By Michael Borella --
In a recent article, we argued that patent law's current subjective and inconsistently-applied mental process doctrine is erroneously based on a false equivalence between human and machine cognition. Notably, inventions that would require machine implementation to be of any practical use at all are mistakenly barred from patentability because they are allegedly directed to "mental" processes. This mental process doctrine has been growing and expanding over the last 50 years despite having at best a poor grounding in both fact and law.
In the meantime, the talk of the Internet these days is ChatGPT, a large language model that is capable of producing remarkably human-like text. Trained on a corpus of millions of human language documents, ChatGPT can generate impressive results from a simple prompt.
For example, I asked ChatGPT to "write a rap song about trickle-down economics invoking Adam Smith and Ronald Reagan." It came back with the following (from which two verses and the chorus are excerpted):
Adam Smith and Ronald Reagan,
Two icons of the trickle down scene.
They preached the gospel of supply and demand,
And said the economy would be freed from government's hand.
The theory goes like this: the rich invest,
And the benefits of their success will be expressed.
Through higher wages and more jobs, it's a win-win,
But some say the reality is much more grim.
It's the trickle down, the trickle down,
Where the wealth flows from the top to the ground.
It's the trickle down, the trickle down,
Where the rich get richer and the poor get poorer, it's the trickle down.
Not bad. ChatGPT can also explain the difference between a patent and a trademark:
A patent is a form of intellectual property that gives the person who holds the patent the exclusive right to prevent others from making, using, selling, and importing an invention for a limited period of time, usually 20 years from the filing date of the patent application. Patents are granted for new, useful, and non-obvious inventions or discoveries.
A trademark is a word, phrase, symbol, or design, or a combination of these, that identifies and distinguishes the source of the goods or services of one party from those of others. A trademark can be a brand name, slogan, logo, or other distinctive mark that is used in connection with a product or service. Trademarks are used to protect brand identity and ensure that consumers can easily identify and distinguish one company's products or services from those of another.
In summary, a patent protects an invention, while a trademark protects a brand or logo.
Also not bad. In addition, ChatGPT can write text in the style of a particular author, essays on virtually any topic, and even computer code.[1] It can also answer questions. Thus, the text that it generates encompasses the functionality of search engines and Wikipedia, but adds a new aspect -- it can create. That is where things get sticky.
Not only are ChatGPT creations cogent enough to make school teachers reach for their aspirin bottles, they may suggest to naïve users that ChatGPT is intelligent, wise, and human-like. While ChatGPT would not pass the Turing Test in its current incarnation, it is good enough to fool the easily fooled.
And those, naïve, easily fooled types might attempt to use ChatGPT's impressive performance on some tasks to justify patent law's mental process doctrine. After all, ChatGPT output sure seems to be based on something at least similar to human cognition.
But even if it quacks like a duck that does not mean that it flies south for the winter. ChatGPT's internal operation is little like that of a human brain, even if you limit your analysis of the human brain to language.
In short, ChatGPT is autocomplete on steroids. Underlying ChatGPT is a transformer-based large language model that is trained by a massive amount of text to predict the next word of a sentence given the first n words of that sentence. In order to get it to respond to queries, a second model is trained, one in which thousands of prompts and associated responses are generated (some automatically, some manually), and human reviewers rate the quality of each response with respect to its prompt. A further model is trained to predict how well a human would rate each response. Then, ChatGPT is trained to generate responses that would be highly-rated by a human.[2]
This does not resemble our current understanding of human cognition. For example, as far as we know human beings use cognitive models to represent real-world objects and their behavior. Based on these models, humans can make predictions of what these objects would do in new situations. Thus, a child might (hopefully) be able to infer that a stove hot enough to boil water will be painful to touch.
In contrast, modern computer language models struggle making such predictions. This is a general problem in artificial intelligence -- unless a model is specifically programmed or trained with data representing a scenario, its prediction of what may occur or what it should do in that scenario may be way off the mark.
For an example of this, ask ChatGPT to write a review of a well-known movie, such as Star Wars or Titanic. The result will be impressive. Then ask it to write a review of a lesser known, newer movie. Prepare for disappointment. A number of comical and disturbing ChatGPT fails have been noted.
Further, human beings have a theory of mind, in that we assign mental states to other persons in order to understand their behavior. ChatGPT does not do this and therefore cannot modulate its output based on the emotions exhibited by a user. Despite the fact that ChatGPT can provide you with a definition of a person, it does not actually understand what a person is and it lacks the ability to emote. It might be able to simulate various emotions but it does not have the ability to have emotions of its own or to exhibit true empathy.[3]
Moreover, ChatGPT does not know whether it is generating text that represents the truth. It has no concept of "truth" and no way of verifying that what it is saying is accurate, much less a moral compass to guide its actions. And it has been known to generate very convincing falsehoods.
