Today, I spoke at the United States Patent and Trademark Office (USPTO) to discuss the impact of Artificial Intelligence (AI) on prior art and the Person Having Ordinary Skill In The Art (PHOSITA) standard. The session brought stakeholders in AI technology together across disciplines to discuss the request for comments issued by the USPTO earlier this year. For those who have expressed interest, my comments on the matter were as follows:
I would like to extend my sincerest thanks to the USPTO for selecting me to speak with all of you about issues that uniquely impact the artificial intelligence community as well as my ability as a patent attorney to meet their needs.
With regard to the specific issues put forth for public comment, with the seven minutes I’ve been given, I will respectfully urge the patent office to avoid tampering with the examination standards surrounding the identification of prior art, the level of ordinary skill, and the legal standard for obviousness. Instead, I hope to persuade the Office to continue dedicating focused effort around enhancing subject matter eligibility guidance for inventors of AI technologies, and I will propose some actionable areas for doing so.
Among the questions presented for public comment, the Office asks whether non-human publications should be treated differently than human publications for prior art purposes. The answer to that should be no. Imposing a barrier on AI authors would cause the Office to misidentify the state of the art at the time of the invention. That’s because whether a disclosure qualifies as prior art is a question that concerns the public; not the author. The question is: “Did the public have access to the disclosure prior to the invention?” The question is NOT “Was the disclosure written by an ‘appropriate’ author?”
The nature of the author bears no relation to what the public can access and is therefore irrelevant to a disclosure’s prior art status. Further, it’s not difficult to imagine how barring a technical contribution to the art based on the nature of the author might be a dangerous precedent that could be wrongfully yet readily extended to harm researchers in controversial fields. I would therefore caution against imposing any test that would bar a contribution to the art based on the nature of the author.
The Office also asks, how does the availability of AI affect what is considered to be ordinary skill? The appropriate answer is that it doesn’t. If you were to ask Justice Kennedy, he might remind you that “a person of ordinary skill is… not an automaton,” as was held in the KSR decision. If you consider an AI to be an automaton then we can stop the inquiry right there.
But if we dig a bit deeper, it’s premature to view AI as having meaningfully impacted the level of ordinary skill. Current AI systems are not generally producing better research papers than human experts, nor are they hallucinating more cogent nonsense than what we can already find on social media. Humans are already producing content across the gamut from sheer brilliance to utter foolishness to blatant deception. To advocate for a change to how we determine ordinary skill is to imply that AI contributions are inadequately represented by the efforts of the human beings that train them, code them, and prompt them for output. Given that humans and their work product are what motivate AI systems to behave in the way that they do, we need not complicate the ordinary skill inquiry. We can simply consider AI systems to be part of the body of work that evidences the skill of the human beings involved in their creation and use.
The notion that AI systems should impact how we determine obviousness seems to imply some concern that AI systems have their OWN motivation to combine references that we should consider. However, a machine with its own motivation, such as an artificial general intelligence, is still science fiction. We do not need to concern ourselves with “what would AI do?” The most functional and useful AI solutions reliably do what humans have trained them to do. Therefore, I would urge the Office to continue evaluating obviousness and the level of ordinary skill with respect to those humans.
If we are looking to make an impact on the community of AI stakeholders, we should instead focus on helping inventors identify whether an AI invention is likely to be considered eligible for patent so they can make well-informed decisions about where to invest their limited capital. While the Patent Office has recently issued some enhanced guidance on the matter, there remains a significant risk that AI inventions will be summarily rejected by examiners as amounting to nothing more than mental steps implemented by a computer.
To be fair, that kind of rejection is often times entirely appropriate. There is no shortage of Applicants that will try to claim nothing more than an AI trained to make judgments that would otherwise be made by humans and leave it at that. Those are not examples that we need to address. We already know that simply using a computer as a substitute for a human is not eligible for patent.
Instead, I would encourage the Office to continue enhancing its guidance on the eligibility of some of the most common and critical functional components in which AI advancements are being made. To that goal, I propose focus on the following five AI subdomains, which I submit are examples of practical applications of AI technology that readily lend themselves to technical solutions that clearly improve the functioning of a computer and are therefore eligible:
- Unconventional AI model training techniques;
- Techniques that enable AI to make inferences and conclusions that are not reasonably discernable by humans;
- Adapting AI problem solving techniques to address specific technical problems;
- Techniques for applying feature extraction to unstructured data for use in model inference or prediction; and
- Improvements to how an AI interprets or produces modes of human communication, such as natural language.
To conclude, I would caution the Office against treating AI technology as something that requires some fundamental adaption to traditional examination procedures, as there is always a suitable human that we can apply well-established, predictable, and reliable procedures toward instead. The public would be better served by mitigating speculation around whether an AI invention would be considered patent eligible, as that would go a long way toward supporting investment in this burgeoning field.