AI Inventorship: Problem or Much Ado About Nothing?
New USPTO guidance on AI inventorship opens a can of worms, but in some ways it also closes it
The year was 1994. Stephen Thaler, just a decade fresh off of getting his PhD in physics, shocked the world when he unveiled his “Creativity Machine.” Thaler, who holds the first neural network patent, filed a patent for his invention, which was granted in 1998. He would not be the last. In 2005, a patent would be granted for an invention generated independently by another program called the Invention Machine. Since then, there have been very few AI-invented patents, and typically the machine is not disclosed as an inventor. Thaler himself has continued to fight for his right to be able to file patents on inventions created by his new system, called DABUS. He has tried, and failed, to pursue rulings allowing for AI patent inventorship at the United States Supreme Court, the Federal Circuit, the United Kingdom Supreme Court, and the European Patent Office. And just this week, USPTO put out guidance on the Federal Register drawing on the “significant contribution” test.
Last year, the debate over AI inventorship started to heat up alongside debates about copyrightability for machines. Although it may seem obscure, failing to accurately represent all inventors on a patent will result in invalidation in most jurisdictions and it cannot be corrected later. It is generally quite difficult to prove defective disclosure of inventorship, so the possibility of doing so in the case of AI is both high stakes and tantalizing to litigators. AI already looks poised to transform other areas of law; perhaps inventorship is on the docket too. But is it much ado about nothing? This week in Nonobvious, we are diving into the issues surrounding inventorship with AI and how they might affect an inventor near you.
It’s Significant, But What’s a Contribution?
The key test for a co-inventor is the “significant contribution” test from Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998). The test has three factors, where a co-inventor must: “(1) contribute in some significant manner to the conception or reduction to practice 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.” Under HIP, Inc. v. Hormell Foods Corp., 66 F. 4th 1346 (Fed. Cir. 2023), without a significant contribution “measured against the dimension of the full invention,” there is no joint inventorship. The new USPTO guidance tries to get around the question of whether the machine has made a significant contribution by cleverly only requiring that at least one human make a significant contribution, and that this contribution be the “conception” of the invention.
The line of tool-assisted invention is quite blurry. Humans have used computational methods to assist in the act of inventing for decades now. But examples have gone far beyond just using Microsoft Word, especially in the life sciences. At a 2023 hearing, Corey Salsberg of Novartis noted that it regularly uses software for drug “screening” where it filters out candidates. Machine learning models from MIT have even proposed a new antibiotic. Self-driving cars test themselves in simulated worlds where they drive simulated miles, resulting in new inventions that are patented. The official position of BIO, a large association of life sciences organizations, is that AI is finding increasing use but that is a tool to aid personal creativity. But if a company just becomes a human-powered machine that tests inventions that come from software, is the patent system still incentivizing innovation?1
The problem for that type of model is that in IP law, protectability comes down to creativity. The TSR nonobviousness standard requires an “inventive step,” rather than the “sweat of brow.” Copyright rejected the same theory with Feist. This tends to be the rule across jurisdictions (though the EU makes somewhat of an exception for copyright for databases under Databases Directive 96/9/EC). So although it may take substantial work to verify the value of an invention that is AI assisted, that work is likely not viable as a reason to extend inventorship. This is not how invention has worked historically, however. No one would deny that Thomas Edison should have been able to patent the lightbulb, yet if when asked how he invented electric light his answer was basically trial and error. Similarly, there are no real cases regarding problems of inventorship stemming from the mere use of technology, such as the use of CAD software for mechanical devices or of machine-learning powered drug screening, even though for both cases those technologies are essential for the modern practice of those fields and involve a significant degree of autonomy from the inventor themselves. This makes it even more difficult to parse, especially since the USPTO guidance allows examiners to consider extrinsic evidence that a claim did not have at least one human inventor, which typically is not done for other type of inventive aids, like technical manuals.
Despite some of the court rulings (or denials of cert), the matter is not settled and is still live, with a spirited academic debate. One article in 2020 put forth the position that the debate was not nuanced enough, comparing machine learning to other computational methods and arguing that all machine learning methods are not truly agentic, requiring human input and thus more resembling tools. Yet others, like Robert Plotkin, have written that the question is a nonsensical red herring that distracts from how the patent system should best incentivize inventors. Some have gone even farther, positing that machines should be able to be patent inventors. A 2017 article published in the American Bar Association’s Landslide magazine argued that machines should be able to count as inventors. There is even an organization, the Artificial Inventor Project, that is exploring this topic. In contrast, some have argued that that these would “wreak havoc” on the patent system and cut against the incentives of the patent system. Clearly, there is a risk that this would result in an overrun patent office, which introduces its own policy problems.
Despite this tension, practitioners have gone out of their way to note for inventors that this is different than the ability to file claims for AI-implemented patents. There is no inherent bar against filing patent for AI matter, and applications of AI are easier to patent than AI inventions themselves. One could imagine providing proof of which software systems were used and their contributions, but USPTO does not require a bill of goods for other tools that are used while inventing, including other software. Perhaps the line comes down to how much the inventor directed or designed the AI system. In the case of the discovery of halicin, MIT researchers created a computer program that operated in a specific domain and functioned as a combination suggestion-screening tool; one imagines this would not introduce problems. Perhaps this is different than just using ChatGPT. The USPTO guidance says that the “construction” of the prompt may be enough, meaning that prompt engineering may be the difference between inventorship or not.
That said, the historical purpose of patents is to incentivize innovation.2 In one article, Erica Fraser argued that the ratio of human to machine input should be the core consideration, but this opens up the question of what types of input we care about—what if the core creativity came from the machine?—and risks creating a line-drawing exercise. But it is a good start for thinking about the matter. And in any case, the USPTO guidance seems to be that as long as a human was involved in one of the steps of the process in an intellectual or creative way, for now, AI-aided inventors are safe for now.
Weekly Novelties
In a bombshell decision, the Federal Circuit held that communications with an intended audience is a public disclosure for purposes of establishing prior art in Weber v. Provisur Technologies (Federal Circuit)
USPTO AI Inventorship guidelines were published. It’s the topic of this week’s newsletter, but it’s important enough to merit its own news item too (USPTO)
Cloudflare fought yet another NPE to a successful jury trial, hailing its success against “another patent troll” (Ars Technica)
An interesting patent landscape for quantum computing (IP Watchdog)
The EPO rejected an appeal from Illumina, resulting in the cancelation of two key DNA sequencing patents in Europe (Juve)
The World Trade Organization did not reach a consensus on a patent pool for COVID-19 treatments and diagnostics, protecting key pharmaceutical inventions from forced licensure (WTO)
Relatedly, the WTO published (and then deleted) a “love letter” poem to patents (Common Dreams)
It is worth noting that USPTO specifically says that just owning a model does not count as a creative input. That does beg the question: what if it is a custom-designed model? Is the act of creating that model a significant contribution? And if so, does tuning a foundational model count?
Vidal mentions in the USPTO guidance that the purpose is to incentivize “human” ingenuity, but this is not abundantly clear, especially given the importance commercial interests are given in the American system and the ability to assign patents to companies.