duties to clients, including the duty of competence (which includes a duty to be competent with technology used to handle client matters before the USPTO)[7] and the duty to "reasonably consult with the client about the means by which the client's objectives are to be accomplished" and "explain a matter to the extent reasonably necessary to permit the client to make informed decisions regarding the representation."[8]

Those rules and regulations apply to the use of AI in at least four different contexts of USPTO patent and trademark practice.[9]  Those contexts can involve different AI tools and considerations, but there in one unifying theme:  use of an AI tool does not relieve a practitioner of compliance with existing duties.  Indeed, the USPTO does not expect any more out of practitioners who use AI tools than it does from practitioners who use assistance of junior attorneys or paralegals -- the practitioner is responsible for reviewing work product and ensuring that it is technically and legally correct.  Notably, however, the USPTO seeks to put a thumb on the scales balancing certain duties owed by practitioners.

First, AI tools are increasingly being used in the drafting of prosecution (and PTAB-related) documents.  Even word processing software such as Microsoft Word are beginning to incorporate AI tools.  But specialized AI tools are also being rolled out that can assist in patent and claim drafting, responding to office actions, and preparing forms (such as IDSs).  While such tools are growing more and more robust and useful, they still have deficiencies that must be double-checked.  Double-checking is not only good practice, the Guidelines make it clear that it is a requirement of both the duty of candor and the signature requirement.  "Therefore, if an AI tool is used in drafting or editing a document, the party must still review its contents and ensure the paper is in accordance with the certifications being made."  A practitioner should also reasonably consult with the client to ensure that the client agrees to the means for accomplishing its goals.

Given the requirement for an attorney or agent to double-check the entirety of a submission, the Guidance does not suggest a general obligation to disclose the use of AI tools in preparing papers.  There are potential exceptions, however.  For example, practitioners often include language to broaden the disclosure of an invention using alternative embodiments and potential substitutions known in the art.  While there may be a question whether this makes the patent attorney a joint inventor, the issue is more concerning if an AI tool has come up with the alternatives.  Not only may it throw inventorship in question (and require disclosure of the AI drafting tool), it may exceed the true scope of the invention.

The Office also counsels that "Practitioners are also under a duty to refrain from filing or prosecuting patent claims that are known to be unpatentable.  Therefore, in situations where an AI tool is used to draft patent claims, the practitioner is under a duty to modify those claims as needed to present them in patentable form before submitting them to the USPTO."  The first sentence is uncontroversial:  it is improper to seek claims that you know -- not suspect, know -- to which the client is not entitled.  But the second sentence is much more opaque.  Whether an AI tool is used or not, if "patentable form" is intended to mean something more than the opposite of "known to be unpatentable" is seems to be placing a higher obligation on the review of AI-drafted claims than that imposed on the review of human-drafted claims.  That would appear to be more an Office request than an obligation.

Similarly, the Guidance points out that the obligation to review an IDS requires more than checking to see that it is in proper format.  Rather, it requires "reviewing each piece of prior art listed in the form."  But the Guidance then asserts that the review requires a practitioner to cull not only those references that are irrelevant but also those that include "marginally pertinent cumulative information."  Of course, the duty of disclosure requires the citation of material prior art and what may be "marginally pertinent" in the eyes of one may be "material" to another.  Practitioners and Examiners often disagree vehemently over the relevance of cited art.  By emphasizing the obligation to eliminate irrelevant art and extending it to "marginally pertinent" art, the Guidance seem to be trying to tip the scales away from citing all potentially relevant art.  That is, the Office seems to be trying to use the vehicle of the Guidance ease the Examiner's burden of reviewing submitted art at the cost of greater risk of violation of the duty of disclosure.  Whether an AI tool is used or not should not affect that calculus.

The issue is that, on one hand, the USPTO has set forth the duty of candor in Rule 56 and elsewhere.  But, on the other hand, the Guidance implicitly incentivizes practitioners to NOT strictly follow Rule 56.  The notion of what a "reasonable Examiner" might consider to be relevant art varies dramatically between Examiners and art units.  Further, there is no practically useful test for what constitutes cumulative art.  Given this, the burden of determining whether to disclose art that falls into either of these gray areas should not fall on practitioners.  Instead, practitioners should be encouraged to err on the side of disclosure when in doubt.

With regard to trademark filings, many of the concerns related to document preparation are the same.  The Guidance brings up one additional example of the perils of AI hallucination, specifically, submission of an AI-generated specimen of use.  Again, the AI tool must be double-checked to make sure it has provided accurate information.

Second, AI tools may be useful in the mechanical process of filing documents with the USPTO.  But the rules seem to throw a wrench into that prospect, at least as far as they currently are written.  Almost all submissions require a person's signature; the Guidance unequivocally states, "It would not be acceptable for the correspondence to have the signature of an AI tool or other non-natural person."  Thus, even if a person has prepared a document, it would be a violation of rules to have an AI tool affix a signature and submit the document.  That is especially true because a non-natural person cannot have a USPTO.gov account.  Thus, for now, AI tools cannot be used to file USPTO documents.