So don't ask ChatGPT for psychological help, dating advice, or who to vote for. And if you do, take its output with a huge dose of skepticism. ChatGPT creations are merely sophisticated pastiches of the writing on which it was trained.
What this all leads to is the inevitable conclusion that ChatGPT's often remarkable language skills are not anything like a human mental process. Therefore, using the existence of ChatGPT to justify mental process doctrine in patent law is disingenuous at best. So let's cut off that avenue of inquiry before anyone foolishly decides to venture down its path.
[1] On the other hand, I asked ChatGPT to "write an article about Alice v CLS Bank in the style of Kurt Vonnegut," and the result was rather bland with mild sarcasm sandwiching a textbook description of the case. So it goes.
[2] See https://pub.towardsai.net/chatgpt-how-does-it-work-internally-e0b3e23601a1 for a more detailed technical description.
[3] Here, I am not trying to contend that there is something magical about human beings, just that AI models still have a long way to go in their ability to simulate human intelligence and may continue doing so in a way that is quite distinct from human cognition.
Wow, I actually agree with a lot here. That's a nice change of pace.
Like you said, ChatGPT and most other contemporary machine learning just reflects the well-known and timeless principle of statistics that having larger quantities of data, and the data being of higher quality, leads to better predictions.
The ubiquity of miniaturized, affordable sensors, plus so much of human activity being transacted online/digitally provides the abundant, high-quality data at low cost. Not to mention that computing performance keeps improving, while storage gets cheaper and more plentiful, so of course ML can take advantage of that.
Again, as you point out, ChatGPT etc. are only as good their input data, which is usually either provided by humans in the first place anyway, or derived along the way from something originally human-generated. So under the hood, it's actually just people doing the grunt work. (Cue the Soylent Green references!) Moreover, that approach bears little resemblance to real human cognition. One of the hallmarks of human reasoning (as I see it at least) is the ability to generalize from very small data sets. That's something quite lacking in ML/AI.
Similarly, another massive problem in the field is that of domain-specific approaches. As good as ChatGPT ever gets at its task, it'll never be able to one day also drive a car. You need a completely separate approach for that. Whereas, the very same humans who compose rap songs and publish essays can easily hop in the car and drive to the store whenever they need groceries.
Last, you rightly note ML/AI models' inability to recognize or comprehend the actual "semantics" of whatever task they're performing. That's another pronounced shortcoming. In the end, no matter how good ML/AI gets at replicating the performance of various discrete human tasks—and even if it does so in ways that are superficially similar to human cognition, like relying on synapses, which is just one small aspect of a brain's structure and functionality—that will hardly mean ML/AI has actually become "human" or sentient. And how could it be otherwise? After all, we're still quite far from having a comprehensive model of actual human cognition itself. You can't emulate something when you don't fully comprehend its mechanism in the first place.
However, taking all the above to manufacture an argument against mental process law seems like a total non sequitur. The doctrine's whole point isn't that a process has to be accomplished by a computer and a human in an identical way. Rather, it's just the basic yes-or-no question of whether in fact the process could be performed by a human (and/or with a physical aid)—after you properly disregard the presence of any generic computers or computer/networking constructs as mere window dressing. And that serves a laudable purpose of preventing the monopolization of business processes and other purely human activities via the simple expedient of artful claim drafting.
Posted by: kotodamas | January 14, 2023 at 09:05 AM
Kotodamas, your last sentence prompts a thought about setting a boundary between what is eligible or patenting and what is not. Should it be everything under the sun except X? Or should it be nothing under the sun except X?
I mean, every patent is a restraint of trade and we want to minimse restraints of trade, right? The only patent that is justified is the one whose restraint of trade is more than compensated for by its promotion of the progress of useful arts. The only ones that can do that are (pace GATT-TRIPS) the ones with claims that are confined in their reach to a non-obvious solution (with new technology) of a problem in technology that defeats the efforts of the univentive skilled person. So, govenments seeking to promote the progress should refuse every patent application except those, right?
Mind you, if and when somebody does invent how to get a computer to "think", they will have solved (inventively and with new technology) a stupendous problem in technology, right?
Posted by: Max Drei | January 14, 2023 at 05:35 PM
I chuckled out loud at this:
"But even if it quacks like a duck that does not mean that it flies south for the winter."
Inside joke a bit, but devastatingly on point.
Posted by: skeptical | January 16, 2023 at 05:32 PM