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.