The US Patent Office (USPTO) recently issued new guidance and three examples for AI-related patent claims, which indicate that claims applying AI to a process are unlikely to render the process patent-eligible at the USPTO under 35. U.S.C. § 101. This briefing:

• Summarizes the key aspects of the USPTO’s new AI guidance and examples.
• Summarizes the varying allowance rates of USPTO art units that examine AI-related patent claims.
• Provides strategic considerations for drafting and prosecuting AI-related patent applications.

It is important to note that while data itself is not patentable under 35. U.S.C. § 101, the use and generation of AI data can be, but the USPTO’s latest examples indicate that uses of AI in patent claims will be treated in much the same way that the USPTO treats software. Patent claims that apply an AI model to a process that on its own would not be patent eligible are unlikely to cause the claims to be eligible. In the new examples, the integrations of AI into claimed processes are considered abstract ideas, and the example claims that are deemed eligible contain additional non-AI elements that contribute to eligibility.

For instance, claim 2 of Example 47 is deemed non-patentable under the new guidance, and refers to an anomaly detection method that includes the training of a neural network that results in a trained backpropagation algorithm and a gradient descent algorithm used to detect anomalies in a data set. Example 47 claim 2 considers the trained neural network as a mental process akin to mathematical calculation performed by a generic computer, and thus an abstract idea (therefore unpatentable under 35. U.S.C. § 101). Example 47 claim 2 also is not considered as integrated into a practical application because the training of the neural network is considered as being performed by a generic computer.

In contrast, claims that are considered eligible in the USPTO’s new examples are eligible because they recite specific hardware features for running an AI model or highlight functions not performed by AI. For example, eligibility in claim 3 of Example 47 hinges on the detection of a packet’s source address and dropping the packet in real time. In Example 48, claim 2 is deemed eligible due to the recitation of synthesizing and combining speech waveforms, which are actions that do not require AI in the claim. Claim 3 in Example 48 is considered eligible in part because a deep neural network’s embeddings allow for subsequent steps not performed by the AI to improve over previous techniques. Thus, AI-related functions contributing to steps that represent technical improvements may be considered eligible, but not solely because of the AI's performance.

The USPTO’s approach shows a tendency to focus on the broader process improved by AI, rather than the implementation of the AI itself. This emphasis is important given the variation in allowance rates across different USPTO Art Units and Technology Centers that examine AI-related patent claims. For example, Art Units 3620, 3680, and 3690 all examine AI inventions and have three-year grant rates below 40%. Conversely, Technology Center 2100 for computer architecture and software and Technology Center 2400 for networking, multiplexing, cable, and security have three-year grant rates above 75%, despite also dealing with AI applications.

Of course, until the USPTO applies this guidance to actual applications, no one can guarantee specific outcomes. Based on recent examinations, decisions, and the new USPTO guidance, below is a list of considerations that may be relevant for drafting and prosecuting AI-related patent applications:

• Emphasize specific technical improvements attributed to AI innovation in the specification.
• Consider whether the use of certain language may cause the application to be assigned to USPTO Technology Centers with lower allowance rates.
• Clearly articulate improvements to the AI itself over mere application of AI to enhance that process.

[View source.]