Patenting an artificial intelligence (AI) invention can be a challenge that depends greatly on jurisdiction and the category of AI to which the invention belongs. Important jurisdictions for patent protection include Europe and China, and, of course, the United States. There are two important categories of AI pertinent to patentability that are termed herein as “technically-applied AI” and “generally-applicable AI”.
Technically-applied AI involves applying AI to other technology areas. AI inventions in this category include, for example, autonomous vehicles and robots, facial recognition, and AI-controlled medical devices. Technically-applied AI inventions are generally patentable in Europe, China, and the U.S.
Generally-applicable AI includes AI innovations that may be applied to any technology area. Obvious inclusions in this category are improvements to fundamental AI and machine learning algorithms, such as algorithms for neural networks and random forests. Patenting inventions directed only to these fundamental algorithms may be possible in the U.S. and also, shortly, in China, but is impossible in Europe except in rare circumstances. In fact, pundits most often cite improvements to such fundamental algorithms as the example of generally-applicable AI that is virtually unpatentable in Europe.
However, much innovation in generally-applicable AI occurs outside of improvements to fundamental algorithms. An example of such generally-applicable AI is a category termed herein as “meta AI”. Innovations that are meta AI apply AI to AI. Meta AI inventions improve and automate the engineering of AI solutions to a broad, if not unlimited, range of technical applications. Examples of meta AI include applying machine learning to selection of a machine learning algorithm for a particular kind of data set, applying machine learning to feature engineering (to automate extraction of features from raw data for training), and hyperparameter value optimization (to automate exploration for optimal parameter values that control learning). Meta AI not only delivers AI models that produce higher quality results, but that can also be executed more efficiently with less consumption of computer resources.
Meta AI is similar to technically-applied AI in that both can involve applying existing AI and machine learning algorithms to a technology area. The difference is that technically-applied AI is applied to improving other technologies, while meta AI is applied to AI methods. Initial experience with meta AI in Europe was encouraging. However, recent experience reveals that Europe treats meta AI as it treats improvements to fundamental algorithms.
A brief description of how Europe, China, and the U.S. treat generally-applicable AI follows. Given the treatment of any form of generally-applicable AI inventions by Europe, patent applicants for most generally-applicable AI inventions should forgo pursuing patents in Europe until the law changes. However, they may apply for patents in the U.S., and based on recent developments, consider applying in China.
In November 2018, the European Patent Office (EPO) set forth the practice for examining AI-related inventions by revising the section on the patentability of mathematical methods in the EPO’s Examination Guidelines. The new guidelines explicitly treat AI methods as mathematical methods. As a result, the guidelines establish two main hurdles to overcome under existing law for mathematical methods, which are termed herein as the eligibility hurdle and the inventiveness hurdle.
Eligibility concerns subject matter eligibility, and is met when a technical means is used. Thus, the eligibility hurdle is easily overcome by citing the use of a computer or the like.
The inventiveness hurdle is more problematic. Inventiveness generally concerns prior art issues. As stated above, AI related features are treated in Europe like mathematical methods. In the EPO, AI related features in patent claims are not considered “technical” by themselves, and as such, are ignored when the claims are analyzed for inventive step, unless the AI related features contribute to technical character of the claims.
The 2018 Guidelines did establish two safe harbors covering ways that AI features can contribute to the technical character of a claim. These are an AI feature that is applied to a field of technology (EPO does not consider AI a technology) and an AI feature that is adapted to a particular technical implementation, as further discussed below. Thus, inventive step analysis can account for an AI-related feature when it is applied to a field of technology or adapted to a specific technical implementation.
With respect to application to a field of technology, the 2018 Guidelines enumerate many example technical applications, which include: “digital audio, image or video … analysis…,” “speech recognition…,” and "optimising load distribution in a computer network….”
With respect to being adapted to a particular technical implementation, the 2018 Guidelines give one example: “The adaptation of a polynomial reduction algorithm to exploit word-size shifts matched to the word size of the computer hardware.”
