Accounting got there first: The blueprint for AI verification

A term is about to get crowded. “AI Verification” will soon be argued over, defined, redefined and claimed by dozens of disciplines. Before that happens, it’s worth saying plainly that a working version of it got there first, quietly, inside accounting, decades ago.

Processing Content

Verification is not agreement, confidence or expertise. It is a specific process: locate the original source, compare the specific claim against it directly, determine whether they match and require that the comparison be capable of coming back negative. A check that cannot fail isn’t a check. It’s a costume.

Accounting did not become a verification profession by talking about integrity. It became one by assuming integrity would sometimes fail and designing institutions that could survive that failure.

Accounting built one of the most mature institutional versions of that process in the world, though not because accountants set out to solve a problem called “verification.” Enron forced the issue. The profession could no longer answer with abstract principles or personal virtue alone, so accountants were compelled to create structural safeguards: systemic solutions that embedded independence into the profession’s architecture.

The rulebook is specific, and it is still in force. The AICPA’s conceptual framework names the failure directly: a “self-review threat,” the risk of relying on work the firm itself produced without evaluating it properly, and a “familiarity threat,” the risk that a long or close relationship with a client makes an accountant too sympathetic to the client or too accepting of its work. Sarbanes-Oxley’s Section 203 answered the familiarity threat with a number: After five years, the lead audit partner and the reviewing partner must rotate off a public company engagement, not because either has done anything wrong, but because closeness itself is the risk. A related rule, Section 206, adds a one-year cooling-off period: a firm may not audit a company whose chief executive, chief financial officer, controller or chief accounting officer worked on that company’s audit for the firm within the prior year. And underneath both rules sits the reason they exist. 

At a Dec. 18, 2001, Senate hearing, Sen. Ernest Hollings said Arthur Andersen claimed Enron paid it $25 million for accounting services and $27 million for consulting. A July 2002 report by the Senate Permanent Subcommittee on Investigations found that Enron’s board had allowed Andersen “to provide internal audit and consulting services while serving as Enron’s outside auditor.” The checker depended financially on the checked, and it was reviewing its own work. Andersen’s people were not incompetent. The audit failed because independence had become theater.

That rulebook didn’t solve everything. It solved one problem, with real precision: The checker cannot depend on the checked. Not “should try not to.” As an enforceable rule, it specifies exactly how long is too long and attaches a prohibition to the money trail. That is a genuine, hard-won piece of institutional engineering, and it predates current AI governance proposals by more than two decades.

It is also where accounting’s model runs out, and the limit matters more than the achievement. Independence rules assume the underlying record is real — the invoice, the contract, the transaction log. Audit methodology was built to catch a distorted account of real events, not a fabricated one. AI verification inherits the independence problem accounting solved, but adds one accounting rarely faced: The evidence itself can now be fabricated cheaply enough that independence alone is no longer sufficient. A rule that guarantees the checker has no financial stake in the outcome does nothing to guarantee the document on the table is real in the first place.

That gap is why accounting is a case study inside something larger, not the whole answer. Medicine solved sterility: A specific discipline against a specific, well-understood failure mode. Aviation solved redundancy: no single point of failure gets to be the only thing standing between a decision and a catastrophe. Law solved adversarial testing: The truth gets established by forcing two motivated parties to attack each other’s account, not by trusting either one. Accounting solved institutional independence. None of these disciplines solved the others’ problems, and none of them alone solves AI’s. AI Verification has to integrate all four, plus the one none of them had to face: verifying the evidence itself, not just the people examining it.

That last piece has a famous precedent. Two women once came before Solomon, each claiming the same child as her own. He did not ask either one to testify more persuasively — testimony is cheap, and a liar can match a truth-teller’s confidence claim for claim.  He ordered the child cut in two and watched how each woman responded. The test worked because it forced honest and dishonest actors to behave differently under pressure, something neither could fake by simply sounding more certain. Solomon did not verify testimony. He designed a situation where testimony stopped mattering. That is the piece accounting’s document-based model doesn’t have: not another layer of independent review, but a situation designed to make truth and falsehood produce visibly different outcomes.

The honest version of the claim is narrower than “accounting solved AI’s hardest problem,” and more durable for being narrower: accounting did not solve AI verification. It independently discovered several of the structural principles that any mature verification discipline eventually needs, decades before AI needed them, and is now watching AI governance rediscover the same principles from scratch, more slowly, with the scars still optional.

One piece of that rulebook transfers directly, right now: independence built into the structure and not dependent on anyone’s good intentions. Even frameworks that require evaluators to be independent of developers, such as the EU AI Act’s rules for notified bodies, stop short of accounting’s machinery: Those rules contain no mandatory rotation and no cooling-off period. Accounting didn’t just theorize about such requirements. It legislated them, enforced them and lived under them for more than twenty years.

Congress’s own FRONTIER Act, introduced in July to require audits of the largest AI developers, goes further than most. It bars a developer from hiring an auditor if either side has a financial interest in the other, and it forbids tying the auditor’s pay to the audit’s result. But as of this writing it is only a bill: It may never pass, and it may change if it does. That annual compliance audit is full of holes. The developer writes the framework, chooses and pays the auditor, and is audited against its own rules. That provision has no rotation and no cooling-off period. For the largest developers, the bill adds federally licensed verifiers, but the developer still chooses them, and they too face no rotation or cooling-off. To watch their independence over time, the bill offers a report: each year, starting one year after enactment, the Government Accountability Office would assess how independent the licensed verification organizations are from the AI industry and recommend changes. Recommendations are not rules. Accounting answered the same problem more than two decades ago: rotate the lead partner after five years. A report on independence is not independence built into the structure.

AI governance already has the principle. What it doesn’t have is the mechanism — the specific, enforceable, inconvenient rule that makes independence survive contact with money. That’s the part accounting had to learn by watching a firm collapse. AI governance still has the choice to read the blueprint instead of living the lesson.