USPTO Plans For AI Implementation

While this blog normally writes about the issues facing the patentability of AI inventions, we want to write about a topic that is high on our radar: the USPTO’s plans for integrating AI into the examination process. While certain programs, such as the USPTO’s Artificial Intelligence Search Automated Pilot Program (ASAP!) program were previously tested, a greater question looms on the availability of examiners to use large language models (LLMs) as a tool for examination.

During a “USPTO Hour” information session earlier this year on April 29th, the Office provided an update on how the agency is driving artificial intelligence (AI) innovation and enabling opportunities to use AI tools to support their mission. A summary of that discussion is below and a link to the video recording can be found here.

USPTO officials said AI is already being used throughout the agency, but they repeatedly described it as a support tool rather than a replacement for employees’ expertise and judgment.

The agency’s goals are to use AI to:

• reduce patent and trademark processing times;
• improve consistency and quality;
• lower costs;
• identify fraud;
• help employees research, summarize, draft, analyze data, and organize work; and
• improve service to inventors, businesses, and the public.

Patents

Traditional AI and machine-learning tools have already been used for years in areas such as classification and relevance ranking. However, ordinary patent examiners are not yet permitted to use generative AI for official examination work.

USPTO is testing its internal generative-AI system with supervisory patent examiners before expanding access. Potential future uses include: 

• interpreting patent claims;
• summarizing inventions;
• checking priority dates;
• mapping claims to prior-art references;
• helping with novelty and obviousness analysis;
• structuring subject-matter eligibility analysis; and
• identifying possible clarity, written-description, or enablement problems.

The proposed system would provide editable drafts, source links, and confidence indicators. Examiners would remain the authors and final decision-makers. USPTO said it would expand the tools only when there is evidence that they improve quality and save time.

Trademarks

The trademark division has moved more quickly in deploying AI tools. One major system, called Class Act, helps classify trademark applications. It can suggest an international class, design-search codes, pseudo marks, mark descriptions, and color claims. Officials said the system had already handled more than 67,000 submissions. USPTO also introduced an image-search capability that lets users upload a design and search for visually similar federally registered trademarks.

The trademark office is also seeking vendors to help develop AI-assisted fraud detection, examination, and quality-review tools. Its internal employees have begun using the agency’s large language model, although examining attorneys were being given access more gradually while policies and safeguards were finalized.

Security and protection of intellectual property

Deborah Stephens, the USPTO CIO, emphasized that protecting intellectual property and agency data is the top priority.

USPTO said its AI systems are being developed with:

• cybersecurity controls;
• restrictions on who may access data;
• contractual limits on vendor use of agency data;
• secure cloud infrastructure;
• FedRAMP-approved commercial tools;
• logs, version controls, and auditability;
• secure endpoints; and
• agency ownership and control of internal models and data.

Officials repeatedly said the systems must not operate as unexplained “black boxes.”

AI policy and intellectual-property law

The policy discussion focused on legal questions created by AI, including:

• whether AI-assisted inventions qualify for patents;
• inventorship when AI is involved;
• what information applicants may eventually need to disclose about AI models or training data;
• how AI may change novelty and obviousness analysis;
• the rapid growth of AI-generated prior art;
• copyright questions involving training data; and
• protection against unauthorized AI replicas of a person’s voice, likeness, or identity.

Officials said USPTO is monitoring court decisions, legislation, and White House policy. They stressed that the ordinary legal standard for inventorship still applies even when AI tools are involved.

Preparing employees

USPTO is building an agency-wide AI change-management program. It has created an AI hub with training, governance information, examples, prompt guidance, and AI playbooks.

The agency also has several internal oversight groups, including:

• an AI Policy Council;
• an AI Governance Council;
• a senior-leader AI working group;
• business-unit AI teams; and
• a formal process for reviewing and approving new AI uses.

Individual divisions are also selecting AI “champions” to help employees learn the tools and provide feedback.

Concerns about AI-generated applications

The panel acknowledged that USPTO has received applications containing gibberish or meaningless material. Officials said it is not always possible to determine whether those filings were produced by AI, but both the patent and trademark divisions are actively studying the issue.

They indicated that additional guidance or public statements about handling AI-generated or unsupervised filings may be issued.

Central takeaway

USPTO is trying to become a leading federal agency in AI adoption, but its stated approach is controlled and evidence-based. The agency wants AI to handle routine research, drafting, classification, searching, and administrative work while employees retain responsibility for legal reasoning and final decisions.

One line in particular from the discussion captured the overall message: “Employee expertise leads; technology supports.” 

We are paying close attention to AI adoption at the USPTO since it will effect what we see in terms of output and work product from the agency. First, the use of AI will increase productivity from the USPTO, which increases workload on the patent owner side of the fence. Secondly, patent practitioners will need to scrutinize responses with a different perspective once we understand the level at which AI is being used to contribute to analysis and content generation. It might be the case that it is up to patent attorneys to spot hallucinations and errors that are already common in AI. For instance, AI often seeks to support the desired outcome depending on the prompt. If the question is “how can I reject this claim?” vs. “should I reject this claim?”, the AI output can be completely different. Therefore, while AI can definitely make our lives and work easier, vigilance cannot be forgotten.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.