Third, AI tools could be used to access USPTO systems, for example, to scrape filings to include in a large language model.  Again, however, an AI tool cannot have a USPTO.gov account and certain actions violate the terms of service of USPTO websites.  As the Guidance warns, "Users should also be extremely careful when attempting to data mine information from USPTO databases.  Using computer tools, including AI systems, in a manner that generates unusually high numbers of database accesses violates the Terms of Use for USPTO websites, and users using tools in this way will be denied access to USPTO servers without notice and could be subject to applicable state criminal and civil laws."  As the Guidance points out, however, the USPTO does offer bulk access to its data for purposes of mass downloading, and those interfaces should be used instead.

Fourth, the Guidance raises the concern of potential unintended disclosure of client confidential information when using an AI tool.  As the Guidance cautions:

This can happen, for example, when aspects of an invention are input into AI systems to perform prior art searches or generate drafts of specification, claims, or responses to Office actions.  AI systems may retain the information that is entered by users.  This information can be used in a variety of ways by the owner of the AI system including using the data to further train its AI models or providing the data to third parties in breach of practitioners' confidentiality obligations to their clients under, inter alia, 37 CFR 11.106.  If confidential information is used to train AI, that confidential information or some parts of it may filter into outputs from the AI system provided to others.

This is an especially acute concern because current large language models are "black boxes" with terms of service that may change without much notice.  But the use of a tool with terms of service that prohibit use of inputs to teach the model may not be sufficient if the owner of the tool is not especially trustworthy.  The Guidance suggests that "practitioners must be especially vigilant to ensure that confidentiality of client data is maintained" when client data is used with an AI tool (but also when client data is stored on third-party storage).  The concern may extend to knowing where the model's servers are based, since disclosure to the model may violate the foreign filing license rules or export control regulations.  In short, the Guidance suggests extreme caution in maintaining the confidentiality of client data.

All in all, the Guidance is intended to reinforce that the use of AI tools does not relieve practitioners of their obligations to comply with existing rules and regulations, even if it may simplify or quicken prosecution tasks.  There are clear risks and pitfalls, and the Guidance helps highlight them without placing additional obligations on practitioners.

[1] 89 Fed. Reg. 25,609 (Apr. 11, 2024).
[2] 37 C.F.R. § 1.56.  There are analogous duties of candor and good faith in 37 C.F.R. § 1.555(a) and 37 C.F.R. § 42.11.
[3] 37 C.F.R. § 11.18(b)(1).
[4] 37 C.F.R. § 11.18(b)(2).
[5] 37 C.F.R. § 11.106(a).
[6] Scope of Foreign Filing Licenses, 73 Fed. Reg. 42,781 (July 23, 2008).
[7] 37 C.F.R. § 11.101.
[8] 37 C.F.R. § 11.104.
[9] In the Guidance, the Office included a fifth context, "Fraud and Intentional Misconduct."  But the Guidance does nothing more than reiterate the previously-stated concerns with the submission (or omission) of materials in violation of the duty of candor and use of AI tools to violate the terms of service of USPTO websites.

The New York Times Response to Microsoft's Motion to Dismiss: An LLM Is Not a VCR

By Joshua Rich --

As discussed in our earlier post, Microsoft opened its motion to dismiss portions of the New York Times's OpenAI case pled against it with an extended analogy to the Betamax case.  It argued that the Times was acting like the Motion Picture Association of America in crying wolf over the death of its industry in light of the development of a new technology.[1]  But no analogy is perfect, and the Times opened its opposition to the motion to dismiss by noting the distinctions between an LLM and a VCR.

The Times's opposition starts with an unambiguous distinction shot across the bow based on the evidence that was cited in the complaint, citing the Betamax case in support of the proposition that "A VCR is not preloaded with millions of programs that can be played back if prompted with an opening scene" while "Defendants' AI models can display works that were copied to train them with the right prompt."  It then notes that VCRs do not generate programming that competes with movies, but ChatGPT returns results that copy and summarize the Times's works (that it asserts users would otherwise obtain from the Times itself).  So the Times concludes, "If VCRs had been built with movies to make movies that compete with movies, or if Sony oversaw the VCR's infringing users, [the Betamax case] would have gone the other way."

The Times then answers Microsoft's accusation that it is standing in the way of technological progress.  The Times offers a solution that is easy to describe, if not to execute:  just teach the LLM on non-copyrighted works.  And it drops a PR bomb of its own in doing so, saying rather than doing so, "Microsoft—the richest company in the world—claims that it was entitled to take what it wanted for free."  Turning Microsoft's rhetoric back on itself, the Times argues that respecting copyright and building an LLM without infringing "will promote a future where both GenAI developers and publishers can flourish."[2]

If nothing else, the introduction serves as a valuable reminder to litigators to consider what the other side might do with an analogy you offer.  Microsoft built the logic and righteousness of its case on an analogy, but appears not to have thought about how the analogy could be turned on its head to rebut its own position.  In the end, the judge is likely to focus on the facts of ChatGPT -- not the analogy to a VCR -- and ignore the rhetorical fireworks back and forth.

