With the recent explosion in popularity of generative artificial intelligence (AI), an increasing number of companies are integrating AI into their products and services. This includes companies designing new AI models, training known AI models on new data and integrating AI models into well-established products and industries.
We often are asked how best to protect such AI innovations and whether AI innovations are patentable. While general software-related patent guidance applies to AI inventions (and AI inventions can indeed be eligible for patent protection), there are a number of AI-specific factors worth considering when deciding whether and how to pursue patent protection for AI inventions. This article discusses considerations under US intellectual property (IP) law, although many of the general principles will be relevant for other jurisdictions.
1. What aspects of my AI are patentable?
Your application of AI likely involves one or more features that are tailored or customized, for example, to achieve a desired prediction accuracy. Examples of such features can include arrangements of layers or nodes, activation functions, loss functions, training frameworks, data cleansing techniques, methods for defining feature vectors and/or methods of using hardware to execute an AI model more efficiently. You may have spent considerable time and effort developing these advancements to produce a well-performing model, and this time and effort can be an indicator that you also have produced patentable subject matter. Even small changes or improvements to an AI technique may be patentable. For example, while large language models have recently become more popular and well known, you may be improving upon these models with bespoke tokenization techniques, self-attention mechanisms or transformer structures. No matter how incremental you believe your contributions to be, it may be worthwhile to consider protecting those efforts.
2. Is patenting the best way to protect my AI?
Before committing to an IP strategy, it makes sense to first consider and weigh the available avenues for protecting your AI. Depending on how the AI will be commercialized and where the novelty of the invention lies, the best approach may or may not include patent protection.
Trade secret protection may be a good alternative to patent protection if your AI model is not publicly identifiable via your commercial product or service – e.g., the AI model is part of a software as a service (SaaS) offering or is used internally and is not reverse-engineerable, you don’t plan to publish the details, and/or the AI model can otherwise be kept secret (for general guidance on trade secrets, see this Cooley GO article).
To obtain patent protection, you will be required to publicly disclose how to make and use your AI model, which may be undesirable if your AI model is otherwise not publicly identifiable. Moreover, if your AI model is not publicly identifiable, a competitor’s infringing model also may not be publicly identifiable – making detection of infringement, and thus enforcement of a patent, difficult. If, on the other hand, the model is publicly identifiable, trade secret protection is not possible, and patent protection should be considered. However, trade secret protection will not allow you to prevent independent development or reverse engineering by others. Thus, if you believe competitors are likely to develop similar technology, or you believe a competitor could reverse engineer your solution, patent protection may be a better option.
As another example, if the novelty of your AI model lies in the data (e.g., the invention involves training a well-known model on new data), patent protection may be difficult to obtain. Specifically, while data can be maintained as a trade secret, “data per se” without any structural components is not eligible for patenting. If, however, at least a portion of the novelty of your AI model lies in its architecture or structure (e.g., the AI model is a newly developed model or includes changes and/or improvements relative to the architecture or structure of a known model, etc.), then those characteristics of your AI model may be suitable candidates for a patent application.
3. Am I ready to pursue patent protection for my AI?
Timing is a strategic consideration in the patenting process. On the one hand, it is advisable to file a patent application as early as possible after the invention is conceived to win the “race to the patent office” and secure as early a priority date as possible. On the other hand, you should wait until your invention is ready for patenting to ensure that you can satisfy the various patentability (enablement, written description) and patent-eligibility requirements. In any event, it can be beneficial to file for patent protection before selling, offering to sell or publicly disclosing aspects of your AI technology, as these activities can bar your ability to obtain patent protection, depending on the circumstances and jurisdiction(s) of interest.
When deciding when to move forward with a patent application for your AI invention, keep in mind that your AI model need not yet be fully developed, validated or perfected. A patent application is, itself, a constructive reduction to practice. You should, however, be able to describe one or more technological improvements that are achieved by your AI model – or that you expect will be achieved by using your AI model. The more detail you can provide on how these improvements are achieved (even if aspirational/planned), the better positioned you are to file. For example, consider whether you can articulate – in concrete terms – the transformations that your AI model applies to generate your outputs from your inputs. Said another way, the explainability of your model may be proportional to the claim scope you can expect to achieve.
As you develop your AI technology, ensuring good communication among your engineering, IP and marketing teams can be critical to preventing premature public disclosure of technological breakthroughs – and to properly timing your patent filings. Also, having good internal documentation processes in place can facilitate timely invention capture and patent application filings. If an invention has been conceived and will be further developed within a year, it may be advisable to initially file a provisional patent application. In the US, a provisional patent application will give you one year to further refine the invention before filing a nonprovisional patent application that adds details derived from any intervening development activities.
4. What details should I include in my AI patent application?
Your patent application should describe the technical details of your AI model – such as, for example, data used, implemented pipelines, model infrastructure, training framework and/or any combination thereof. Although your AI model’s field of use can be important context to include in your application, a discussion of the structure used to implement your AI model can often better distinguish your AI model from prior art models. The more structural details of your model that you provide in the patent application, the more flexibility you will have to evidence the novelty and nonobviousness of your invention during examination.
Describing how the unique parts of your AI model are tied to the output(s) your AI model generates can strengthen your patent application. For example, rather than listing the components of your AI model in isolation, your patent application can include a description of how data flows through and interacts with those components to generate an output. You also may want to consider describing the computer hardware with which your AI model interacts during operation. For example, if you describe a novel loss function, you also can explain how computer resources – such as a processor and memory – can use that loss function to perform an optimization. Including these details can provide flexibility during examination and help ensure that a person reading your patent application can implement your invention(s) without undue experimentation, fulfilling the enablement requirement.
Including a description of alternative architectures, use cases and scenarios also can strengthen your patent application. For example, if other models, normalization algorithms, activation functions or feature vectors may be embodied by the invention and used by competitor systems, it can be beneficial to describe these in your patent application. This can increase the potential scope of protection you may obtain with your patent application, reducing the opportunity for a competitor to design around your protection.
5. In which jurisdictions can I obtain patent protection for my AI?
As noted above, this article focuses on US IP law. While many of the general principles apply elsewhere, be mindful that the relevant rules, requirements and timelines in other jurisdictions differ from those of the US. Thus, when preparing a patent application for your AI, you should consider the jurisdictions where you would like to obtain protection, and tailor your patent application for success in those jurisdictions. Your patent lawyer can work with you to ensure your patent application meets the requirements for the jurisdictions you are interested in.
Last reviewed: October 28, 2023