Does directors' liability make the White House AI accord binding? Gavin Baker's claim, checked against the accord's full text, Delaware case law and liability-insurance terms

① Who said it, and what was said

On September 29, the heads of Google, Anthropic, Meta, OpenAI, xAI and NVIDIA signed a voluntary accord at the White House. The accord carries no penalties. On October 1, Gavin Baker argued that the accord is binding after all. Baker founded Atreides Management, a fund that invests in tech companies. In Baker's view, the force comes from corporate governance: "If they ignore a report from the third party auditors then a bad faith finding from a court is a real possibility." (Gavin Baker, 2026-10-01)

In Baker's account, if directors ignore an outside auditor's report, a court could find bad faith, and the insurers that write directors' liability cover could then refuse to pay. "Bad faith" here is a legal term, and it doesn't mean bad motives. The term means directors knew they were supposed to oversee something and deliberately didn't. Liability policies usually have clauses that exclude this kind of conduct, which is why Baker reasons that insurers could refuse to pay. Baker also says this route matters more in the near term than almost any regulatory move.

How far it spread. Among the voices SecondSource tracks, this was the most repeated line of the week. SecondSource's daily brief and its October 4 deep dive both quoted it.

What to check it against. The accord's full text, Delaware case law and empirical statistics on the directors' duty of oversight, and a Delaware Supreme Court ruling on the bad-faith exclusion in liability insurance.

Which sentence could fall. "More meaningful than regulation in the near term." If none of the signatories has named an auditor and a committee by early December, the route has nothing to run on in the near term.

Why it matters. Liability underwriters are deciding whether their questionnaires should ask about AI audits. Compliance heads who buy frontier models are deciding whether contracts should separately require audit reports. The first thing to come out before December 6 is whether any signatory names an auditor and a committee.

② The receipts

The accord's text asks boards to set up a committee that receives audit reports (primary full text, the hardest evidence). The fourth tier reads: "Designate an independent committee of the board of directors to oversee and receive reports from the teams operating the controls and the internal and external auditors and evaluators" (full text as republished by Zvi Mowshowitz, 2026-09-30). The full text also says plainly that the accord is not law yet: "Over time, it may make sense to codify these steps into laws or regulations." ⚠️ The full text was available for this check only through that one republication. The White House press release is in the American Presidency Project archive. The republished full text doesn't require audit results to be made public, and it doesn't say who the auditors are.

Delaware law really does have this route. The 1996 Caremark case established that directors have a duty of oversight. Directors must make sure the company has a system for watching its major risks; if they don't, shareholders can sue them. In the 2019 Blue Bell case, the Delaware Supreme Court held that failing to build oversight of a mission-critical risk can amount to bad faith. Blue Bell is the situation in which a board never built an oversight system at all. Anthropic and OpenAI are both Delaware public benefit corporations, and shareholders also have a statutory right to inspect board records. SecondSource's October 4 deep dive worked through these cases and statutes from the originals. The mechanism Baker describes is not wishful thinking.

But the route rarely succeeds. An empirical study published in September on Columbia Law School's blog found that about 25.6% of these oversight claims in Delaware courts survived the early stage of the suit without being dismissed. That share hasn't risen since Blue Bell. Getting past that stage only means the case continues; it doesn't mean winning. In the same study, the 28 claims that argued only that directors had ignored warning signs all failed (CLS Blue Sky Blog, 2026-09-02). ⚠️ This is a single study. Our read: Baker's scenario, a committee that exists but ignores a report, is closest to the type in which all 28 claims failed. It is not the Blue Bell situation, where no oversight system existed.

An insurer's refusal to pay waits on a final judgment (the weakest link in this column). A 2021 Delaware Supreme Court ruling held that, under the terms of that policy, only a final adjudication that can no longer be appealed could trigger the exclusion. Whether this kind of clause covers a court's finding of bad faith depends on how each policy is written. The original ruling hasn't been reviewed, so that point can't be confirmed. Cases can also settle before any judgment: the Boeing board-oversight case settled for US$237.5 million, all of it paid by insurers. Our read: because a refusal to pay has to wait for a final judgment, the threat rarely gets a chance to bite in practice. ⚠️ For both the ruling and the Boeing settlement, only summaries on law-firm and industry blogs have been reviewed, not the originals.

The party that moved at the same time was a regulator. On September 30, The Washington Post got a senior official at the US Federal Trade Commission (FTC) to confirm: "The Federal Trade Commission has opened a broad investigation into the safety of artificial intelligence systems made by Anthropic and OpenAI" (The Washington Post, via the Spokesman-Review, 2026-09-30). The FTC is acting under powers it already has, not under the accord. The report doesn't say when the investigation began, and it doesn't tie the investigation to the accord.

Neither precondition is public yet. A search on October 11 turned up no signatory that had named an outside auditor or designated a committee. An empty search doesn't show that none exists, and only that one search was run. Zvi Mowshowitz, whose Substack carried the republished full text, points to a precedent: "OpenAI had exactly this committee, but they either were kept out of the relevant loops or failed to act."

⚠️ Disclosure: Baker holds positions in signatory companies, and Baker's post says outright that he opposes creating a new regulator.

③ The receipts say: unresolved

The claim ruled on, in paraphrase: the accord has real binding force through directors' liability, and in the near term that is more meaningful than almost any regulatory move.

