AI due diligence on executives: what to trust

AI due diligence on executives is the habit of asking an AI assistant who a person is, what they have done and whether they can be trusted, before a funding round, a board seat or a deal. An assistant is useful for the first orientation and for deciding where to look. It is not a record, and what it says about an executive should be treated as a lead to check, never as a finding to rely on.

That matters on both sides of the table. An investor who pastes a name into an assistant gets a fluent paragraph that can mix two people, quote a role that ended years ago or invent a detail that was never published. An executive who never looks at those answers does not know what the first impression of the deal looks like.

This guide is written for both readers. It sorts the jobs people give assistants into safer and riskier ones, lists the ways answers about executives go wrong, and shows an executive how to audit and correct them honestly. It rests on the provider and regulator pages named in the text, as read for this guide, and on our experience on files, written as tendencies and never as promises.

Why an AI answer about an executive is a lead, not a record

An assistant writes sentences that sound like a profile. It does not look up a profile. That is the whole problem in one line, and OpenAI states the risk plainly on its help page about whether ChatGPT tells the truth. The page says the system can produce incorrect or misleading outputs, can sound certain when it is wrong, and can produce invented quotes or sources. It advises treating ChatGPT as a first draft and not a final source, and checking cited sources by opening them.

Related readAI Answers About Job Candidates: Audit and Fix

The US National Institute of Standards and Technology uses a sharper term. In its generative AI risk profile (NIST AI 600-1, section 2.2), it defines confabulation as a system generating and confidently presenting erroneous or false content in response to prompts, and it notes that people may believe false content because of the confident nature of the response. Confidence in the tone of an answer says nothing about whether the answer is right.

For an executive this has a practical meaning. The paragraph an investor reads may be accurate in nine sentences and wrong in one, and the wrong one is usually the one that sounds most specific. A reader cannot tell which. OpenAI's help page for ChatGPT search says responses that use search contain inline citations and that the cited sources can be found with a Sources button at the end of the response. A citation lets a reader check a claim, but it does not prove the page says what the sentence claims.

The jobs people give assistants, and how far each can be trusted

Not every use is equally risky. The useful question is what the answer will be used for. The table sorts the common jobs by reliability and by what to do instead of relying on the assistant.

Assistants are best at language tasks, such as summarizing material the user supplies, and weakest as the source of record for an identity. The most dangerous row is the fourth, because "are there any red flags?" invites the assistant to answer a question it cannot know. If the honest answer is "nothing in the pages I could reach", a fluent system may still phrase it as "no known concerns". That sentence reads like a clearance and is not one.

Related readAI answers about doctors and lawyers: what to fix

Seven ways answers about executives go wrong

These are the failure modes we see most often when we read assistant answers about senior people. Each has a different cause and a different fix, which is why a general "the AI is wrong" complaint rarely leads anywhere.

Merged identity

Two people with the same or similar name become one profile, with the career of one and the controversy of the other. This is the most damaging error for an executive with a common name, and it also fools an investor who does not check. The fix is disambiguation: consistent facts about who you are, on pages you control, so that a system has something to anchor on.

Stale role

The assistant states a job the person left years ago, or omits a recent one. This usually follows the public web: an old biography that was never updated, a conference page that still lists a former employer, a directory entry nobody owns. A new role announced on only one page can lose to five old pages that repeat the previous one.

Invented credential

A degree, an award, a board seat or a patent that does not exist. It can also be the opposite, a real credential credited to the wrong person. Both are guesses that a fluent system fills in when the pages are thin. Investors who see an impressive credential in an assistant's summary should confirm it with the institution or the person, not with a second assistant.

Fabricated quote or source

A quotation attributed to the executive that they never said, or a link or article title that does not exist. OpenAI names invented quotes and sources as a known risk. An investor should open every cited page and search it for the quoted words. A quote that cannot be found on the cited page is a warning about the whole answer.

Related readWrong company information in ChatGPT: how to fix

Unsupported criticism

A negative claim with no source behind it, or a source that reports an allegation and a summary that states it as fact. This is the failure that carries legal weight, and the legal routes are a separate topic. The point here is narrower: an unsupported criticism in a diligence context can end a conversation before the executive knows it was said.

Silence about a real record

The assistant says nothing about a genuine, public matter, such as a regulatory action, a court judgment or a prior insolvency, because it did not reach the page or did not weigh it. An investor who reads silence as a clear record is exposed. An executive who hopes that silence will last is making a mistake too: formal checks do not rely on an assistant, and the record will be found. We would not help anyone try to hide it.

