USPTO’s AI-Based Search Tools Send Warning to Patent Applicants - Bloomberg Law News

The US Patent and Trademark Officeâs Federal Register notice announcing its opt-in automated search pilot program is another step in the growing use of artificial intelligence in patent examination. It should prompt applicants to understand the long-term potential of AI-based prior art searching tools, their likely impact on the examination process, and what actions they can take to position todayâs applications for a potentially disruptive future.

The automated search tool uses an applicationâs cooperative patent classification designation, specification, claims, and abstract to identify up to 10 potentially relevant prior art documents, according to the October 2025 notice. Applicants can consider these references before formal examination and, if appropriate, amend their claims to expedite prosecution and overcome obstacles posed by the AI-identified references.

This pilot program isnât the PTOâs first foray into AI-based patent searching tools. It introduced a âsimilarity searchâ option four years ago that, in theory, enabled examiners to identify additional references similar to the application under review or other relevant references.

To date, these tools havenât fundamentally altered the patent examination process. But the rapid pace of AI advancement suggests their influence is likely to grow.

AIâs Possible Impact

There is no clear, unambiguous definition of an âAI search toolâ nor how it differs from prior search tools. But any system that falls under the AI-tool umbrella will leverage a custom-trained model, an existing AI model, or a combination of both to identify prior art references that may anticipate or render obvious an applicantâs patent claims.

These tools will differ based on operational modes and programmed features. Broadly speaking, they likely will fall into two types. One is examiner assistant tools that improve the search results presented to an examiner for review. The other is agentic tools that can autonomously search relevant databases and evaluate results in an iterative, reasoned manner to produce either traditional search results or initial office action drafts (in more advanced cases).

For examiner-assistance tools, the key differentiator from traditional search methods is the use of learned associations developed during training. Language models learn to encode the semantic meaning of an input passage â here a patent application â into a high-dimensional vector space in which different directions correspond to different semantic features, such as the relevant technology fields, component attributes, and system functionality.

These tools can then compare the application with vectors from prior art references to identify relevant references or portions of references for review by the examiner. Through this process, the tool can automatically understand the full scope of a patent disclosure or prior art document without relying solely on the traditional linguistic matching process of searching for identical or synonymous text across applications.

Agentic systems would likely operate similarly, except that the instantiated âAI agentâ would be trained and configured to review the identified references, conduct additional searching if needed, and generate a summary report â or even draft an office action that includes both the identified references and the reasons the agent believes them to be relevant.

In either case, the systems likely would have similar effects on the examination process. Specifically, they could increase the number of on-point prior art references identified and cited by the patent office and reduce the time it takes the office to examine individual filings.

However, this per-filing speed-up may not translate into meaningfully shorter pendency at the PTO, as the increase in better prior art may require applicants to engage in additional rounds of prosecution and/or appeals. The higher volume of better, more pertinent prior art likely will cause applicants to resort to narrowing claim amendments more often and even abandoning weak cases.

How to Prepare

Although there is no silver bullet, the most proactive step applicants can take is to file robust, detailed specifications focused on particularized improvements in technical systems, processes, devices, and methods.

Ideally, applicants should have a clear engineer- or scientist-level of understanding of the technical features and operation of the system, device, or method theyâre seeking to patent. This includes providing detailed examples and clearly articulating how the invention operates or is intended to operate.

While applicants arenât required to constructively reduce their invention to practice before filing, those that are closer to that goal will be better positioned for success. More prophetic and less developed patent applications may lack the disclosure necessary to support strong claim amendments needed to overcome the increased volume of relevant prior art that AI-based search tools may uncover.

This warning is especially relevant for computer- and software-related inventions, which often are filed earlier in the product development cycle compared with mechanical, chemical, or biological inventions. As with all technologies, the key disclosure elements arenât the outcomes that a system, device, or method achieves, but what the invention is and how it achieves those outcomes.

This article does not necessarily reflect the opinion of Bloomberg Industry Group Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.

Author Information

Michael R. Anderson is a patent attorney with Marshall, Gerstein & Borun in Chicago.

Interested in writing? Review our author guidelines and submit pitches to Insights@bloombergindustry.com.

Learn more about Bloomberg Law or Log In to keep reading:

See Breaking News in Context

Bloomberg Law provides trusted coverage of current events enhanced with legal analysis.

Already a subscriber?

Log in to keep reading or access research tools and resources.