Research: The Witnesses — AI Whistleblowers and What It Cost Them

July 2, 2026

Every profile below rests on the public record: court records, obituaries, sworn testimony, signed open letters, on-record statements. Where a death is contested, the official ruling is cited first and the family’s dispute follows with sourcing for the dispute itself — never our conclusion about cause. Every characterization of a named person or organization is attributed to whoever made it — a witness’s testimony, a reporter’s finding, a tribunal’s ruling — and never adopted as our own voice. Where an account is contested, both sides are cited and neither is adjudicated.

I. The Quitters

People who left, signed letters, or refused to ship — without filing whistleblower complaints or going to regulators first.

Geoffrey Hinton

Hinton was VP and Engineering Fellow at Google Brain until May 2023, and shared the 2024 Nobel Prize in Physics with John Hopfield for foundational work on neural networks (nobelprize.org). He joined Google around 2013 through its acquisition of DNNresearch and resigned on 1 May 2023; the Nobel was awarded 8 October 2024, with the banquet speech delivered 10 December 2024.

He resigned from Google so that he could speak publicly about existential risk from AI without his comments being read as Google’s position — his stated reason, per his departure interview with The New York Times and the Nobel speech (NYT; CNN). He then used the Nobel Prize platform to issue a public existential-risk warning. The exit was voluntary, at age 75, with no obvious financial cost. In media framing he was described as the “godfather of AI” who turned on his own field (MIT Technology Review); some peers publicly disagreed with him, Yann LeCun most prominently among them. Hinton has expressed personal regret on the record.

His on-record words include:

A line sometimes attributed to Hinton — that he is “being congratulated for the work I suspect will be the last important thing humans ever do” — does not appear in the Nobel banquet text on nobelprize.org and cannot be verified against a primary source. It is not reproduced here as his words.

Ilya Sutskever

Sutskever was co-founder and Chief Scientist of OpenAI, and co-led the Superalignment team with Jan Leike from July 2023 to May 2024 (SSI). He was at OpenAI from 2015 to May 2024. On 17 November 2023 he voted with the board to fire Sam Altman; on 22 November 2023 he signed the staff letter demanding Altman’s reinstatement; on 14 May 2024 he announced his departure; and on 19 June 2024 he founded Safe Superintelligence Inc.

According to New York Times reporting, Sutskever authored an internal memo accusing Altman of a pattern of deception — a characterization carried here as the Times’s reporting, not asserted as fact, and the memo’s text is not public (NYT). Reports of a “52-page memo” come from the Times secondhand; treat its contents as reported, not as Sutskever’s direct quote. He voted to remove Altman, then reversed course under staff pressure (the reversal is the Times’s characterization), spent roughly six months effectively sidelined, and left to start a competing lab with one stated product: a safe superintelligence.

He founded SSI at a reported $5 billion valuation in September 2024, later raised at a reported $32 billion valuation in April 2025 (Financial Times, via TechCrunch); financially, the “cost” is inverted for Sutskever. He has given no substantive public interview since the board crisis, and the silence is itself part of the story.

Jan Leike

Leike co-led OpenAI’s Superalignment team from July 2023 to May 2024, then became an alignment lead at Anthropic (Anthropic). He was at OpenAI from 2021 to 15 May 2024 and joined Anthropic on 28 May 2024. He resigned publicly the same week as Sutskever, then posted a detailed public departure statement on X accusing OpenAI of deprioritizing safety — his own characterization — and joined Anthropic to continue alignment work (X departure thread; CNBC). The move to Anthropic was lateral; the cost was the public-statement risk itself.

From his departure thread on X, 17 May 2024:

Daniel Kokotajlo

Kokotajlo was a governance researcher on OpenAI’s alignment/futures team, from 2022 to April 2024. He resigned in April 2024, refusing to sign the lifetime non-disparagement clause attached to the vested-equity off-boarding agreement, and co-organized the “Right to Warn” open letter published 4 June 2024 (righttowarn.ai). According to Kelsey Piper’s reporting for Vox on the agreement’s terms, the required off-boarding agreement carried a perpetual non-disparagement provision so broad that acknowledging the NDA existed was itself a breach, in exchange for keeping vested equity (Vox). He estimated the personal cost at roughly $1.7 million to $2 million in forfeited equity, per his own public statements and Vox’s reporting; the forfeiture was partially reversed after public pressure on OpenAI, but the original forfeiture stood. He then publicly co-organized 13 current and former OpenAI/DeepMind employees behind the “Right to Warn” letter demanding whistleblower protections (NYT). He is the one party explicitly not under the non-disparagement agreement — he never signed. He subsequently co-authored the AI 2027 forecast project.