When it gets to the legal arguments, the Times notes that the issues are almost identical to those raised by OpenAI (save for OpenAI's limitations argument, which Microsoft did not raise), and the Times response is essentially the same.  One difference between the Times's two responses, however, is that the Times repeatedly argued that OpenAI's case citations were inappropriate, whereas it argues that Microsoft may be correct but they relate to later stages of the case, whether on summary judgment or post-trial.  The basic argument is still the same:  the Times should be permitted to take discovery to support its claims, not cut off on a motion to dismiss.

First, the Times starts its argument against Microsoft's attempt to obtain dismissal of the contributory copyright infringement count by noting that six of the eight cases Microsoft relies on were decided later in the case.  The Times then notes that it pled averments of the elements of the claim disputed by Microsoft; while it did not cite evidence, that is generally not required in federal court pleadings.  As with its response to the OpenAI motion to dismiss, the Times does go beyond the pleadings to cite articles that would support the averments of the complaint, but the judge is not likely to require amendment of the complaint to show proof of the plausibility of the factual allegations.  Furthermore, in a situation like this where most (if not all) of the evidence of is in a defendant's hands -- especially evidence of intent -- a judge would be loathe to allow that defendant to prevail without permitting the plaintiff to investigate through discovery.

Second, the Times repeats basically the same arguments about violations of the Digital Millennium Copyright Act as it raised against OpenAI.  While Microsoft (and OpenAI) focuses on how removal of copyright management information does not facilitate third-party infringement, the Times argues the stripping of CMI during training of the LLM and synthetic search retrieval conceals Microsoft's (and OpenAI's) own infringement.  That is, the Times argues Microsoft seeks to focus on the wrong actor and act of infringement for the DMCA violations and that Microsoft's own knowledge of those violations is where the judge should look.  Again, the Times argues it has pled Microsoft's knowledge and should be allowed to take discovery to support its allegations.  As with the contributory copyright infringement claim, the court is likely to allow it to do so.

Finally, the Times addresses its "hot news" state law misappropriation claim by basically focusing on the fact that its pleadings are sufficient, not on whether it will ultimately prevail. It provides scant argument or details beyond a recitation of the allegations of the complaint and notation that Microsoft relies on cases decided at later stages of the case.  Again, as we suggested in our discussion of OpenAI's motion to dismiss, the Times's misappropriation claim is likely to survive the Microsoft motion to dismiss, but may not survive summary judgment.

[1] The first sentence of Microsoft's brief reads, "'The VCR is to the American film producer and the American public as the Boston strangler is to the woman home alone,' warned Jack Valenti, then head of the Motion Picture Association of America, in his 1982 testimony to the House of Representatives."

[2] The Times cites Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023) and A&M Recs., Inc. v. Napster, Inc., 239 F.3d 1004 (9th Cir. 2001) as support for its argument.  Neither case is really on point, which Microsoft likely will point out in its reply.

For additional information regarding this topic, please see:

• "The New York Times Calls Out OpenAI on Its Motion to Dismiss," March 18, 2024
• "Microsoft's Motion to Dismiss NY Times Lawsuit over ChatGPT: How is ChatGPT Like a VCR?" March 7, 2024
• "OpenAI's Motion to Dismiss NY Times Lawsuit over ChatGPT: Do They Want to Win or Influence Public Opinion?" March 5, 2024
• "The New York Times Case against OpenAI is Different. Here's Why." February 19, 2024
• "An Analogy for the Current Wave of AI Copyright Lawsuits," November 12, 2023

The New York Times Calls Out OpenAI on Its Motion to Dismiss

By Joshua Rich and Michael Borella --

Responding to the OpenAI brief that read more like a press release than a traditional motion to dismiss, the New York Times attacked OpenAI's approach from the very first sentence, calling the factual background of OpenAI's brief "grandstanding about issues on which it hasn't moved."  The Times echoed the concern we raised in our post on OpenAI's motion to dismiss, that the motion "introduces no fewer than 19 extrinsic documents, none of which can be properly considered on a motion to dismiss, in a submission that for nearly 10 pages reads more like spin than a legal brief."

But none of that stopped the Times from responding to some of the factual arguments or introducing new documents.  Clearly, the Times was most concerned by the public relations value of the accusation that it "hacked" OpenAI to fabricate a copyright lawsuit.  Without citation to the factual record, the Times spends much of its introduction on refuting that claim, which wasn't a basis for the legal arguments in OpenAI's motion.  Then, when it actually gets down to addressing OpenAI's arguments, it props up the contributory copyright infringement claim with a citation to a Gizmodo post and a Wired article rather than relying on the allegations of its complaint.  So while OpenAI is a far worse offender in refusing to focus on the legal issues and the facts of the complaint, the Times was not above employing a similar strategy.

When it got around to addressing the merits of the arguments in OpenAI's motion, the Times argued clearly and forcefully.  Those arguments suggest that the challenged counts of the complaint are likely to survive dismissal, even if they are vulnerable on the ultimate merits.

First, OpenAI had argued that a copyright infringement claim against the teaching of its large language model (LLM) was time-barred by the statute of limitations.  The Times raised two counters.  First, even if the statute of limitations might otherwise bar such claims for the initial training of the OpenAI model, it would not apply to the dataset used for development of more recent models of GPT 3.5 and GPT4, the LLM that underlies ChatGPT.  That is, the Times views the LLMs as different versions that each constitute a copying of the Times's articles rather than consecutive iterations of the same work.

