Artificial intelligence (AI), once just a topic of science fiction, has been breaking one milestone after another and finding its way into new and different fields of use. AI systems now not only can drive a car and scan x-rays for cancer, but also create abstract art, and beat you at poker.
Perhaps unsurprisingly, AI technology’s increased use has also been accompanied by an explosion in AI-related patent application filings and the United States Patent and Trademark Office (USPTO) is now calling for comments on how to treat AI-related patent applications.
The Request for Comments (RFC) includes twelve questions on knotty topics such as:
- Who (or what) can qualify as an inventor;
- Who owns AI-created inventions; and
- How does a patent applicant comply with existing statutory requirements for patentability for AI-based inventions when those statutory requirements were largely developed with human creators in mind.
Key Questions Surrounding Inventorship/Ownership Raised by the Request for Comments
Inventorship is important in U.S. patent law. Beyond just who gets named on the face of a patent, inventorship can determine who owns the patent. Currently, U.S. patent law recognizes only humans as capable of being inventors. The RFC asks the community’s input on how to define inventorship for AI-based inventions and inventorship for inventions discovered by AI.
In the United States, any individual contributing to the conception or reduction to practice of an invention claimed in a patent application is deemed to be an inventor. When a new AI-based invention is created, inventorship could arguably reside in:
- The computer scientist who conceives of and implements a machine learning algorithm used by the AI;
- The data scientist who carefully selects, prunes, and labels training data used to train a model for the AI;
- The engineer who identifies a real-world application in which the model can be used; and
- Any combination thereof.
More problematic, however, is the treatment of inventions discovered either wholly or in part by the artificial intelligence. Since inventorship resides only in humans, the USPTO is ill equipped to handle the ramifications when an AI discovers a new malaria drug or designs a new, easier-to-handle container.
While attributing inventorship of an AI-generated invention to the AI’s human creator may seem like a reasonable solution, in some cases the human’s actions and the realization of the invention might be so far separated in time and intervening events, and proximate cause questions, that it can be hard to square naming the human as an inventor.
What might happen if a human puts into place an AI system that runs for years and in ways that might not have been anticipated, that AI system – maybe even after the human has forgotten about it or passed away – outputs something that if done directly by a human would unquestionably be an invention? If the human was an independent inventor or employed as an inventor, and a different entity owns the AI system, who should own the patent rights?
It is good the USPTO is asking these questions now, and it will be interesting to see what comments the office receives. One important consideration of inventors and companies doing research is that of certainty. Investments in inventive research and patent filing efforts tend to favor predictability as to the boundaries of what is patentable and who owns the patents.
Intersection of Black Box Concerns and Written Description/Enablement Among Several Other Important Questions USPTO Will Explore
Additionally, the USPTO has posited several questions regarding how existing statutory requirements apply to artificial intelligence inventions. This category of questions presents fewer problematic issues than inventorship and ownership. The USPTO’s interest includes comments directed towards: how existing rules might apply in AI-contexts, such as:
- The written description (to show that the inventors had possession of the invention at the time of filing) and enablement requirement (to teach how to make and use the invention);
- Obviousness; and
- Subject matter eligibility.
Written Description and Enablement
With respect to written description and enablement, AI-based inventions may make use of models or networks that evolve as part of a training process that is probabilistic in nature, making reproducibility challenging or impossible. How can a patent application convey how to make and use the invention, when it might be impossible to describe how the AI system created the invention?
Obviousness is a barrier to patenting – a claim must be novel and nonobvious to be allowed and those are typically tested using the “person of ordinary skill in the relevant art” standards. The USPTO is interested in comments whether the “person of ordinary skill in the relevant art” should reflect the capabilities possessed by AI when considering a combination of existing inventions would have been “obvious.”
In many fields for many problems, humans are still better than AI systems. But when an AI system can easily solve problems that humans cannot, do we discount nonobvious human inventions because an AI system would have found it to be obvious?
Subject Matter Eligibility
And finally, whether there are any patent eligibility considerations unique to AI inventions – for example, which parts of an AI invention should be eligible for patent protection (the training process, the machine-learning model itself, the use of the trained model in a real world application, etc.) Ultimately, though, existing statutory rules and guidelines are well-equipped for dealing with the forthcoming influx of AI-related patents.
As with other types of software patents, AI-related inventions merely claiming a desired outcome or simply performing routine steps of collecting, analyzing, and presenting data are likely not patentable, and that does not change whether invented by humans or AI. Conversely, inventions robustly described in the specification and claimed as a particular solution integrated into a real world application are more likely to be patentable.
That said, there is still a tricky middle-ground where AI is involved.