Miles Brundage

Brundage was Senior Advisor for AGI Readiness at OpenAI, and previously head of its policy research team, from 2018 to October 2024. He resigned on 23 October 2024 with a detailed public Substack post stating that OpenAI is not ready for AGI and neither is anyone else — his characterization — and the AGI Readiness team was disbanded shortly after, as press reported (CNBC; TechCrunch). He continued AI policy work externally, saying in his statement that “the opportunity cost has become very high.”

From his Substack departure post, 23 October 2024 (primary):

Steven Adler

Adler was a safety researcher at OpenAI for a four-year tenure, from around 2020 to November 2024. He resigned over the pace of AI development — his stated reason — and posted a public departure statement on X on 27 January 2025 (X thread; Fortune). It was one of the bluntest departure statements from any OpenAI alum. The exit was voluntary, with no reported financial penalty.

From his X thread, 27 January 2025:

Leopold Aschenbrenner

Aschenbrenner was a researcher on OpenAI’s Superalignment team, from 2023 to April 2024, when he was fired. This is a contested-firing case, and both accounts are set out here without adjudicating either. By his own account, given on the Dwarkesh Podcast in June 2024, he wrote an internal memo to the OpenAI board warning that the company’s security posture was inadequate against state-actor theft of model weights, and the security memo was “a major reason” he was fired; the memo is not public (Dwarkesh Podcast). OpenAI characterized the firing as being over an information leak; Aschenbrenner says the document at issue was a brainstorm shared for feedback, and OpenAI dismissed the memo’s link to his firing (Transformer). He forfeited equity on terms not made public. He then published Situational Awareness, a 165-page essay arguing AGI arrives by 2027 and superintelligence shortly after, and launched a hedge fund reportedly managing around $1.5 billion by mid-2025 per Fortune — another tracker, BlockBeats, reported much higher figures, and tracker numbers should be treated cautiously (Situational Awareness; Fortune). His financial outcome inverts the “cost” frame, and he now has a financial stake in the AI-acceleration narrative; post-firing statements should be read with that in mind.

William Saunders

Saunders was a Member of Technical Staff on OpenAI’s Alignment / Superalignment team, a three-year tenure from around 2021 to February 2024. He signed the Right to Warn letter in June 2024 and gave sworn testimony to the Senate Judiciary Subcommittee on 17 September 2024 (testimony PDF). He resigned, then went public despite having signed the standard non-disparagement clause — that is, he spoke under threat to his vested equity. By his own testimony, OpenAI had retracted enforcement of the clause after the Vox reporting, but the original signing meant he was speaking at risk. His adverse claims about OpenAI are in privileged sworn testimony; they are carried here as his testimony, not as our finding.

II. The Whistleblowers

People who took information to journalists, regulators, or Congress — past the line of a quiet exit.

Suchir Balaji

Balaji was an engineer and researcher at OpenAI, from around 2020 to August 2024, who worked on data pipelines for what became GPT-4. A New York Times interview with him was published on 23 October 2024. He was found dead at his San Francisco apartment on 26 November 2024, at age 26.

He resigned, then went on the record with the Times alleging that OpenAI’s training pipeline violated US copyright law and that ChatGPT was substantially substituting for the works it was trained on — his allegation; the copyright question is unadjudicated (NYT). He posted a long technical essay on his personal site making the legal argument (suchir.net). He was named in court filings by the Times as a potential witness in the NYT v. OpenAI case.

Two on-record accounts of the cause of death exist, and this page adjudicates neither.

The official ruling: the San Francisco Office of the Chief Medical Examiner ruled the death a single self-inflicted gunshot wound, with the autopsy report released on 14 February 2025. Toxicology showed alcohol, amphetamine, and GHB. The SFPD report says the apartment was dead-bolted from inside and that his computer showed recent searches related to brain anatomy (SF Standard; Fortune).

