Research: The AI-Lab NDA and Non-Disparagement Cluster

July 2, 2026

The AI-lab NDA cluster is not a story about gag orders in the abstract. It is a story about a specific instrument: a perpetual non-disparagement clause attached to vested equity, deployed during off-boarding under time pressure — typically a seven-day signing window inside a sixty-day forfeiture deadline — in a private market with no liquid secondary, against employees whose only path to realizing the equity ran through the issuer’s own tender offers. Every claim below about a real person or organization carries a source; characterizations are attributed to whoever made them, never adopted as our own voice. The contract itself has not been published as a primary document — quoted clauses come via journalists who state they saw it, so quoted clause language is treated as journalism, not as direct primary source.

The mechanism — financial architecture and equity leverage

OpenAI’s compensation model granted profit participation units (“PPUs”) rather than traditional stock — equity-equivalents whose only realization path was through OpenAI-controlled tender offers. Departing employees were presented with:

The underlying reporting is Kelsey Piper’s May 17 and May 22, 2024 Vox coverage, corroborated by Bloomberg, CNBC, Gizmodo, and TechCrunch on the same leaked-document trove (Vox, May 17, 2024; Piper’s May 22, 2024 document thread).

The escalation tactic

When at least one employee retained outside counsel and pushed back on the equity-cancellation threat, OpenAI did not relent. According to Piper’s May 22 thread, OpenAI shifted from “your units are cancelled” to “any ex-employee who doesn’t sign the restrictive release will be excluded from all opportunities to sell their equity” — the units would notionally still exist but could never be liquidated, since OpenAI controlled the only secondary market (Vox / Kelsey Piper, May 22, 2024). The effective economic outcome was identical.

Documents that surfaced executive signatures

Per Vox’s May 22, 2024 document drop, the modified equity-cancellation framework was approved over the signatures of Sam Altman, Jason Kwon (then Chief Strategy Officer), Diane Yoon (then VP People), and Brad Lightcap (COO) (Vox / Kelsey Piper document thread; Gizmodo). The documents show these signatures; Altman had stated on May 18 that he “did not know this was happening” (see the reform timeline below). The record is presented as documents-show-X and Altman-said-Y; no assertion is made that any officer lied or acted with intent to deceive.

Why this combination is unusual

Three structural features distinguish this instrument from ordinary corporate non-disparagement:

  1. Lifetime duration. Most enforceable non-disparagement clauses run for the term of severance benefits or a fixed period of one to five years. OpenAI’s ran forever.
  2. Mutually undisclosed existence. The confidentiality-of-the-release-itself provision meant alumni could not even tell each other, let alone the press, that the instrument existed. This is the specific feature that triggered the Dodd-Frank Rule 21F-17 SEC complaint (see below).
  3. Equity asymmetry in a controlled secondary. Because PPUs trade only through OpenAI-administered tenders, “you keep the equity but cannot sell it” is functionally a clawback while preserving deniability about clawbacks.

That “weaponization” reading of the instrument is our characterization; the underlying mechanics are sourced.

Per-lab inventory — what each contract type says, as of 2026

OpenAI

Anthropic

Google DeepMind

Meta AI

xAI

Inflection AI (Microsoft acquihire, March 2024)

Character.AI (Google licensing deal, August 2024)

The reform timeline (May 2024 → ongoing)

Pre-trigger context

The Vox break

Altman’s statement

The contradicting documents

The retroactive release

Regulatory escalation

Statutory response

The “Right to Warn” letter — full signatory list, June 4, 2024

The primary source is righttowarn.ai — the letter’s own URL, registered for and maintained by the signatories — corroborated by TIME, the Washington Post, CNBC, Axios, Deadline, and Android Authority (all June 4, 2024).

Letter title

“A Right to Warn about Advanced Artificial Intelligence”

The four principles demanded

The verbatim demands of the letter’s signatories:

  1. No disparagement restrictions on risk-related criticism. Companies must not enter or enforce agreements prohibiting risk-related criticism, nor retaliate for it.
  2. Verifiably anonymous reporting channels. A process for current and former employees to raise risk-related concerns to the company’s board, regulators, and independent organizations with appropriate AI expertise.
  3. A culture of open criticism. Support for current and former employees to raise risk-related concerns publicly, with protections for legitimate trade-secret interests.
  4. No retaliation for public sharing after internal failure. Once internal processes for raising concerns have failed, no retaliation for sharing risk-related confidential information.

Named signatories (7)

Anonymous signatories (6)

The letter notes the anonymity explicitly; the underlying identities have not been published. The anonymous-signatory count is from righttowarn.ai’s own page.

Endorsers (named)

Total signatures

13 employees plus 3 endorsers.

Refusers and the price they paid

Daniel Kokotajlo

William Saunders

Leopold Aschenbrenner

Steven Adler

Helen Toner

Geoffrey Irving (cross-lab refuser)

Chris Painter (cross-lab refuser)

Cullen O’Keefe / Evan Hubinger / Gretchen Krueger

Signers who broke anyway, and what happened

Jacob Hilton

Beth Barnes

Jeremy Schlatter

William Saunders

Pattern note

No public case has surfaced of OpenAI actually clawing back vested equity from a signer who later spoke. Altman’s claim — “we have never clawed back anyone’s vested equity” — appears, on the public record as of May 2026, to be accurate as a historical fact. What the mechanism produced was self-silencing, not enforced clawbacks. This is the cluster’s key empirical finding: the instrument worked by deterrence, not by enforcement actions.

Cross-domain comparison — is the AI-lab pattern unusual?

