The inventors of the screwdriver would have had no trouble describing their invention in a patent application. They had a bit, a shaft, and a handle—and it’s not hard to show how those fit together. Their patent application would have handily met the statutory “enablement” requirement to explain how to make the invention, under 35 USC §112(a), in exchange for patent rights. The same goes for inventors building better mousetraps: they are easy to understand and easy to describe. It is harder to meet the enablement requirement for a new machine tool you have invented, and harder still for a jet engine or a complicated computer program. But for artificial-intelligence (AI) inventions, such as the Netflix movie-recommendation system, Section 112 may seem like an insurmountable barrier. How do you explain how to make a system that learns and grows as it operates? Yet this is what patent attorneys and agents in the AI field must do to meet growing client demand.
AI is a prominent and expanding research area. Tesla, Google, and many other groups are working feverishly on self-driving cars. Apple’s Siri, Microsoft’s Cortana, and other digital assistants gain new capabilities seemingly every day. International Data Corporation (IDC) forecasted last year that the AI industry will take in more than $57 billion in 2021. As research progresses, so do efforts to patent the resulting inventions—to the tune of more than 9,000 US patent applications in AI-related areas published in the last three years alone. Clients develop AI solutions to particular problems and want patents that protect those solutions.
But AI solutions are unlike screwdrivers or mousetraps. A screwdriver, once made, doesn’t grow or change. By contrast, after inventors make an AI, they have to train it to solve a particular problem. The AI ends up being made partly by human construction, and partly by learning about that problem. To support meaningful patent coverage, therefore, a patent application must explain not only how to make the untrained AI (a proverbial blank slate) but also how to make that initial AI into all the myriad machines the AI could learn to be. Your clients’ AIs will learn one way, and their competitors’ AIs will learn a different way. To block those competitors, a patent should cover as many of the possibilities as it can. How can you write one patent application that describes all the options without giving the application the length of a Russian novel?
The law helps you fill the gap between the blank slate, on the one hand, and the constructed AI problem-solvers you claim in a patent, on the other. Section 112 permits you to rely on the knowledge of a person who is knowledgeable in the relevant technology area. This person, or artisan, knows that AI machines have to be trained. Therefore, you can claim the fully trained AIs as long as you tell the artisan reading your patent not only how to set up the blank slate but also how to help the AI learn. As long as the artisan does not have to conduct excessive experimentation, you can meet the enablement requirement. Therefore, ask your inventors for details about how the training is conducted. Incorporating their answers into your patent applications will help you meet the requirement to describe how to make the problem-solving AI that your client is seeking to protect.
These answers can also strengthen patentability arguments beyond enablement. Under Alice v. CLS Bank, software-related patent claims may be required to recite a “new and useful application” of an algorithm to be patent-eligible under 35 USC §101. Reciting not only the blank slate (algorithms and data structures) but also the trained AIs (practical uses), provides additional ways to argue for patent-eligibility of your clients’ AI inventions. Including technical details of the training can reduce the chance that your application will be considered as a pure business method under Alice, rather than as a technological development.
As with any patent application, make sure you know the problem the invention solves, and comprehensively understand the solution. This way, your application can tell the story of a meaningful innovation. Once you know that story, you are ready to ask the inventors about details. A good starting point is to find out which type of blank-slate AI the invention uses, of the wide range of AI technologies. For example, neural networks imitate the brain in simple ways, regressors model represent complicated mathematical relationships, and decision forests break down patterns into combinations of simple choices. The remaining detail questions will be influenced by the type of blank slate.
Then, ask questions about the training process. Almost all present-day AI systems rely on external training data the AI cannot obtain for itself. Ask questions to connect the dots between the sources of training data and the final AI. Example questions include:
- What types of data are used, and where do they come from? Do the data include human judgments or not?
- How are the data preprocessed for training?
- What training algorithm is used to change the blank slate into the final AI?
- Is the AI trained only once, or is it retrained periodically?
These are only some of the many questions we have asked inventors to get details about the untrained AI and the training process. The relevance of each question depends on the situation. Studying artificial intelligence and machine learning will give you the background to decide which questions are pertinent to which inventions. Asking the right questions and including the answers in the patent applications you draft will help your clients benefit from their inventions. And clearly teaching the public how to apply artificial intelligence to solve practical problems will help spur technological progress and improve quality of life for all of us.
All opinions are the authors’ and not necessarily those of Lee & Hayes, Purdue University, Indiana University, IUPUI, or any other party.