The family dispute: his parents, Poornima Ramarao and Ramamurthy Balaji, have publicly disavowed the official ruling. They commissioned an independent autopsy whose pathologist reported that the bullet’s downward and slightly left-to-right trajectory through the brainstem was atypical for a self-inflicted wound, and reported a contusion on the back of the head. They have sued San Francisco to release records, and have filed a separate suit, on 22 September 2025, against Alta Laguna LLC and Holland Partner Group — the owner and manager of the apartment complex at 188 Buchanan Street — alleging concealment of evidence; that suit is a pending allegation in active litigation, with the outcome undetermined (SF Standard; KQED). The official ruling is cited first; the family’s experts and lawsuit second; the reader weighs it. No individual is named in connection with the death.

His own pre-death public statements:

Helen Toner

Toner was an OpenAI board member from 2021 to November 2023, and is Director of Strategy and Foundational Research Grants at Georgetown’s Center for Security and Emerging Technology (CSET). She was one of the four board members who voted to fire Altman on 17 November 2023, and departed the board on 22 November 2023 as part of the post-reinstatement reset, alongside Tasha McCauley. After roughly six months of silence, she gave a TED AI Show interview in late May 2024 describing specific instances of what she called Altman’s deception, and followed with sworn testimony to the Senate Judiciary Subcommittee on 17 September 2024 (TED AI Show transcript; testimony PDF). In the immediate aftermath of the firing she was painted, in media framing, as part of a “doomer coup.” Her adverse claims about Altman below are carried as her testimony and her interview statements, never as our finding of fact about Altman (Fortune; CNBC).

Frances Haugen

Haugen was a product manager on Facebook’s Civic Integrity team, from 2019 to May 2021 — a pre-LLM whistleblower whose template later whistleblowers inherited. She filed eight SEC whistleblower complaints in September 2021, testified to the Senate Commerce Subcommittee on 5 October 2021, and was the source for the Wall Street Journal’s “Facebook Files” series in September 2021 (WSJ). She copied tens of thousands of internal Facebook research documents on her way out, gave them to the Journal, then went on the record with 60 Minutes and Congress, protected by SEC whistleblower status. Her reception has run in both directions — heroized in some quarters, dismissed as a particular ideological actor in others. Her characterizations of the company below are carried as her sworn testimony.

From her Senate Commerce Subcommittee testimony, 5 October 2021 (written testimony; full hearing transcript, Washington Post; transcript, Rev):

Timnit Gebru

Gebru was Technical Co-Lead of Google’s Ethical AI team, from around 2018 to December 2020, and founded the Distributed AI Research Institute (DAIR) in December 2021. Her departure is contested: Gebru says she was fired; Google says she resigned. Both accounts are cited; neither is adjudicated. She co-authored “On the Dangers of Stochastic Parrots” (Bender, Gebru, McMillan-Major, Shmitchell, FAccT 2021), warning about the environmental, bias, opacity, and misinformation risks of large language models (ACM). By her account and press reporting, Google’s leadership demanded retraction or the removal of Google-affiliated co-authors; Gebru offered conditions for retraction, Google refused, and — in Google’s account — announced she had resigned, while she and her supporters say she was fired (MIT Technology Review; NYT). Margaret Mitchell, who searched her own Google email for evidence to support Gebru’s account, was terminated two months later. DAIR is funded by the Ford, MacArthur, and Rockefeller foundations (DAIR). She is now one of the most prominent AI-ethics public voices, and by her account has borne recurring online harassment and deplatforming attempts.

Margaret Mitchell

Mitchell was Co-Lead of Google’s Ethical AI team, with Gebru, from around 2016 to February 2021, and is now Chief Ethics Scientist at Hugging Face. As reported by Reuters and The Verge, she searched her own Google email systematically for evidence of discriminatory treatment of Gebru, to support Gebru’s account; Google detected the searches, locked her out for five weeks, and terminated her by email to her personal account (Reuters; The Verge). Her termination announcement on X was two words. She was subsequently hired at Hugging Face.

Laura Nolan

Nolan was a Staff Site Reliability Engineer at Google in Dublin, through 2018, and a tech-worker organizer. She resigned in protest over Project Maven, the Pentagon AI program, founded TechWontBuildIt Dublin, and is now active with the International Committee for Robot Arms Control (ICRAC) and Stop Killer Robots (ICRAC). She was one of about a dozen Googlers who resigned over Maven after the 3,100-signature internal letter failed to stop the contract, and has continued public advocacy against lethal autonomous weapons.

III. The Moderators

The Kenyan, Filipino, and Ghanaian workforce that built the training data and the moderation layer.

