Meeting the Written Description Requirement for AI Algorithms: Lessons from a Recent PTAB Decision

Patent practitioners are well aware of the importance of a thorough written description under 35 USC Section 112. However, when it comes to AI inventions, ensuring that all aspects of an algorithm are sufficiently described can be particularly critical. A recent PTAB decision involving Microsoft’s attempt to patent certain AI algorithms for LinkedIn serves as a valuable case study.

The invention in question (Appeal 2023-000931, Application 15/215,251) relates to techniques for performing skill-based recommendations of events to users. Claim 1 is reproduced below with limitations at issue bolded:

1. A method, comprising:
obtaining user attributes for a user of an online professional network, including one or more job titles held by the user and one or more professional skills of the user that have been endorsed by other users of the online professional network;
matching a location of the user and one or more of the user attributes to event attributes of a set of events;
applying a neural network to the one or more user attributes and the event attributes to calculate a set of relevance scores representing relevance of the events to the user;
identifying a plurality of user connections for the user in the online professional network;
combining the plurality of user connections for the user with an attendee list for the event to determine a subset of the user connections that are on the attendee list;
displaying via a graphical user interface (GUI) a first subset of the events as recommendations to the user, based on the set of relevance scores, the displaying including rendering, for each event, at least four user interface elements, the first user interface element being a title, the second user interface element being a location, the third user interface element being an indication of a number of the subset of user connections that are on the attendee list, and a fourth user interface element being a user interface element which, when selected, adds the corresponding event to a calendar of the user; receiving a response to the displaying of the first subset of the events; aggregating the response and other responses to the event from other users of the online professional network into an overall response to the event;filtering the overall response to include responses from connections of the other users; using the filtered overall response to calculate an additional relevance score representing a relevance of the event to an additional user of the online professional network, based on the overall response;outputting the event, via the GUI, as a recommendation to the additional user;receiving a selection from the user of the fourth user interface element;and responsive to the receiving, adding the corresponding event to the calendar of the user.

While the specification had literal support for “applying a neural network to the one or more user attributes and the event attributes to calculate a set of relevance scores representing relevance of the events to the user”, the examiner asserted that the specification did not contain sufficient additional details as to how this was carried out, and thus the claim lacked sufficient written description so that one skilled in the art would recognize that the inventor had possession of the invention. Note, there was not lack of enablement alleged. The crux of the issue was that while the Specification outlined the desired outcome of the neural network's application, it failed to detail the operational mechanism by which the neural network would achieve this outcome. The PTAB affirmed the rejection as follows:

Accordingly, we find that the Specification merely describes a desired result of applying a neural network to one or more user attributes and event attributes to calculate a set of relevance scores representing relevance of the events to the user. The Specification fails to describe how a neural network functionality operates within a computer to calculate the set of relevance scores. In paragraph 38 of the Specification, Appellant describes nine different examples of how the relevance scores may be calculated, some through pure arithmetic computation and others, through devices, such as the neural network and support vector machines. Although in paragraphs 27 through 30, the Specification does describe how computing mathematical model results in the relevance scores, these paragraphs nevertheless fail to disclose how using a dedicated device, such as a neural network accomplishes this. The latter requires more description especially given the generic nature of the claims

The PTAB's decision underscores the importance of the specification sufficiently supporting algorithmic claim elements for practitioners: merely stating the desired result of an algorithm is not enough. The inventors cannot merely make a wish for the result, they have to disclose a way to achieve it to show possession. A prudent drafter will ensure that the Specification goes beyond the goals and provides a detailed description of how the algorithm functions within its environment to achieve the claimed result. This level of detail is essential, even if the workings of the algorithm might seem obvious to those skilled in the art, and even if a person skilled in the art can make and use the invention without undue experimentation. One approach to consider when drafting patent applications for AI inventions is to provide a roadmap that delineates not just what the algorithm is designed to do, but precisely how it accomplishes its tasks by giving additional details and examples.

So, when it comes to drafting patent applications for AI inventions, don’t just set forth wishes and hopes for what the algorithm is supposed to do, but also link in additional details for achieving the results. Hope is not a plan, and wishes are not a strategy!