Tone drift

Each sentence is defensible, yet the paragraph leaves a different impression from the sources: "faced scrutiny" in place of "was named in a complaint that was dismissed". No single claim is false, so it is fixed by giving the system a clearer, dated source to summarize.

What the record systems hold that an assistant cannot replace

The sources that carry weight in diligence are records with an owner, a date and a procedure. The details differ by country and by sector, so this section uses two US examples read for this guide and stays general beyond them.

The SEC's page for searching filings describes EDGAR as free public access to millions of informational documents from public companies and others. It lists company search, a full-text search of filings, and filters by date, company, person and filing category. An assistant summarizing "what the filings say" is paraphrasing something the reader can open directly.

Related readGoogle AI Mode reputation: audit what it says

FINRA's BrokerCheck, according to its site, shows employment history, certifications, licenses and violations for brokers and investment advisors. An investor weighing a financial professional can check the register rather than ask an assistant whether there are problems. The joint investor alert on artificial intelligence fraud from the SEC, NASAA and FINRA, dated January 25, 2024, points to the same step: use the Check Out Your Investment Professional tool on Investor.gov to confirm whether a professional is registered and to review any disciplinary history.

Beyond those, the families of records are familiar: company registers, court dockets, sanctions lists, professional licensing bodies and the person's own documents. The principle is the same in every country: if a consequential claim has a record, read the record.

Why the way an investor asks changes the answer

A language model continues the conversation it is given. A question that carries a premise tends to pull the answer toward the premise. That is a general property of how these systems are used, and we describe it as a tendency we see when reading logs, not as a measured rate.

Compare three ways of asking about the same executive. "Summarize this person's background" asks for an orientation. "What controversies has this person been involved in?" presupposes that there are some, and an assistant that finds thin material may stretch it. "Is this person trustworthy?" asks for a judgment that no public source contains, and the answer will be built from whatever tone the pages have.

The executive's audit should therefore include neutral, leading and trust-seeking versions of the same question. If the neutral version is accurate and the leading version invents a controversy, an investor who types the second one will see it. The aim is to know the range of answers, not to chase one bad output. Answers differ between runs and between assistants, so a finding that appears in most runs matters more than one that appears once.

Related readWhat Siri, Alexa and Google Assistant say about you

How an executive audits these answers before a raise or a deal

An executive should look before the investor does, with the investor's questions and the investor's tools. The goal is a log of claims, each with an assistant, a date, a source and a verdict.

1. Write the investor's questionsTen to fifteen questions: who is this person, what are their roles and exits, what has been written about them, are there regulatory or legal matters, and who is the same name. Include a neutral, a leading and a trust-seeking version of the three most important.
2. Run them on several assistantsUse at least two assistants that search the web and one that answers mostly from training, on more than one day, signed out of any account that has memory of you, so you see what a stranger would see.
3. Log every claim with its sourceSave the full text, the date and the cited pages. Split each answer into single claims so that one wrong sentence does not hide inside nine right ones.
4. Open every cited pageCheck what the page actually says. Mark each claim as supported, unsupported, outdated or about someone else.
5. Trace each error to its causeA stale page, a missing page, a same-name page, a summary that overstated a source, or a claim with no source at all. The cause decides the fix.
6. Fix the sources, then re-runCorrect the pages you control and ask the owners of the others. Re-run the same questions after the sources have had time to be re-read, and compare with the first log.

One caution keeps the exercise honest: do not try to coach an assistant with hidden text, fake pages or invented profiles. Those tactics break platform and search engine policies, an analyst is likely to notice them, and a discovered manipulation turns a minor error into a character question.

What to fix, what to explain and what to leave alone

Not every finding calls for action. The table below gives the decision rule we use. It follows one principle: change what is wrong, add context to what is incomplete, and never try to suppress what is accurate, lawful and relevant to a business decision.

The last row is the one executives find hardest. A real record that is material to a funding or hiring decision belongs in the conversation, in your words and early. A formal search will find it. A disclosed fact with context is a question that is answered. An undisclosed fact discovered by an analyst becomes a question about candor.

What investors, boards and hiring committees should do differently

If you are the one reading an assistant's answer about an executive, a small number of rules cover most of the risk. They apply whether the assistant is a consumer chatbot, a search feature or a vendor's report that was assembled with one.