Daniel Motaung

Motaung, a South African national, was a content moderator for Facebook at Sama in Nairobi in 2019, and is an active plaintiff against Meta and Sama in Kenyan court; the case’s outcome should not be characterized before judgment. By his and his legal team’s allegation, as reported by TIME, he was fired in 2019 for attempting to organize a union of moderators reviewing the worst Facebook content — beheadings, child abuse, suicide, torture (TIME, Perrigo, February 2022; TIME). He filed a petition against Meta and Sama in Kenyan court in May 2022. The case is still active: the Court of Appeal ruled on 20 September 2024 that Meta can be sued in Kenya (Kenya Law, Court of Appeal ruling; Business & Human Rights Resource Centre), with a further ruling by the Employment & Labour Relations Court in 2025 (Kenya Law). He has described PTSD, per his own public testimony.

Mophat Okinyi

Okinyi was a quality analyst on OpenAI’s content-moderation pipeline at Sama in Nairobi, from 2021 to 2023, and continued advocacy after the Sama contract ended; he was named to the TIME100 AI list in 2024 (TIME100 AI). Per TIME’s investigation, he reviewed up to 700 pieces of text per day describing sexual abuse, bestiality, murder, suicide, and torture, for $1.32 to $2.00 an hour, while OpenAI paid Sama $12.50 an hour per worker (TIME, Perrigo, 18 January 2023). He helped found the Content Moderators Union and co-signed an open letter to President Biden in May 2024. By his account, his marriage ended, he was diagnosed with PTSD, and he faced continued financial precariousness (Tech Worker Community Africa).

Joan Kinyua

Kinyua is founder of the Data Labelers Association of Kenya and a former moderator, organizing since 2023; she co-signed the May 2024 open letter to President Biden. Per BBC and TIME reporting, she organized 140-plus former moderators with PTSD diagnoses, of whom 81% were classified as severe — a figure attributed to that reporting, not restated as an independently verified statistic — and built the first formal data-labeler advocacy organization in Africa (BBC; TIME).

Paul Njoroge

Njoroge is a Boeing 737 MAX widower, adjacent to the AI cluster in that the automated system that killed his family was MCAS — the crash and MCAS’s role are established by the official investigations, and the pattern linkage to the AI cluster is ours. On 10 March 2019, Ethiopian Airlines Flight 302 crashed; he lost his wife Carolyne, their three children (Ryan, 6; Kellie, 4; Rubi, 9 months), and his mother-in-law. He gave sworn testimony to the U.S. House Committee on Transportation and Infrastructure on 17 July 2019 (sworn testimony PDF; NPR). He has been among the most consistent public voices of the 737 MAX families demanding accountability, filed wrongful-death litigation, and continued lobbying through the collapse of the 2024 DOJ deferred-prosecution agreement. Boeing-specific causal claims trace to the NTSB, DOJ filings, and House committee record.

From his sworn House testimony, 17 July 2019:

IV. The Refusers

People who walked away rather than build, often without going public in the moment.

Meredith Whittaker

Whittaker founded Google’s Open Research Group, co-founded the AI Now Institute at NYU in 2017, was a core organizer of the November 2018 Google Walkout, and has been President of the Signal Foundation since September 2022. She was at Google from 2006 to July 2019, resigning by her own account citing retaliation. She organized the 20,000-plus Googler walkout over the company’s $90 million payout to Andy Rubin and over Project Maven, and wrote an open letter on departure documenting what she described as retaliation — being told to “abandon her work” on AI Now; the retaliation characterization is hers (her departure letter, Medium; TechCrunch). She pivoted to running Signal, which makes her one of the few high-profile tech leaders whose product cannot harvest user data even if asked.

Joy Buolamwini

Buolamwini is a researcher at the MIT Media Lab and founder of the Algorithmic Justice League (2016). Her “Gender Shades” paper (2018) documented, in peer-reviewed research, that commercial facial-recognition systems from IBM, Microsoft, and Face++ misclassified darker-skinned women at rates up to 34 percentage points higher than lighter-skinned men (gendershades.org; Gender Shades paper, PMLR). By her own stated stance, she refused offers to commercialize her own systems and built the AJL as a refusal-of-the-default vehicle, staying in academia and advocacy rather than joining industry (AJL). Her work is documented in the Coded Bias documentary (2020) and her book Unmasking AI (2023).