That distinction plays into the Times's second argument -- that it cannot be charged with constructive knowledge of OpenAI's alleged infringement, especially because it would have been quite difficult to investigate prior to the launch of ChatGPT.  A copyright owner is not barred from asserting claims that arise before the statute of limitations if (1) it had not discovered the infringement, and (2) it could not have, with due diligence, discovered the infringement. Only with the launch -- and prompting -- of ChatGPT was the Times able to definitively determine that the OpenAI LLM had been trained with Times articles.  At least, that is a strong argument for avoiding dismissal at this phase of the dispute.  We will see what discovery reveals about the Times's actual or constructive knowledge down the road.

Second, OpenAI had argued that the Times's contributory copyright infringement count should be dismissed because the Times had not identified a specific direct act of infringement.  The Times started by attacking OpenAI's argument head-on, asserting that case law suggests that a plaintiff is not required to identify specific infringements to state a claim for contributory copyright infringement.  The Times next argued that public reports show that members of the public have used ChatGPT to circumvent paywalls and obtain copyrighted Times articles, hinting that discovery would show that there is a factual basis for contributory infringement.  Even more fundamentally, the Times had alerted OpenAI to the threat of contributory infringement and, according to the complaint, OpenAI chose to do nothing to stop that infringement. Each of those bases should preserve the contributory infringement claim.

Third, OpenAI had challenged the Times's Digital Millennium Copyright Act over the lack of evidence of removing copyright management information.  Among other things, the Times accuses OpenAI of removing copyright notices and links to terms of service from articles when teaching its LLM and not including them in output provided in response to user prompts.  The Times disputed the legal arguments that OpenAI made and also the factual claims about what had been removed from the Times's articles.  Again, it is likely that the Times's arguments will be persuasive and discovery will proceed.

Finally, and perhaps most tenuously, the Times argues against OpenAI's motion to dismiss its state court misappropriation claim.  The Times's arguments are likely strong enough to preserve its claim for now, but its arguments highlight just how far this situation is from the past cases in which a misappropriation claim has not been preempted by copyright law.  In those cases, a party reported "hot news" that would diminish in value if the reporting were delayed by minutes or hours.  Here, the Times is attempting to apply the same doctrinal framework to Wirecutter reviews and reports that may be delayed weeks or months from the time they appear in the newspaper or on the Times website (and, for that matter, weeks or months after the products are received for review.  Again, the Times appears to be on the better side of the law on dismissal, but likely will have a far greater challenge on summary judgment.  That is, unless it can show that expeditious reporting of such reviews is necessary for their value, the Times is likely to lose on this count.

Microsoft's Motion to Dismiss NY Times Lawsuit over ChatGPT: How is ChatGPT Like a VCR?

By Joshua Rich and Michael Borella --

Like OpenAI before it, Microsoft has sought to dismiss portions of the lawsuit the New York Times has brought against it over ChatGPT.  While raising some of the same arguments, Microsoft takes a more traditional path with its motion relying on the facts pled in the Times's complaint.  However, it also relies on an extended analogy to the VCR (and the Supreme Court's "Betamax" decision) to suggest a framework for the court's analysis.

Microsoft's brief opens with a "Preliminary Statement" about the Motion Picture Association of America's resistance to the VCR in the 1980s and the "all-out effort by television and movie producers to stop a groundbreaking new technology."  Those efforts ultimately gave rise to the Supreme Court's "Betamax" decision,[1] in which the closely-divided Court decided that Sony did not contributorily infringe copyright because the ability to "time shift" when a television program was watched constituted a non-infringing fair use of a VCR.

But Microsoft characterizes it as an attempt by existing media to throttle a disruptive technological advance in the crib, just as it accuses the Times of doing for large language models (LLMs) here.  It boldly asserts "copyright law is no more an obstacle to the LLM than it was to the VCR (or the player piano, copy machine, personal computer, internet, or search engine)," despite the key issue -- whether the use of the Times's copyrighted works to train the ChatGPT LLM is a transformative or fair use -- not being raised or argued in the motion.

Instead of the central issue of transformative use, which Microsoft says it will "show on summary judgment," citing the Authors Guild v. Google case[2] relating to digitization of printed books and Google v. Oracle case[3] regarding use of Java APIs, Microsoft's motion to dismiss raises three of the four issues raised in OpenAI's motion to dismiss.  In doing so, however, Microsoft sticks closely to the facts alleged in the Times's complaint.  That is, after the extensive Preliminary Statement related to issues that will be raised later in the case, Microsoft presents a very typical factual background and argument.

First, Microsoft argues that the Times's contributory infringement claim fails as a matter of law because the Times was not able to identify an end user that was directly infringing its copyrights.  Microsoft argues that the contributory infringement "claim is therefore the exact one the Supreme Court rejected when it blessed the VCR."  That, of course, is not true.  ChatGPT provides the end user with the data that includes copyrighted works in the LLP; in contrast, a VCR is content-agnostic.  That is, OpenAI chose the works that it used to train the LLM and provide response to end user prompts on ChatGPT, whereas Sony did not dictate what end users were taping on VCRs.  Furthermore, the VCR did not require training on a massive corpus of copyrighted works in order to even exist.