Pentagon / intelligence community

Wall Street / SEC-regulated firms

Big Tech generally

Pharma / biotech (Theranos)

Verdict on whether the AI-lab pattern is unusual

Yes, on three specific dimensions:

  1. Equity clawback as default enforcement. Most NDA regimes require litigation to enforce. OpenAI’s was self-executing — break the clause, lose the equity, no court required.
  2. Confidentiality of the agreement itself. This is unusual outside the intelligence community and is the specific feature that made the clause invisible to public scrutiny until Piper’s reporting.
  3. The private-secondary asymmetry. In a publicly traded company, “you keep the equity but can’t sell it” is meaningless because the holder can sell on the open market. OpenAI’s PPUs trade only through OpenAI-administered tenders. This converted “lose your equity” and “can’t sell your equity” into economically equivalent threats.

The cluster is not unusual in its rhetoric or its legal theory. It is unusual in the mechanism of enforcement.

Unanswered questions

  1. The full Vox document set. Piper’s May 22 X thread shows fragments of the leaked agreements. The full document set has not been published in primary form (the Vox archive does not include scans, only quoted excerpts). The question is whether the full text exists in any public-records venue or remains in journalist hands.
  2. SEC complaint identity. The July 13, 2024 whistleblower complaint to the SEC, brought by Stephen M. Kohn, was filed anonymously. Whether the complaint is still active, has produced a Wells notice, or has resulted in a settlement is not public as of May 2026. Whistleblower attorney work product would be subject to confidentiality regardless.
  3. Mutual-clause beneficiaries. Who exactly is still bound by mutual non-disparagement with OpenAI? Plausible candidates named in reporting include the Anthropic cohort (Dario Amodei, Daniela Amodei, Jack Clark, Sam McCandlish, Tom Brown, Chris Olah), the September 2024 exits (Mira Murati, Bob McGrew, Barret Zoph), and Ilya Sutskever (May 2024 exit, now SSI). None of this is documented; mutual NDAs are bilateral by definition, so these names remain speculative and are not asserted as fact.
  4. xAI structure. Whether xAI’s employment agreements include OpenAI-style equity-tied non-disparagement is not public. Musk’s litigation history against Altman and OpenAI (Musk v. OpenAI / Altman, filed February 2024) creates an incentive for xAI to either avoid the practice (to claim moral superiority in the Musk-versus-Altman narrative) or replicate it (because Musk-controlled entities historically use aggressive confidentiality). Open question.
  5. Anthropic’s current paperwork. No public refuser case. No public statement. No leaked clause. The lab’s reputation as the “safety-first” alternative depends on contrast with OpenAI, but the empirical basis for the contrast on this specific dimension is not documented.
  6. DeepMind’s UK posture. UK employment law (the Employment Rights Act 1996, plus the 2025 Employment Rights Bill) treats post-employment non-disparagement more skeptically than US law. Whether Google DeepMind’s UK-based employees have ever been asked to sign US-style perpetual clauses is not public.
  7. The seven-day window. Piper reports a roughly seven-day signing window. Whether this varied by employee, role, or year is not documented. If it varied, the variation itself would be evidence of negotiability — and therefore of which employees had leverage and which did not.
  8. S.1792 status. Stalled in Senate Judiciary as of May 2026. Whether it will receive a floor vote in the 119th Congress before the next election cycle is the open political question for the cluster’s regulatory arc.
  9. FAR violation. Whether OpenAI’s pre-May-2024 NDA violated the Federal Acquisition Regulation’s anti-gag provision (applicable to federal contractors since 2017) is not adjudicated. A FAR violation could produce contract debarment — a much more material consequence than NLRB or SEC remedies. No public enforcement action has been reported.
  10. The 6 anonymous Right to Warn signers. Their identities, by the letter’s own design, will likely not surface unless they choose to be named. If they have since spoken publicly under their own names without acknowledging the prior anonymous signature, that is a category of fact that exists but is not provable from public sources.

Sourcing principles applied to this dossier

The bottom line

The load-bearing spine of this cluster is documented to a publishable standard: OpenAI’s pre-May-2024 exit paperwork tied a lifetime non-disparagement clause to vested equity realizable only through OpenAI-controlled tenders, with a confidentiality-of-the-agreement provision that hid its own existence — surfaced by Kelsey Piper (Vox, May 17 and 22, 2024), corroborated across Bloomberg, CNBC, and Gizmodo, and confirmed on the record by a signer under oath (Saunders, Senate Judiciary, September 17, 2024). The reform arc (Altman’s May 18 statement, the May 24 retroactive release, the SEC 21F-17 complaint, the Grassley and Warren-Trahan letters, and S.1792) is all sourced to primary or on-record material.

The single most defamation-sensitive claim — that the equity-cancellation framework carried the signatures of Altman, Kwon, Yoon, and Lightcap, contradicting Altman’s “I did not know” statement — is held throughout to the documents-show-X, Altman-said-Y construction. The record does not assert that Altman lied or knew, and it preserves the exculpatory finding (no clawback was ever executed) with equal weight.

The cluster’s payload is the enforcement-by-deterrence mechanism: the instrument silenced without ever having to be enforced, which is why it left almost no litigation trail and why the human cost is measured in self-censorship, not court records. That is our characterization; the underlying facts stand on their own. Publication-grade on the OpenAI spine, appropriately hedged on the per-lab comparanda (the Anthropic, xAI, DeepMind, and Meta clusters, where the honest finding is “not public” and this page says so), with speculation about individual mutual-clause beneficiaries quarantined as speculation.

Get updates on the Evil Robots series

Newsletter essays on AI escape, deception, and the humans who built them.