The United States Patent and Trademark Office (USPTO) has issued a new guidance document that is intended to help identify when a process or system that uses artificial intelligence (AI) tools may be eligible for patenting.

The new guidance adds three AI-specific examples to the USPTO’s set of subject matter eligibility (SME) examples, which are intended to help patent examiners and inventors understand whether the USPTO will consider certain inventions to be eligible for patenting.

The USPTO’s previous SME examples include many computer-related inventions. For computer-related inventions (such as those that use AI), the SME examples typically address whether specific claimed inventions would be classified as an ineligible “abstract idea,” or as an invention that is eligible for patenting (if also novel and non-obvious).

Only one previous SME example was AI-specific. That example (SME example 39) addresses a method for training an AI model. The USPTO’s new guidance adds examples involving using an AI model.

The AI-specific examples discuss patent claims that are directed to:

• Using a neural network to detect anomalies in data sets or network traffic (SME example 47)
• Using a neural network to analyze speech signals and separate foreground speech from background speech or noise (SME example 48)
• Using an AI model to help personalize a medical treatment plan to individual characteristics of a particular patient (SME example 49)

SME example 47 includes three claims. The first claim covers a specific hardware configuration implementing the AI model and is classified as eligible. The other two claims, both method claims, provide an informative comparison:

• The claim classified as ineligible is directed to a method that describes, in general terms, receiving data, training a neural network on the data and then using the trained network to detect anomalies in a new data set. The USPTO described this claim as broadly covering steps that are mere “mental choices or evaluations,” recited at a “high level of generality.”

• The claim classified as eligible is directed to a method that uses a trained neural network to detect an anomaly in network traffic. The claim also recites an action — specifically, dropping malicious network packets in real time and blocking future traffic from the source address of the anomalous traffic.

Key point: The guidance notes that the eligible claim of example 47 describes a mental process, but the guidance considers the claim to be eligible because the additional steps requiring action provide a “practical application” that improves the technical field of network intrusion detection and security.

SME example 48 includes three claims that draw similar distinctions between claims that recite a “high level of generality,” and claims that recite a practical application and reflect a specific technical improvement:

• The first claim recites “using a deep neural network (DNN) to determine embedding vectors V.” While this may seem reasonably specific, the guidance points out that the claim omits any details as to how the DNN solves a technical problem. Thus, an examiner is likely to treat that claim as invoking a generic DNN merely as a tool for applying an abstract mathematical concept on a computer. Thus, the guidance did not find this claim to be eligible for patenting.

• The second and third claims attempt to make the first claim patent eligible by adding several limitations. The guidance treats one of those limitations (“transmitting the mixed speech signal x' for storage to a remote location”) as insignificant extra-solution activity — a nominal addition to the claim that does not meaningfully limit the claim. Nevertheless, the guidance categorizes the second claim and the third claim as patent eligible. That is, the guidance reasons that other limitations recited in these claims integrate any abstract mathematical concepts into a practical application.

To come to this conclusion, the guidance first notes that the patent’s disclosure describes a technical problem that is not sufficiently addressed by present solutions, and that the invention offers an improvement over those other solutions. The guidance further reasons that the claims reflect this technical improvement by reciting details of the solution. Thus, the claim reflects the improvement discussed in the disclosure.

Key point: This example highlights the importance of drafting the specification to describe the technical problem to be solved, the shortcomings of present solutions to the problem and the improvement offered by the invention — in reasonable detail — and to reflect those details in the claims.

In SME example 49:

• Claim 1 covers a medical treatment method that includes generating a genotype dataset for a patient, identifying the patient as being at high risk of a condition based on a score that is generated by an AI model from the dataset, and administering an “appropriate treatment” to the patient. Although the claim included additional detail about the model and the dataset, the guidance considers this claim to include a “law of nature” in that it describes a “naturally occurring relationship” between a patient’s genotype and risk of having a medical condition. The guidance also considered the claim to include a “mental process” in that its generation of a score was a mathematical calculation. Thus, the guidance classified this claim as ineligible.

• The guidance did not consider the step of “administering an appropriate treatment” to add anything significant to claim 1. However, claim 2 specifically defined the method of treatment as administering “Compound X eye drops.” The USPTO considered claim 2 to be eligible because it recites a “particular treatment for a medical condition”, which provides a practical application.

Key point: In the new SME examples, the USPTO did not base its eligibility findings on how an AI model was used, but instead on what actions were taken after the model was used. This suggests that the USPTO is not likely to consider methods or systems that merely use AI to process data to be eligible. To be eligible, patent claims directed to methods and systems that use AI should recite a specific hardware structure, a practical application such as a particular medical treatment or causing a system to take a discrete action, or an improvement to a technical problem that is described in the specification.

[View source.]