1. Treat the output as a list of claims to verify, not as a summary of the person.
2. Open every cited page and look for the claim on it. A citation that does not support the sentence is a reason to distrust the sentence.
3. Run the same question on a second assistant and compare. Disagreement is information. Agreement between two assistants that drew on the same weak page is not independent confirmation.
4. Verify identity first. Before any claim is accepted, confirm that the pages are about the right person, using facts the person has supplied.
5. Read silence carefully. An absent record in an assistant's answer is not evidence that no record exists.
6. Put the question to the person. A short, written list of what you found, and an invitation to correct or add, is fair to them and improves your file.

A related risk comes from the other direction. The SEC, NASAA and FINRA alert notes that deepfake videos may imitate a company's CEO to announce false news and manipulate a stock's price, and that fraudsters can use AI to clone voices, alter images and create fake videos. Its guidance is to verify communications by contacting the party through details you find independently, not through the message itself, and not to rely solely on AI-generated information. For diligence, the practical rule is that a video, a voice message or a "statement from the CEO" that arrives through an unverified channel is evidence of nothing until it is confirmed at the source.

Related readWhat Microsoft Copilot says about you: check and correct

Neighboring problems this guide only touches

Adjacent topics have their own treatment on this site, so this guide stays on the diligence question.

A plain web search of a name shows ranked pages that the reader judges, while an assistant shows a synthesis that the reader tends to accept. The two should agree, and where they do not, the gap is worth examining. A formal founder search is run by an analyst against registers, litigation databases and sanctions lists, with identity matching and a written report; a good analyst treats an assistant's output as one more thing to verify. Monitoring is a fixed set of questions run on a schedule, and for a diligence window a dated baseline and one re-run before the decision are usually enough.

Assistants also differ in how they cite and in what feedback controls exist, which affects how you read citations and where you send a report, not how the audit works. Same-name confusion deserves its own mention for executives with common names: it often reaches an investor before the executive knows, because the assistant sounds sure and the merged profile includes real, specific details. For the service side of all this, see the page on AI search reputation.

A worked scenario: ten working days before a term sheet

This is an invented scenario with generic parties. A chief executive expects an investor's partner to run a due diligence review of the leadership team in about three weeks. The executive has a common name, a former role at a company that later restructured, and a stale conference biography that still lists an employer from six years ago. The assumptions are that the executive controls the personal site and company site, can ask the conference organizer for a change, and cannot edit third-party pages.

Related readChatGPT wrong information about my product: how to fix

In an outcome like this, the stale employer usually disappears from some assistants and not others within the window, because they re-read pages at different times, and a provider report may still be open. The point of the ten days is not a perfect record. It is that the investor meets an accurate, consistent account from the executive first, with the real events told in the executive's own words.

Your next step: run the investor's questions this week

Start with the six steps above and a spreadsheet, and do the first run before a deal is announced. Keep the log as a plain record of what you saw and when. If you find an answer that is false and damaging, the provider routes and legal options are separate steps, taken in order, with the evidence you have just collected.

Questions people ask

Can an investor rely on ChatGPT to check an executive?

No. An assistant can suggest who to look into and what to ask, but OpenAI itself advises treating ChatGPT as a first draft and verifying important information from reliable sources. Anything that matters, such as a role, a credential, a regulatory matter or a legal dispute, should be confirmed in the filing, register, court record or document that holds it.

What should an executive do if an AI says something false about them before a deal?

Save the answer with its date and cited sources, work out whether the cause is a stale page, a same-name page or an unsupported claim, then fix the sources you control and report the error to the provider with evidence. If the false claim is serious and damaging, take legal advice. Tell the investor's team about it plainly, rather than hoping they will not see it.

Does a reassuring AI answer mean there is no problem on record?

No. An assistant may not have reached the page that holds a record, or may have weighed it lightly. Silence in a summary is not a clearance. Formal checks use registers, court records and identity verification, which do not depend on what an assistant happened to say. Executives should assume that a real, public record will be found.

Is it legitimate to ask an AI provider to correct an answer about me?

Yes, for inaccurate, outdated or unlawful content, and providers publish routes for it. It is not a legitimate route for removing something accurate, lawful and of public interest. A correction request works best with clear evidence of what is true, and it is not certain to succeed.

How can I tell whether a video or message from an executive is real?

Do not rely on the message. The SEC, NASAA and FINRA investor alert warns that AI can clone voices and imitate a company's CEO in video, and advises verifying through contact details found independently rather than links or numbers in the message. Confirm with the company through a channel you already trust before acting on any urgent statement.

How many assistants should be checked in an executive audit?

At least two that search the web and one that answers mostly from training, because they draw on different sources and respond differently to the same question. Repeat the runs on different days, since answers vary between runs. A claim that appears in most runs and on several assistants deserves attention first; one that appears once is worth logging, not chasing.