Nonetheless, Microsoft's argument may succeed because of how the Times has pled its contributory infringement claim.  Rather than arguing that Microsoft encouraged end users to obtain copyrighted works, the Times argued that Microsoft materially contributes to end user infringement by offering products that it knows are capable of distributing infringing works. The argument is a bit tenuous because the Times did not identify an infringing end user or that Microsoft knew the end user was infringing.  The Times may argue that it could not have provided that information because the information passes only between ChatGPT and the end user.  We will have to see how the Times responds to the motion.

Second, Microsoft argues that the Times failed to state a claim under the Digital Millennium Copyright Act (DMCA) because it failed to identify that Microsoft had removed any specific copyright management information (CMI) from a Times work before distributing that work.  Just as with the contributory infringement count, Microsoft argues that the Times did not identify any specific article reproduced or CMI removed.  That appears to be true at this juncture.  Again, however, this information is not easily available (if at all) to the Times.  Further, Microsoft argues that it had an honest belief that training the LLM was a non-infringing fair use.  The court may choose to allow the Times to take discovery to determine if it can find instances of Microsoft removing the CMI from works reproduced by ChatGPT, whether in training or final output.

Finally, Microsoft argues that the state law misappropriation claim is preempted by the Copyright Act.  Traditionally, courts have accorded the Copyright Act broad preemptive effect.  Unless the information is highly time-sensitive "hot news," courts have been loath to find that copyright infringement is not a sufficient (and sole) remedy.  That does not seem to be the case with ChatGPT -- it does not provide information based on Times articles immediately, only after the LLM has been taught on articles.  Thus, the Times's misappropriation count seems the most likely one to be dismissed.

Like OpenAI, Microsoft is using its motion to dismiss as more than just a vehicle to get rid of certain counts.  Instead, it is seeking to create a framework for the case analogous to the Betamax case in addition to arguing for the dismissal of three counts of the complaint.  By doing so in a traditional brief format, it may obtain better results than OpenAI.  Nonetheless, Microsoft's VCR analogy to LLMs is shaky when placed under scrutiny and seems to be a ploy to leverage Hollywood's overblown and ultimately misplaced reaction to early home recording.

[1] Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417 (1984).

[2] Authors Guild v. Google Inc., 804 F.3d 202 (2d Cir. 2015).

[3] Google LLC v. Oracle Am., Inc., 141 S.Ct. 1183 (2021).

OpenAI's Motion to Dismiss NY Times Lawsuit over ChatGPT: Do They Want to Win or Influence Public Opinion?

    By Joshua Rich and Michael Borella --

In response to the lawsuit the New York Times has filed against it, OpenAI has sought to dismiss portions of the complaint.[1]  But instead of filing a traditional motion to dismiss that argues that the allegations of the complaint are insufficient to support legal liability, OpenAI went on the offensive.  It sought to recast the case by including an extensive factual background that went far beyond -- and sometimes is incompatible with -- the allegations of the complaint.  In short, the motion to dismiss seems more like an attempt to influence the public view of the Times case, rather than an effort to narrow the scope of the dispute.

On December 27, 2023, the Times filed its lawsuit against Microsoft Corp.; OpenAI, Inc.; and related companies in the U.S. District Court for the Southern District of New York, alleging that those companies have improperly trained large language models (LLMs) on the Times's copyrighted works, and that the defendants then deployed these models for public use.  The Times's complaint includes counts of direct copyright infringement, vicarious copyright infringement, contributory copyright infringement, violations of the Digital Millennium Copyright Act (DMCA), unfair competition by misappropriation, and trademark dilution.  As we have discussed earlier, a notable aspect of the complaint is its specificity -- particularly, it provides several examples of the ChatGPT LLM being able to reproduce Times articles almost verbatim.

On February 26, 2024, OpenAI responded with a motion to dismiss the claims of contributory copyright infringement (Count IV), DMCA violations (Count V), and unfair competition (Count VI).  OpenAI also seeks to have portions of the direct and contributory copyright infringement claims that occurred outside of the statute of limitations dismissed as time-barred.

OpenAI's motion to dismiss is noteworthy for a number of reasons.  Motions to dismiss generally follow a well-trodden path of reciting the factual averments of the complaint, then arguing that those facts, even in the best of circumstances for the plaintiff, do not form the basis of an actionable claim.  That is because the court's standard in reviewing a motion to dismiss is to assume the truth of all of the facts in a well-pled complaint and all reasonable inferences drawn from those facts.  The court will then dismiss the complaint (or parts of it) only if those facts do not form a plausible basis for the claims asserted or if there is some legal bar to the claims that is apparent without any fact finding.  OpenAI did not take that approach.

