Legal research platforms have begun to ship audit trails alongside their AI agents, which is essentially a log of the steps the agent took, the searches it ran, and the sources it opened while preparing a research report. The feature is marketed as a verification mechanism and a defensible record.
On a recent panel, I was asked whether Midpage offers something similar. The answer is yes, though it is not much of a differentiator: the ability to drill down to see what the agent did is table stakes on virtually any AI platform today.
But the question rests on a premise worth examining. Can an audit trail fulfill an attorney’s obligation to verify legal research? The short answer: no, not really.
Where AI research goes wrong
A research log answers a narrow question: what searches did the agent run and what sources did it consult? That is useful if verification means confirming where the agent looked. But the errors that plague briefs are rarely errors of process. Setting aside outright fabrication, they are errors of judgment about what a source says:
- The case is real, but the quoted line is from the dicta or dissent.
- The quotation is accurate, but the court was reciting the losing party’s argument.
- The holding turned on a different procedural posture, an earlier version of the statute, or a standard of review that does not apply.
- The case is on point and was later narrowed by a decision the summary omitted.
Each of these is a judgment about the content and import of an opinion. None of these nuances is visible in a record of which searches were run and which documents were opened.
The empirical work on legal AI recognizes this distinction. The Stanford RegLab evaluation of AI research tools counted a real case cited for a proposition it does not support as an error, and that category is where much of the risk exists. The sanctions decisions adopt the same view. In Noland v. Land of the Free, the court found that nearly all of the quotations in the opening brief were fabricated, that many cited cases did not discuss the topics for which they were cited, and that a few did not exist at all. The database that tracks these decisions sorts them into fabricated content, false quotes, and misrepresented material.
A closed database of real opinions can flag a citation to a case that doesn’t exist, but nothing in a log of the agent’s activity can catch a case that has been mischaracterized or overstated.
How a memo is verified
The analog of an agent’s activity log is an associate’s search history. Legal AI platforms like to describe their agents as tireless associates who never miss a citation. Even assuming such an associate exists, a supervising partner does not verify her memo by reviewing her search history. If he is fulfilling his duties of candor and competence, he reads the memo’s treatment of the authorities, and when something looks wrong, he pulls the case to read it directly.
The gap is wider with an agent, whose account of its own reasoning is itself generated text. Researchers studying AI research agents have found that provenance records typically capture what an agent did but fail to encode which sources substantiate each claim and how. What is needed is auditability at the level of the claim, not a log at the level of the action.
A log is not independent verification
Even an imperfect audit trail might seem better than none. But the risk is that a process log feels like diligence and lends a false sense of verification. A 2025 review of automation bias in high-stakes domains, including law, concluded that explanations often increase the perceived acceptability of a system without improving decision accuracy, and can reinforce misplaced trust, especially among less experienced professionals. Earlier controlled studies found that explanations increased acceptance of the AI’s recommendation whether or not it was correct.
Nor is it clear that reviewing a log satisfies an attorney’s ethical obligations. ABA Formal Opinion 512 requires independent verification or review of generative AI output. A ritual built around the agent’s process shifts reliance from the agent’s output to the agent’s account of its output. An activity log is a timesheet for the agent. It shows that the work was done; it cannot show that the work is right.
Would a court accept it?
The strongest case for an audit trail is defensive: if a bad citation slips through, the firm can point to the audit trail to show what the agent did. Courts have already answered this approach.
In Park v. Kim, the Second Circuit held that Rule 11 requires attorneys to read, and thereby confirm the existence and validity of, the authorities on which they rely, adding that it could think of no other way to ensure an argument is warranted by existing law.
In Wadsworth v. Walmart, the attorneys had used their own firm’s AI platform; the court observed that an excerpt could be a losing party’s argument, an overruled case, or dicta, and that the duty to check sources is unchanged by the transition to AI. The court noted the firm’s remedial policies with appreciation but did not excuse the signing attorneys.
In Johnson v. Dunn, the firm had an AI policy requiring verification of every output; the court treated the attorney’s disregard of those internal controls as an aggravating factor, and refused to discount the sanction because real authorities happened to support the propositions.
A log showing where the agent looked does not rebut any of this. To the contrary, it suggests that the lawyer did not read the case.
Verify the claim, not the process
Effective verification tools operate at the level of the claim and the citation. As an example, BriefCatch’s RealityCheck checks a brief’s citations against legal databases and evaluates whether the quoted language and cited authorities support the propositions for which they are offered, flagging fabricated quotations, misstated holdings, and unsupported propositions. That is the right unit of verification — the proposition and the authority — not an agent’s account of the route between them.
Midpage adopts the same approach on the research side. Explanations reduce over-reliance on AI output when they lower the cost of verifying the answer, so Claude with Midpage returns answers in which every citation is a live link that scrolls to the passage being relied on. The links work without a Midpage account, so a supervising partner, a client, or a court can open the authority and confirm that it says what the answer claims, even if they don’t have a Midpage account. Either way, the lawyer ends up where the courts require: reading the opinion.