Last year, the U.S. Patent and Trademark Office (USPTO) issued a request for comments (RFC) on patenting artificial intelligence (AI) based inventions.  Topics of the RFC included AI's impact on inventorship and ownership, patent eligibility, disclosure, prior art, and the level of ordinary skill in the art.  The USPTO received 99 comments from a variety of stakeholders.  In parallel to that effort, the USPTO also released a second RFC related to the impact of AI on copyright, trademark, database protections, and trade secret law.  Stakeholders provided a similar number of comments.

With this feedback in hand, the USPTO collated it into a 56-page report.  This article focuses on issues related to the drafting and prosecution of AI-related patent applications.  Readers are encouraged to read the rest of the report to obtain the USPTO's summary of the remaining issues.

The report begins with a high level overview that can be expressed as three main points:

1.  Most commenters agreed that AI-based inventions are a subset of computer-implemented software inventions, and thus the current framework for examining such inventions is suitable for examining AI-based inventions.  Still, some commenters exhibited concern that it may be difficult to meet the requirements of 35 U.S.C. § 112(a) for certain AI-based inventions.

2.  Most commenters agreed that AI may lead to changes in how the standard of a person of ordinary skill in the art (POSITA) is interpreted.

3.  There were some concerns that AI would lead to a proliferation of prior art making it difficult for examiners to conduct searches for relevant prior art.

A brief list of issues of note follows.

Definitions

Interestingly, there was no clear consensus on what constitutes an AI-based invention.  Of the definitions provided, the one that seems to be the most helpful (in my subjective opinion) states that:

AI inventions can be categorized (in no particular order) as follows:

(a) inventions that embody an advance in the field of AI (e.g., a new neural network structure of an improved machine learning (ML) model or algorithm)

(b) inventions that apply AI (to a field other than AI)

(c) inventions that may be produced by AI itself.

Patent Eligibility

As noted, the consensus is that, as a special form of computer-implemented invention, AI-based inventions can be fairly evaluated under the current patent-eligibility of 35 U.S.C. § 101 as interpreted by the courts.  Some commenters indicated that AI-based inventions could be characterized as falling into the abstract idea subject matter exception, as a method of organizing human activity, a mental process, or a mathematical concept.  But to the extent that these inventions provide technological improvements and "amount to significantly more than the abstract idea," they would be found to pass through the § 101 filter.

Indeed, this approach is consistent with the Supreme Court's refusal to draw a line between what is and is not eligible.  Each invention is to be considered on its own merits, and there are no "magic words" or drafting techniques that can be used to guarantee a claim is eligible.

Written Description and Enablement

While many commenters did not believe that AI-based inventions had any special written description requirements, one commenter stated that "AI inventions can be difficult to fully disclose because even though the input and output may be known by the inventor, the logic in between is in some respects unknown."  This is true, in the sense that a learning model can be viewed as a "black box" that receives input (e.g., words or images) and produces output (e.g., a classification).  One can view the constituent elements that make up the model (e.g., nodes, connections, and weights), but these elements can number in the millions.  As a consequence, they cannot be used to concisely characterize the invention, nor should they as the specific values of weights can be different each time the model is trained.

Regarding enablement, the USPTO notes that "the amount of guidance or direction needed in the specification to enable the invention is inversely related to the amount of knowledge in the state of the art, as well as the predictability in the art."  But there is no clear understanding of the predictability of AI systems.  Some contain an inherent amount of intentional randomness that allow these systems to produce superior results when compared to more deterministic systems.  Others may be predictable if their training data and training technique is fixed.

One possible way of addressing both of these concerns is to allow AI models that involve machine learning to meet the written description and enablement requirements with proper specification of how the models can be trained.  If, from the specification, a POSITA can glean enough knowledge to reasonably reproduce the training technique with their own data, the invention is likely to be well-described and enabled.

Level of Ordinary Skill in the Art

While the ubiquity of AI may impact the level of ordinary skill in the art, commenters seems to agree that the present legal framework is "adequate to determine the impact of AI-based tools in a given field."  As the level of skill in the art generally increases over time, use of AI-based tools would be just another way for the level of skill to rise, albeit at a potentially faster rate than what we have seen in the past.

Prior Art Considerations

While most commenters believed that there are no specific concerns regarding prior art and AI-based inventions, some indicated that AI itself may be used to generate massive amounts of prior art that would be challenging to thoroughly search.  Others pointed out that prior art unique to AI may eventually exist as AI systems evolve toward more general intelligence.  A key to these considerations would be proper examiner training.

Data Protection

A critical component of AI goes beyond the algorithms and models, and is embodied in the data sets that the algorithms use to train the models.  Currently, such data sets have relatively weak protection (mostly in the realm of trade secret law).  Some commenters would like to have provisions that allow incumbent companies to maintain proprietary rights over data sets that they have collected, while still allowing newcomers to use these data sets to train their own models.

Examination

There were suggestions that the USPTO coordinate with the world's other patent offices when considering its approach to AI-based inventions.  Of note was that the Japan Patent Office and Korean Intellectual Property Office have established specific and dedicated AI examination units.  Further, the Intellectual Property Office of Singapore has "reportedly created an expedited examination path for AI technologies."