Instead, OpenAI ranged far afield factually, not limiting itself to the factual averments of the complaint.  In an attempt to do so without drawing the judge's ire, it sought to invoke several of the loopholes that allow the consideration of information outside the express language of a complaint, stretching the exceptions until they would swallow the basic rules on motions to dismiss.  So in the end, the motion's factual background section itself may draw the judge's wrath for a transparent attempt to circumvent the basic rules of motions to dismiss.

Indeed, the factual background section of the motion to dismiss reads more like a press release than a legal pleading in that it introduces new factual allegations, argues that the facts alleged in the complaint are incorrect, suggests affirmative defenses, and contains an overblown self-serving assessment of the company's importance.[2]  But OpenAI actually flat-out admits the most basic premise of the complaint:  that it copied the Times's work and used it to teach the ChatGPT LLM (albeit more than three years ago). That fourteen-page exposition on the history of ChatGPT then bears little, if any, relationship to the four substantive arguments raised in the second half of the motion.

In its first substantive argument, OpenAI plainly asserts that any direct copyright infringement that "occurred more than three years ago" should be dismissed because it is time-barred under the statute of limitations set forth in 17 U.S.C. § 507(b).[3]  However, three years is not necessarily three calendar years; federal courts follow the "discovery rule" that begins the three-year clock when the plaintiff discovers, or reasonably should have discovered, the infringement as opposed to the date of actual infringement.  It is likely that the Times will respond by saying that it did not know -- and could not reasonably know -- of OpenAI's infringement until it could query ChatGPT.

OpenAI states that the Times' copyright infringement position is based on OpenAI's use of the WebText database -- which contains Times works -- to train an early version of GPT (the LLM underlying ChatGPT), and that this training took place more than three years ago.  That is really much more of a misappropriation model of wrongdoing than copyright infringement; OpenAI clearly made a copy of the Times's articles when it first copied them, but it arguably continues to do so every time it creates a derivative work based on those articles.  In addition, OpenAI does little to address why it apparently believes that the Times knew or should have known of these activities.  Further, OpenAI does not specifically address a related allegation in the complaint regarding another database containing Times works that was used to train GPT, Common Crawl.^[4]

Regardless of when OpenAI began using any of these databases, it is well understood that newer versions of GPT are likely to have been trained and retrained on the same or similar sources.  As implied by OpenAI's own website, further ingestion of the Times's works may have occurred during some or all training procedures that have led to numerous iterations of ChatGPT.  Indeed, OpenAI no longer publicly identifies its training data sources or training procedures.  Thus, it is hard to understand how the Times should reasonably have known about all of these possible copying events.  And more fundamentally, OpenAI provides the court with no guidance on how to narrow the case and tease apart the time-barred original infringement and unbarred creation of derivative works in a simple manner.

Second, OpenAI argues that the Times has failed to state a claim for contributory copyright infringement.  Such a claim requires that the plaintiff allege that the defendant knew or had reason to know of copyright infringement of another and that the defendant actually has knowingly induced, caused, or materially contributed to this infringement.  In the complaint, the Times alleged that:

Defendants knew or had reason to know of the direct infringement by end-users because Defendants undertake extensive efforts in developing, testing, and troubleshooting their LLM models and GPT-based products.  Defendants are fully aware that their GPT-based products are capable of distributing unlicensed copies or derivatives of copyrighted Times Works.

Here, OpenAI makes two points.  The first is that the Times' allegations are too generalized and contemplates only the mere possibility of infringement by OpenAI's end users.  That is, there is no concrete end-user direct infringer to which OpenAI's inducement can be tied.  Of course, information about end users' prompts and the result returned would be in OpenAI's hands (if anyone's), not the Times.  The second point, which is rather specious, is that any such contributory infringement would violate OpenAI's terms of use.  It is a curious argument because it does not deny that it may have violated the law, but only that end users may have violated their contractual obligations to OpenAI if they follow OpenAI's lead.  To the extent that OpenAI is contending that it just held the door open and its end users walked through on their own accord, such an angle is unlikely to be persuasive.

Third, in seeking to establish a DMCA violation, the Times alleged that ChatGPT removed "one or more forms of copyright-management information in each of The Times's infringed works, including: copyright notice, title and other identifying information, terms and conditions of use, and identifying numbers or symbols referring to the copyright-management information."  OpenAI responds by, among other things, stating that the Times' allegations do not include the specific copyright-management information that was included in each work, and thus paint with too broad of a brush.  Here, OpenAI has some recent authority on its side, as the judges in parallel cases of Andersen v. Stability AI Ltd. and Tremblay v. OpenAI, Inc. have set the pleading bar to require a higher level of specificity for analogous LLM infringement claims under the DMCA.  It remains to be seen how the Times will seek to distinguish those cases.

Finally, with regard to unfair competition by misappropriation, the Times essentially alleged that OpenAI's use of the Times' content results in a loss of revenue for the Times.  The argument, made in more detail in the complaint, is based on the Times's concern that integration of GPT into a search engine (here, Microsoft's Bing) will obviate the need for readers to subscribe to the Times or visit its website.  In other words, instead of referring users to the Times, the search engine will generate a result that incorporates aspects of the Times's works without giving the Times money or credit.

