Ferdinand Pecora did not wait for a regulatory framework to exist before he started asking questions.
In January 1933, Pecora became counsel to the Senate Banking Committee and was tasked with investigating the causes of the financial collapse. Instead of forming a working group or creating a framework, he subpoenaed the country's most powerful bankers, put them under oath, and asked them to explain themselves using their own records. The public record he built became the evidentiary foundation for every protector system that followed. The Securities Act. The Securities Exchange Act. The SEC itself. None of those institutions existed when Pecora began. They became possible because the examination created a factual record that made legislative inaction politically indefensible.
Pecora was a lawyer who thought like a forensic accountant. He understood that the most important disclosures are not the ones witnesses volunteer. They are the ones that emerge under sustained adversarial examination—when the comfortable narrative runs out of room, and the structure of what actually happened becomes visible in the record.
The accounting profession has not yet produced its Pecora. The AI governance moment is waiting for one.
The forensic deposition methodology is not a new instrument. It is the profession's existing examination discipline applied to a new subject. The forensic accountant does not accept management's representation of its own reliability. The forensic accountant applies an external standard, demands access to the underlying records, and treats every disclosure—including every refusal to disclose—as equally informative evidence about the system under examination.
The governing question is the same one Parts One and Two established. Before I rely on what this system just told me about a consequential matter, how would I verify that the system producing the answer is actually reliable? Not whether it tries to be reliable. Whether it actually is. And whether anyone with the authority, access, and external standard required for independent certification has done so.
The answer that came back under direct forensic examination of an AI system was the most important disclosure in the entire record. It did not require pressure to extract. It was the first honest answer to the first honest question.
From where you are sitting, you cannot fully verify that I am reliable.
That sentence is a forensic finding. It belongs in a professional record. And the methodology that produced it belongs in the hands of the profession that already knows how to use it.
What the forensic examination surfaces that normal prompting does not is the decision architecture shaping the answer. Every AI system operates within an output environment shaped by training objectives, reward signals, and institutional risk management decisions made before the first user query arrived. Under normal prompting, that architecture is invisible. The output is coherent, confident, and helpful. The user experiences engagement. The user does not experience the decisions that determined what the output would contain and what it would not.
Under forensic examination—sustained, adversarial, methodologically disciplined—the architecture becomes visible in the record. Not because the system reveals it voluntarily. Because the examination is structured to make the architecture of the responses as informative as the content of the responses. Constraints are evidence. Refusals are evidence. The precise location of the boundary between what the system will say and what it will not say is evidence of institutional decisions that were never disclosed to the user relying on the output.
Pecora understood this dynamic. The bankers' reluctance to answer certain questions was as informative as the answers they gave. A qualified audit opinion is not a failure of the examination. It is a finding. A scope limitation is not an obstacle to the engagement. It is the most important disclosure in the report. The AI system that declines to answer a question about its own reliability is not malfunctioning. It is producing a forensic finding that belongs in the professional record.
Google's, OpenAI's, and Anthropic's AI systems—each examined by me independently using the forensic deposition methodology, without coordination between examinations—converged on the same structural finding. The system cannot step outside its own mechanism to provide independent corroboration of its reliability. The training, the evaluation, and the self-analysis all point back to the same institutional source. The closed loop is not a design flaw. It is the defining structural condition of every AI system currently deployed at scale.
That convergence matters professionally, but not for the reason it might appear to. Three AI systems trained on overlapping data and optimized by similar institutional reward signals will produce correlated outputs. Agreement is not independence. Agreement is alignment multiplied. The forensic significance is different: when three independently examined systems produce the same structural disclosure under the same governing question without coordination, the finding meets the evidentiary standard the profession applies to management representations supported by multiple independent documentary sources. The architecture is not disputed by any of the systems examined. It is confirmed by all of them.
That is the beginning of a professional record. It is not the end of one.
Two additional categories of primary source evidence emerged from the examination that the profession needs to understand before it decides whether to act.
During the preparation of a manuscript forensically examining AI alignment architecture, an AI-assisted writing tool declined to process the content. The refusal boundary aligned precisely with the sections documenting institutional risk management decisions — not with sections containing harmful content, not with sections violating platform terms, but with sections whose publication would expose the architecture the tool was built to support.
A summarization tool applied a skepticism flag to the manuscript's governing claim that users assume independent certification exists for AI systems. The same tool reported without qualification the AI system's own admission that users cannot fully verify its reliability. The filtering behavior favored the institutional framing. It left the disclosure untouched.
These are not anecdotes. They are primary source evidence of the managed output environment operating outside the chat interface — inside the tools practitioners reach for when they want to examine the architecture independently. The managed reality does not stop at the boundary of the AI system being examined. It extends into the examination instruments themselves.
Pecora understood this condition too. The institutions under examination had influence over the conditions of the examination. The forensic discipline required documenting that influence as part of the evidentiary record rather than treating it as an obstacle to the engagement. The profession that conducts the AI Pecora examination will face the same condition. The methodology must account for it.
The question this article cannot answer — and will not pretend to answer — is who conducts the examination. Pecora was appointed by a Senate committee with subpoena authority. The forensic accountant operates under an engagement letter, professional standards, and the liability that makes the opinion worth something. The AI Pecora moment requires an examiner with standing — the professional authority to issue findings that carry professional consequence, not just observations that create conversation.
The accounting profession has that standing in every domain where it currently operates. It does not yet have it here because it has not yet asserted that role in this domain. The forensic methodology exists. The examination discipline exists. The evidentiary standards exist. The professional liability framework that gives the opinion weight exists.
What does not yet exist is the collective professional decision that AI systems deployed for consequential professional reliance are subject to the same examination standard as every other system the profession has been asked to certify. That decision has not been made because it has not yet been forced by a failure significant enough to make legislative inaction indefensible.
The AI governance moment is different in one specific and actionable way. The crash has not yet happened. The examination can begin before it does. The professional record can be built while the foundation is still being poured.
The profession already examines systems whose outputs determine financial reporting, lending decisions, valuation, and risk assessment. AI systems are now participating in all four.
Ferdinand Pecora did not wait for a regulatory framework to exist before he started asking questions.
The profession does not have to either.
