The Second AI Checked the First. The Source Still Didn’t.

Using a second AI to review the first AI can improve a draft.

It does not necessarily verify the evidence.

A 2026 disciplinary matter before the U.S. Patent and Trademark Office makes the distinction unusually concrete. According to the published record, a patent attorney used one generative AI tool to assist in drafting proposed claim constructions and then used a second AI tool to review the draft. The resulting filing still contained numerous citations, quotations, and parentheticals referring to portions of the patent specification, figures, and prosecution history that either did not exist or were inaccurately attributed.

The first AI generated a representation of the source record.

The second AI reviewed that representation.

Neither step replaced returning to the authoritative source.

AI-to-AI Review Is Another Inference Layer

Organizations increasingly use one model to critique another. A second model can identify inconsistencies, challenge reasoning, improve style, detect omissions, or flag suspicious claims.

Those are useful functions.

But review and verification are different institutional acts.

An AI reviewer is still making inferences from whatever evidence it is given. If it receives the first model’s draft without independently resolving the underlying source objects, it may evaluate plausibility rather than truth. It can tell the institution that an answer looks supported while never establishing that the cited support exists.

This is especially dangerous when the output contains the surface features of verification: citations, quotations, pinpoint references, confidence language, or a second model’s approval.

The process can look more rigorous while remaining detached from evidence.

Verification Must Reattach the Claim to the Source

In the GovKM Continuity Topology, the critical early relationship is:

Source → Evidence.

A citation is not the Source. It is a representation that claims a relationship to the Source.

Verification asks whether that relationship is real.

Does the cited section exist? Does the quotation appear there? Does the figure show what the draft says it shows? Does the prosecution-history document contain the asserted statement? Is the version correct? Is the passage being used in a way the source supports?

Those questions require reattachment to the evidentiary object.

A second generative model can help perform that work only if the architecture actually gives it the source and requires it to resolve each claim against that source. Merely asking another AI to review the prose is not source verification.

Intrinsic Evidence Makes the Lesson Stronger

Most public discussion of AI hallucination in legal work has focused on fabricated cases, statutes, and external authorities.

The USPTO matter is useful because the errors reached into the intrinsic record of the patent dispute itself: the specification, figures, and prosecution history.

That broadens the continuity lesson.

The risk is not only that AI invents an external authority. AI can also misrepresent an authentic document already inside the working environment.

The Source can be real while the claimed relationship to it is false.

This is a different failure mode from a nonexistent citation. It is a valid-object, invalid-relationship problem.

More Models Do Not Automatically Create More Provenance

Layered AI workflows can create a false sense of independence.

One model drafts. Another critiques. A third scores. A retrieval system adds citations. A compliance model checks policy. The final output may look like the product of several independent controls.

But if every stage consumes representations derived from the same unsupported claim, the workflow can amplify confidence without increasing evidentiary grounding.

Continuity depends on whether each critical claim can still be traced to a source object whose identity, version, authority, and meaning remain recoverable.

This is why model diversity is not equivalent to evidence diversity.

Human Review Has the Same Requirement

The lesson should not be interpreted as “humans must manually redo everything AI touches.”

Human review can also fail when reviewers inspect a polished representation instead of the source.

A lawyer can read a draft and miss a fabricated pinpoint citation. A manager can approve a report without opening the dataset. An analyst can accept a summary whose quoted policy language is subtly wrong. An auditor can review a dashboard without checking the transaction evidence beneath it.

The continuity requirement applies to humans and machines alike:

when the validity of a consequential claim depends on source evidence, verification must return to that evidence.

The Record Should Show What Was Verified

Organizations should also preserve the verification relationship in the resulting record.

For consequential AI-assisted work, it may be useful to distinguish:

generated claim; candidate source; source resolution; exact supporting passage or data; verification result; reviewer identity or process; unresolved conflict; correction; and final promoted statement.

That does not require turning every task into a bureaucratic exercise. The level of control should match consequence.

But a system should not label content “verified” when all it can establish is that another model found it plausible.

Verification Is a Continuity Function

The USPTO disciplinary order is ultimately a professional-responsibility matter, not a technology benchmark. The attorney corrected the errors promptly after they were identified; the district court imposed no sanction, and the published account records no client prejudice.

The GovKM significance is architectural.

The workflow contained drafting and review. What failed was the continuity relationship connecting the claims back to the authoritative record strongly enough to justify institutional promotion into a filed document.

That distinction will matter in every sector where AI prepares material that can become authoritative: legal filings, clinical documentation, financial analysis, compliance reports, procurement evaluations, intelligence products, policy summaries, and executive decisions.

The GovKM Proposition

Verification is not another opinion about the output. Verification is a governed return to the evidence that makes the output supportable.

AI can review AI. AI can challenge AI. AI can even help verify AI.

But the verification architecture must re-establish Source → Evidence rather than merely adding another inference layer between the institution and the source.

The second AI checked the first.

The missing check was the one that mattered most.