OpenAI's response here is more traditional than its other arguments:  that such state law claims are preempted by the Copyright Act. That argument is well-established in precedent and relies only on legal arguments (not any factual information) and is thus more likely to succeed than the other, more novel arguments.  OpenAI also addresses one of the complaint's specific example of alleged misappropriation relating to the Times' Wirecutter product recommendations (as source of revenue for the Times).  Specifically, OpenAI analogizes the Wirecutter recommendations to compilations of stock recommendations that were found not to be misappropriation in a previous case.  In a nutshell, OpenAI argues because that the Wirecutter recommendations are not being sold by OpenAI as its own work and the Times did not obtain the recommendations through "efforts akin to reporting," the allegations fail the misappropriation test.  That portion of the argument is less likely to succeed because it would force the judge to delve into the factual similarities and differences.

OpenAI's motion to dismiss seems like it is more of an attempt to convince credulous reporters[5] (and, by extension, the public) of the justice if its ways.  It refuses to accept the Times's characterization of the facts and instead provides its own history and arguments.  But those assertions are not closely tied to the legal arguments OpenAI makes.  So at best, OpenAI is seeking to give the judge a roadmap to what he will see later; at worst, it is a cynical ploy to win over hearts and minds outside of court.  The legal arguments in the motion to dismiss, in contrast, are written in a more traditional manner but have varying likelihoods of success.  We will know more soon, when the Times responds. 

[1] Microsoft has now also moved to dismiss portions of the Complaint, with a far more traditional approach.  We will review that motion in a forthcoming post.

[2] For example, "OpenAI was founded in 2015 to advance digital intelligence in the way that is most likely to benefit humanity as a whole," and "The new GPT-3 model was by far the most powerful language model ever created."  While the GPT family is indeed a very powerful set of large language models, this assertion should be taken with more than a lost shaker of salt.  A standard response employed for years by high tech companies accused of IP theft amounts to "Don't stop us bro, we're innovating."  This response is often made as those companies are raking in huge profits or setting up an environment in which they can do so.

[3] Notably, Microsoft does not make this argument.

[4] OpenAI contends that prior to the three-year period, "it was common knowledge that WebText2 and Common Crawl included numerous articles published by the Times."  That assertion is not drawn from the complaint.

[5] To be fair, few reporters are trained in the law well enough to distinguish between the standard for summary judgment and dismissal. Even some lawyers get it wrong.

USPTO Issues Guidance about Inventorship of AI-assisted Inventions

By Aaron Gin and Michael Borella --

On February 12, 2024, the U.S. Patent and Trademark Office released detailed guidance regarding inventorship of inventions created with the assistance of artificial intelligence (AI).  The guidance, signed by Kathi Vidal, Under Secretary of Commerce for Intellectual Property and Director of the USPTO, is scheduled to be published in the Federal Register on February 13, 2024.  The guidance focuses on how inventorship should be determined in U.S. patents and patent applications when AI, including generative AI, plays a role in the conception and/or reduction to practice of a given invention.

The 27-page document demonstrates the Patent Office's effort to navigate the challenges AI poses to traditional notions of inventorship and creativity, and comes in response to AI-specific requests for public comment, public listening sessions, and President Joe Biden's "Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence."

At the heart of the USPTO's guidance is the principle that AI systems, even when they have a role in facilitating or contributing to the creation of inventions, cannot themselves be named as inventors on patent applications.  Instead, only natural persons—humans—are capable of inventing under U.S. law, consistent with Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022).  Thaler makes clear that "the AI system's contribution is not inventorship, even if the AI system's contributions were instrumental in the creation of the invention."  Id.

Accordingly, the guidance firmly states that inventors as well as joint inventors named on U.S. patents and patent applications must be natural persons.  Specifically, the guidance provides that "based on the holding in Thaler that an 'individual' must mean a natural person, it is clear that a 'joint inventor' or 'coinventor' must also be a natural person."  Furthermore, patent applications that name a machine (e.g., an AI) on an Application Data Sheet, an inventor's oath or declaration, or a substitute statement as either an inventor or a joint inventor will be considered to have improper inventorship.

Interestingly, the guidance acknowledges that AI systems could perform acts that, if performed by a human, could constitute inventorship under existing U.S. patent law.  Furthermore, in the scenario where an AI could otherwise be considered an inventor, the inability to list the AI as an inventor or joint inventor does not inherently render the invention unpatentable.

Detailed principles and examples are provided to assist in determining when a human's contribution to an AI-assisted invention is significant enough to warrant inventorship.  Namely, the guidance reiterates the factors of Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998), which include that each inventor must:

(1) contribute in some significant manner to the conception of the invention;

(2) make a contribution to the claimed invention that is not insignificant in quality, when that contribution is measured against the dimension of the full invention; and

(3) do more than merely explain to the real inventors well-known concepts and/or the current state of the art.

The guidance also delves into the public policy of the U.S. patent system.  Specifically, it outlines the importance of a human contribution in the innovation process, because "the patent system is designed to encourage human ingenuity," citing the IP Clause of the Constitution and the legislative history of the 1952 Patent Act ("Inventions eligible for patenting 'include anything under the sun made by man.'") Graham v. John Deere Co., 383 U.S. 1, 9 (1966).