The half that holds: the mechanism is real. This is the mechanism Baker describes. If a board receives audit reports and leaves no record of acting on them, that record could become evidence against the directors when shareholders later inspect the board's records.

The half that doesn't hold up: "more meaningful than regulation" has no receipts. Only shareholders can set the route in motion, and only after something has gone wrong. Every "ignored warning signs" claim in the study lost. Settlements get paid by insurers anyway. The day after the signing, the party reported to have opened an investigation was the FTC, using powers it already had. Of this evidence, the liability-insurance link is the weakest, since the evidence for it is summaries only. This half of the conclusion therefore rests mainly on the litigation statistics and the FTC report.

What would prove this wrong:

• ① A signatory names its auditor and its committee, and underwriters start asking about AI audits: the ruling would lean toward "holds," but not yet far enough to change.
• ② A public report that a committee or an audit finding delayed a model release: the ruling changes to "holds."
• ③ By December 6 (SecondSource's own observation deadline), none of the six has published either, while the FTC investigation has already issued a formal request for information: the "more meaningful than regulation" half changes to "doesn't hold."

How to use it. A vendor's statement that it has signed the accord has backing only when three things come with it: a named auditor, the board committee that receives the reports, and a description of the audit's scope that a buyer can actually obtain. Scope of this column: it rules on Baker's sentence, not on whether the accord should have been signed.

④ Where the claim sits

This draws on an existing tree; no new one is drawn.

The tree records whether the tools that check a model can be trusted when an AI company says the model is safe. The tree shows that the industry has already swapped the checking tool it uses to back releases, and that the new tool has no reading yet under conditions where the model deliberately evades it. The tree bears on this column because Baker's sentence can hold only if a board first receives an audit report with real content. What such a report can reveal depends on which tools the auditor uses and how far the audit reaches. That is why the audit-scope description is one of the three things that, under "How to use it" in ③, back a vendor's statement that it has signed.

⑤ Checking the answers

A new call this issue. By December 6, will any of the six signatories announce, through its own official channels, both a named outside auditor and the board committee that takes on the accord? Our call: no. Announcing only one of the two, or appearing only in media accounts, counts as no. The call settles only on the six companies' official newsrooms and investor-relations pages. If even one company announces both, this call counts as missed, and the ruling leans toward "holds" accordingly (see item ① under "What would prove this wrong" in ③).

Older calls. None of the calls this column opened in September has come due yet.

Settled calls

Would an analyst on CoreWeave's second-quarter call ask how much capacity Nvidia's backstop had absorbed?

① What the call said then: CoreWeave is a cloud company that rents GPU compute to AI companies. The backstop is Nvidia's commitment to buy the compute that CoreWeave can't rent out. The column opened this entry on August 2. The card puts the question as "whether CoreWeave's next earnings call, for Q2 2026 and expected mid-August, gets asked about how much Nvidia backstop capacity it is absorbing, and whether the company gives an auditable number for utilization or absorbed capacity." Utilization here is the share of CoreWeave's compute that is actually rented out. The entry recorded no confidence level at the time.

② The actual reading: the sources named in advance were CoreWeave's Q2 2026 results and the transcript of its August 11 earnings call. The word "backstop" appears nowhere in the transcript, and no analyst in the Q&A asked how much the backstop had absorbed. CoreWeave disclosed neither utilization nor how much Nvidia had absorbed. CoreWeave said only that near-term capacity is "effectively sold out." The transcript is a paid-database version that readers can't get publicly, and SecondSource checked it there on September 30. What is public is CoreWeave's second-quarter results.

③ Ruling, and at which layer: came true, though the result gives the prediction only limited support, for the reason below. The entry set out three outcomes in advance. If analysts asked how much the backstop had absorbed, the first half would hold. If CoreWeave disclosed checkable figures for utilization (the share of compute rented out) or for how much the backstop had absorbed, the second half would hold. If neither happened, the prediction that "the backstop keeps the financials from showing real demand" would come true. This time it was the third outcome. What came true is that prediction. What fell flat is the premise the entry borrowed when it opened: the prediction of Bill Gurley, of the venture firm Benchmark, that analysts would keep asking for eight straight quarters. The first time they should have asked, nobody did. Nobody asking could also simply mean that analysts don't care about the question, so this result supports the prediction only weakly. This call also came due on September 15, and SecondSource checked it only on September 30, 15 days late.

④ Which tracked judgments moved: none yet. The result supports the judgment that the backstop masks the demand signal, but, as ③ says, only weakly. That judgment still awaits reassessment. The recorded fact about Nvidia's backstop terms hasn't changed and needs no update.

⑤ Does the standard change: Yes. One call bundled two questions, so the half that missed could be hidden behind the half that came true. From now on, "will analysts ask" and "will the company disclose" become two separate calls, each with its weight written down in advance, so that the half that misses doesn't get buried.

See the receipts →

Was August's wave of incident disclosures by AI labs mostly old incidents coming to light at once? Can't judge. On the due date, Congress received the replies to its letters, and neither of the two developments that would have overturned the judgment happened. But the entry never wrote down what would count as coming true, so there is no yardstick to judge by. See the receipts →

Will Anthropic's enterprise market share fall to 35% or lower? Postponed. No readable figure exists yet; the call waits for the next enterprise LLM adoption report from the venture firm Menlo Ventures. See the receipts →

All of SecondSource's calls and their verdict dates → the judgment index