Artificial intelligence’s “black box” decision-making presents challenges for AI and machine learning innovators who want to file for patents.
Artificial intelligence’s “black box” decision-making presents challenges for AI and machine learning innovators who want to file for patents. Practitioners need to consider the unique properties of AI technology to secure meaningful and enforceable patent protection for these inventions.
The problem: “Black box” AI algorithms
Machine learning relies on training a piece of software to make decisions by providing feedback on the output it produces while processing a set of training data. The programmers create the initial structure of the software and define the feedback heuristics used to train it, but the software produced by the training process is often a jumble of weights and interconnections between nodes in a neural network or some similarly human-illegible chunk of math. Thus, while AI systems created via ML often exhibit highly effective decision-making, they do so without providing their creators (or anyone else) any meaningful insight as to the underlying logic of the system.
Patents and artificial intelligence
A patent is a bargain with the public. An inventor applying for a patent is required to disclose the nature of the invention, including instructions for building it, to the public — in exchange, he or she is granted a limited monopoly, usually for a period of 20 years. The public is granted access to the workings of the invention in perpetuity thereafter.
In the context of AI and ML, inventors are faced with hard decisions about how to satisfy this disclosure requirement. What constitutes proper disclosure about the workings of a “black box” AI invention when the inventor herself does not always understand how it works?
Further, how much detail is needed to describe the intricate (and commercially sensitive) training protocols by which a mature ML system is created?
A national patent office will examine the invention through the lens of public policy, including policies against the monopolization of abstract ideas: How can the inventor overcome the issues outlined above and describe the workings of the algorithm so as to satisfy the patent examiners — and the courts — that inventors are entitled to a patent for their inventions?
Considerations for patenting an ML system
One fundamental question that requires consideration is what parts of the technology are being claimed as the invention? An inventor will need to make choices about how much they want a patent to focus on protecting the processes by which the system is created, trained and validated, as opposed to focusing on the final product deployed after these processes have run their course.
Another issue that requires consideration is the question of subject matter eligibility. To qualify as proper subject matter for a patent in these jurisdictions, an AI-based technology generally needs to be something more than an “abstract idea.” The patent office needs to be convinced that a specific “technical problem” is being overcome by the invention in order to have the patent issued. (See Alice Corp. v. CLS Bank International). This hurdle is easier to overcome if the application includes claims directed to specific implementations of the technology, such as:
- Specific hardware (e.g. sensors, remote devices, autonomous vehicle controls, processor architectures);
- Specific details about the training data or how that data is processed by the system;
- Specific data structures implementing an AI or ML system (e.g. a neural network);
- Specific heuristics being used for decision-making and/or training feedback; and
- Technical improvements to the functioning of a computer.
With regard to AI and ML inventions, the United States Patent and Trademark Office and U.S. courts have provided some guidance to patent eligibility: as one example, a patent application for an ML system that simply describes possible specific data structure implementations (e.g. neural networks) may be rejected by the USPTO (See Ex Parte Kirshenbaum, Board of Patent Appeals and Interferences, Appeal 2007-3223), whereas an application that restricts the claimed monopoly to such a specific implementation may be allowed (See Ex parte Bramlett, Patent Trial and Appeal Board, Appeal 2015-002707).
By including technical implementation details of the AI or ML system in the patent claims and specification, the inventor should be able to overcome common pitfalls. However, there are trade-offs involved in this choice. Specifically, if patent claims limit an invention to a specific implementation, the patent may be difficult to enforce: first, because a patent owner may not be able to detect when a competitor is using an invention as part of their product or service and, second, because it may be easy for a competitor to design around the specific implementation claimed in the patent by employing an alternative implementation.
These considerations of enforceability take practitioners outside of the purely technical realm and require practitioners to consider the details of the market in which the business operates. Before drafting the claims, a patent agent should be fully briefed on these market factors, including:
- Who the competitors are;
- Which features of the new technology competitors are most likely to infringe;
- How such infringement could be detected; and
- Which other market actors a company would (or could) sue in the event of infringement.
An effective patent strategy to protect AI-related inventions should take into account all of the issues of patentability and enforceability set out above, tailored to the realities and needs of the business.
Practitioners need to consider the unique properties of AI technology to secure meaningful and enforceable patent protection for these inventions.
Stacy Rush is an associate in the Toronto office of Ridout & Maybee LLP – a boutique intellectual property firm. Stacy’s practice focuses on assisting with the preparation and prosecution of patent applications in the areas of software and computer-related devices, gaming, and medical device technologies.
Matt Norwood is an associate in the Toronto office of Ridout & Maybee and assisted in the writing of this column. Matt’s focus is directed towards litigation of intellectual property disputes and advising clients on IP licensing matters. Matt also assists with the drafting and prosecution of patent applications with an emphasis on electrical and computer related technology.