The USPTO highlights some "guiding principles" in determining AI-assisted inventorship, which include:

(1) A natural person's use of an AI system in creating an AI-assisted invention does not negate the person's contributions as an inventor.  Rather, the natural person can be listed as the inventor or joint inventor if the natural person contributes significantly to the AI-assisted invention;

(2) Merely recognizing a problem or having a general goal or research plan to pursue does not rise to the level of conception.  A natural person who only presents a problem to an AI system may not be a proper inventor or joint inventor of an invention identified from the output of the AI system.  However, a significant contribution could be shown by the way the person constructs the prompt in view of a specific problem to elicit a particular solution from the AI system;

(3) Reducing an invention to practice alone is not a significant contribution that rises to the level of inventorship.  Therefore, a natural person who merely recognizes and appreciates the output of an AI system as an invention, particularly when the properties and utility of the output are apparent to those of ordinary skill, is not necessarily an inventor.  However, a person who takes the output of an AI system and makes a significant contribution to the output to create an invention may be a proper inventor.  Alternatively, in certain situations, a person who conducts a successful experiment using the AI system's output could demonstrate that the person provided a significant contribution to the invention even if that person is unable to establish conception until the invention has been reduced to practice;

(4) A natural person who develops an essential building block from which the claimed invention is derived may be considered to have provided a significant contribution to the conception of the claimed invention even though the person was not present for or a participant in each activity that led to the conception of the claimed invention.  In some situations, the natural person(s) who designs, builds, or trains an AI system in view of a specific problem to elicit a particular solution could be an inventor, where the designing, building, or training of the AI system is a significant contribution to the invention created with the AI system; and

(5) Maintaining "intellectual domination" over an AI system does not, on its own, make a person an inventor of any inventions created through the use of the AI system.  Therefore, a person simply owning or overseeing an AI system that is used in the creation of an invention, without providing a significant contribution to the conception of the invention, does not make that person an inventor.

The guidance applies to utility patents, plant patents, design patents, and corresponding applications.

The guidance also states that the duty of disclosure and the duty of reasonable inquiry both apply to the question of proper inventorship.  This will incentivize inventors and their representatives to be forthcoming about the contributions that humans and/or AI systems may have had with respect to the conception of inventions.  Moreover, the guidance notes that examiners and other USPTO employees may make requests for information regarding inventorship when such an individual "has a reasonable basis to conclude that one or more named inventors may not have contributed significantly to the claimed subject matter."

Nonetheless, further rounds of USPTO guidance may be released "as appropriate."  These may discuss the impact of AI on other aspects of patent practice, including subject matter eligibility, obviousness, and enablement.

By emphasizing human contribution and providing specific principles and examples, the guidance gives the clearest picture yet available from the USPTO regarding its policies and practices at the intersection between AI and inventorship.

For additional information, please see:

• Unpublished Version of USPTO Inventorship Guidance
• Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence

USPTO Contributes Patent and Trademark Data to Help Launch National AI Research Resource Pilot Program

By Aaron Gin --

This week, the U.S. National Science Foundation (NSF) alongside public and private sector collaborators launched the National Artificial Intelligence Research Resource (NAIRR) pilot program.  NAIRR seeks to advance AI discovery and innovation by making resources available from 11 different federal agencies and over 25 private companies.  The pilot program was formed as a direct reaction to President Biden's 2023 AI Executive Order (specifically section 5.2(a)(i)) and targets several of the EO's broad topics including "Ensuring Responsible and Effective Government Use of AI" and "Promoting Innovation and Competition."

With an estimated six-year budget of $2.6 billion, NAIRR seeks to provide researchers with access to AI-relevant data and compute resources, which are being made available by agencies including the NSF, Defense Advanced Research Projects Agency (DARPA), National Aeronautics and Space Administration (NASA), National Institutes of Health (NIH), and National Institute of Standards and Technology (NIST) to name a few.

The contribution of the U.S. Patent and Trademark Office (USPTO) is expected to include provision of its "bulk datasets," which include U.S. patents and trademarks in a structured format.  Such data offers a "rich and diverse" dataset of complex language, which may be used for the development of robust language models.  The USPTO is also making available "research datasets," which include patent and trademark information relating more generally to "intellectual property, entrepreneurship, and innovation topics."  The datasets and related documentation is specifically formatted for academic researchers.  The patent and trademark information is provided for NAIRR users in a variety of different formats, including the official gazette, multi-page PDFs, full text/XML version, and others.

Kathi Vidal, Under Secretary of Commerce for Intellectual Property and Director, USPTO commented with respect to the NAIRR Pilot: "The USPTO is committed to advancing US innovation including AI innovation and AI-assisted innovation.  We are focused on maximizing AI's benefits and distributing them broadly across society, confining the risks through technical mitigations and human governance and empowering a diversity of individuals and organizations to participate in AI innovation."

Broadly speaking, the information and resources provided by the USPTO through the NAIRR pilot program may help democratize access to AI resources and assist researchers in areas related to intellectual property, data governance, and patent data.

For additional information, please see:

• NSF NAIRR Pilot News Release
• NAIRR Pilot