Most generally-applicable AI inventions will not fall within these safe harbors. This is true even for generally-applicable AI inventions that enable algorithmic efficiencies, which most kinds of generally-applicable AI inventions do. Algorithmic efficiencies is a term used herein to refer to improvements realized by a software-implemented algorithm over other algorithms that perform the same or an equivalent function, where the improvements use generic computer resources more efficiently. Such improvements do not rely on a specific adaption to hardware, such as adaption to a GPU. Examples of algorithmic efficiencies include reaching inferences more quickly with less processor usage or generating features that require less storage but that retain inferencing accuracy.
With meta AI inventions, the resulting algorithmic efficiencies can be even more significant than inventions related to other forms of generally-applicable AI. In the EPO, algorithmic efficiencies may underpin inventiveness in other areas of software technology. Unfortunately, the EPO has purposely disregarded algorithmic efficiencies for all generally-applicable AI.
To address patentability of AI and other emerging technologies, the Chinese Patent Office (CPO) updated its Patent Examination Guidelines in February 2020. As the EPO’s 2018 guidelines, the CPO’s 2020 guidelines bar patenting generally-applicable AI, except that in China both the hurdles of eligibility and inventiveness serve as bars. In China, directing a claim to a computer that executes an AI feature avoids potential classification of the claim as an ineligible mental activity. However, an eligible invention must also be a technical solution to a technical problem, and AI-based features are not considered to be technical in and of themselves. As such, without an application to a technological area other than AI, a generally-applicable AI invention is not a technical solution and is considered ineligible.
Concerning inventiveness, the CPO’s new guidelines follow an equivalent approach to that of the EPO. An AI feature must be applied to another technology area other than AI to be accounted for in an inventiveness analysis.
However, in November 2020, the CPO proposed new guidelines that are promising for eligibility and inventiveness of generally-applicable AI. Under the proposed guidelines, a step or means fulfilled by a computer automatically qualifies for eligibility.
The proposed guidelines also establish that algorithmic features that lead to an improvement of a computer's internal performance are pertinent to inventiveness. The CPO’s proposed guidelines do not prevent consideration of algorithmic efficiencies in an inventiveness analysis for AI. In contrast, as explained above, Europe disregards algorithmic efficiencies of AI features as pertinent to inventiveness.
In the U.S., algorithmic efficiency is generally pertinent to the patentability of software inventions, and it is particularly important to eligibility. With the adoption of the pertinence of algorithmic efficiencies to inventiveness, China’s treatment of generally-applicable AI could become very similar to that of the U.S.
The U.S. is more favorable to the patentability of generally-applicable AI, and China appears promising. However, applicants should refrain from filing an application directed to only generally-applicable AI in Europe at this time, unless the safe harbor of being adapted to a particular technical implementation applies. It is expected that very few generally-applicable AI inventions fall within this safe harbor.
In the U.S. at least, we recommend explaining how algorithmic efficiencies are achieved by the claimed inventions in application specifications. A general assertion of algorithmic efficiency may not be sufficient. An application specification should explain, in detail, how steps and operations pertinent to key claim limitations achieve algorithmic efficiency for specific computer resources.
Algorithmic improvements that are not directly related to efficient use of computer resources may not by themselves be sufficient for patentability, even in the U.S. For example, an algorithm may improve prediction quality, but this improvement by itself may be insufficient. While such improvements should be touted in application specifications, they should not be exclusively relied upon for patentability, where possible.
Disclaimer: This article is purely a public resource of general information that is intended, but not guaranteed, to be correct and complete. It is not intended to be a source of solicitation or legal advice and is for informational and entertainment purposes only. The information is not intended to create, and receipt does not constitute, an attorney-client relationship. The laws of different jurisdictions may be implicated, and facts and circumstances can vary widely. Therefore, the reader should not rely or act upon any information in this article, but should instead seek legal counsel for individualized legal advice. For more information, please contact a firm attorney through www.hickmanbecker.com.