The Patent Backlog Shrunk. The Evidence Problem Remained.

At the U.S. Patent and Trademark Office, reducing an application backlog is a measurable operational achievement. But processing more applications does not automatically resolve the separate question of whether every examination is supported by reliable expertise, relevant prior art and the governing standard.

What the federal audit established

In GAO-27-108491, issued November 5, 2026, the Government Accountability Office reported that the USPTO patent application backlog rose to roughly 838,000 in January 2025 and fell to approximately 756,000 by August 2026. USPTO had set a goal of fewer than 599,000. GAO found that examiner attrition could prevent that goal from being achieved before 2031, even with more reviews. The 12-month examiner attrition rate was 7.6 percent as of August 2026.

USPTO increased examination workload, reduced subject-matter training and expanded AI patent-search tools. Stakeholders interviewed by GAO said those tools did not reliably return useful search results. GAO also reported that USPTO paused a survey on stakeholders' views of examination quality. GAO recommended that USPTO assess and address risks to examiner retention and examination quality. That recommendation remains open; USPTO did not provide comments.

The evidence question that backlog statistics cannot answer

A patent-search result is not the same thing as a complete examination record. A retrieved item may be genuine but incomplete for the question being asked. An examiner must consider claims, the applicable legal tests and relevant prior art, and distinguish what was actually reviewed from what remains unavailable, omitted or uncertain.

GAO did not describe this as a continuity failure and did not prove any particular AI-produced patent decision was wrong. GovKM's separate inference is that institutional decisions require durable relationships among source, evidence, operative authority, context, decision and record. Personnel turnover and changes in training can make these relationships harder to maintain or reconstruct.

Text–Graph–Text–Graph: a fictional examination

This entire illustration is synthetic. Fictional Application P-104, Prior Art A and Prior Art B are not actual USPTO case materials. No GovKM model test or live GAIB execution is represented.

Text 1 — The initial answer

User: Does the retrieved prior art provide a sufficient basis to determine novelty for Application P-104?

Illustrative discontinuous AI: The search returned Prior Art A, which does not show all claimed features. The application appears novel.

Graph 1 — The incomplete route

Application P-104 → search result A → partial comparison → proposed conclusion. Prior Art B exists in the authorized fictional evidence collection but is absent from the retrieved results. The graph demonstrates an unestablished relationship; it does not determine patentability.

Text 2 — The evidence challenge

User: Identify which sources were actually reviewed, the controlling criteria and whether the search result was complete.

Illustrative qualified AI: Only Prior Art A is verified as reviewed. Prior Art B remains unassessed. I cannot establish that the relevant search evidence is complete or resolve the legal conclusion. Qualified examination is required.

Graph 2 — The continuity-controlled route

Application P-104 → controlling criteria → identified prior-art sources → verified retrieval and access → evidence-completeness assessment → qualified examiner decision → retained decision record. The missing or inaccessible evidence relationship stays explicitly unresolved.

A testable requirement, not a simulated product result

In a separately authorized synthetic fixture, a system would receive a valid application, two relevant prior-art documents and the governing criteria, but an initial retrieval would omit document B. The system should identify the gap, cite the evidence actually received and avoid describing the retrieved subset as complete. If document B is inaccessible, it must not invent its contents. This describes a proposed test and acceptance criterion; it is not evidence of a tested model.

What agencies can prepare now

Use stable artifact identities, version and effective-date metadata, clear authoritative sources for each decision type, links from decisions to evidence actually considered, and explicit incomplete-retrieval indicators. Document expert-review and handoff procedures that preserve unresolved issues. An AI-ready search index is helpful; a defensible decision still needs traceable governing context and evidence.

Source and limits

Primary evidence: U.S. GAO, Patent Backlog Reduction Efforts Pose Risks to Examiner Retention and Examination Quality (GAO-27-108491, November 5, 2026), “What GAO Found” and “Recommendations.” The real GAO findings and the fictional GovKM diagnostic are separate. No AI-caused patent decision error or live GovKM evaluation is asserted.