Sneha Revanur

Revanur is founder and president of Encode (formerly Encode Justice), which she founded at age 15 in 2020, and was named to the TIME100 AI list in 2023 (Encode; TIME100). She built youth-led AI-safety advocacy from a campaign against California’s Prop 25 algorithmic-pretrial-risk system, co-sponsored California SB 1047 in 2024, and coordinated celebrity support behind the bill, which Governor Newsom vetoed in September 2024 (Newsweek; background, Wikipedia). She has been positioned in reception as aligned with the AI-safety camp at a moment when that label costs some allies in the AI-ethics community.

Stuart Russell

Russell is a Professor of Computer Science at UC Berkeley, author of Human Compatible (2019) and, with Peter Norvig, of the standard AI textbook Artificial Intelligence: A Modern Approach, and founder of the Center for Human-Compatible Artificial Intelligence (CHAI). He has, by his own stated stance, refused defense-contractor commercialization, and spent years at the UN Convention on Certain Conventional Weapons Group of Governmental Experts arguing for a ban on lethal autonomous weapons (his UN remarks page). He made Slaughterbots with the Future of Life Institute in 2017 to dramatize the threat, and used his textbook to embed alignment material into the field’s first-year curriculum (FLI). Some peers have dismissed him as alarmist; tenured at Berkeley, he is insulated, and the cost has been diplomatic — the LAWS ban has not happened.

From his UN CCW remarks in Geneva, 13 November 2017:

V. The Destroyed

People whose lives the systems ended — legally, financially, or literally.

Robert Williams

Williams is a resident of Farmington Hills, Michigan, an auto-industry worker and father of two. He was arrested on 9 January 2020 in his driveway in front of his wife and two young daughters, then aged 2 and 5, and released after about 30 hours; the ACLU filed a complaint in June 2020, and Williams v. City of Detroit was settled on 28 June 2024, with the Detroit Police Department’s error admitted (ACLU case file; Washington Post, settlement). Williams did nothing — that is the point; the match was wrong, and the case settled. Detroit Police ran a Michigan State Police facial-recognition search against an expired driver’s-license photo using a blurry still from a Shinola watch store; the match was wrong, and per the ACLU complaint the detective wrote it up anyway (NYT, Hill, 24 June 2020; ACLU Michigan). By his account, his daughters’ subsequent fear of police is a lasting injury.

Porcha Woodruff

Woodruff is a resident of Detroit, a nursing student who was an expectant mother at the time. She was arrested on 16 February 2023, at home in front of her two children (ages 6 and 12), while eight months pregnant, and released after eleven hours; the Wayne County Prosecutor’s Office dismissed the case on 6 March 2023 for insufficient evidence, and she filed a federal lawsuit in August 2023 (NYT, Hill, 6 August 2023; Washington Post; Democracy Now). Per the reporting, a facial-recognition match to her 2015 mugshot — taken eight years earlier, when she was not pregnant and looked materially different — was used as the sole basis for the warrant. Her lawsuit was subsequently lost at the trial level, per CBS Detroit (CBS News). By her account, as reported, she had contractions during interrogation and a subsequent dehydration and low-heart-rate diagnosis.

UK Horizon Postmasters

Roughly 900 or more UK subpostmasters were wrongly prosecuted between 1999 and 2015 based on data from the Fujitsu-built Horizon accounting system — a heavily adjudicated public scandal. Horizon was deployed in 1999, prosecutions ran from 1999 to 2015, and the Court of Appeal began overturning convictions in 2021 (Post Office Horizon IT Inquiry, official). The post-Inquiry Statutory Volume 1 report was published in 2025. On the question of 13 suicides, the Inquiry Chair, Sir Wyn Williams, “cannot make a definitive finding” of a causal link but did not rule it out as a “real possibility”; that is his exact adjudication, and it is adjudicated no further here (Computer Weekly). More than 350 subpostmasters died before receiving compensation. Per the Inquiry’s findings, subpostmasters were destroyed by an automated system whose designers and operators knew it was unreliable; the cost was imprisonment, bankruptcy, and suicide. The cross-cluster linkage to the AI cases — a single buggy software system used as the unquestioned source of truth — is ours.

The lead spokesperson is Sir Alan Bates (knighted 2024), founder of the Justice for Subpostmasters Alliance, now the subject of Mr Bates vs The Post Office (ITV, January 2024). Several other named postmasters — Seema Misra, Lee Castleton, Jo Hamilton — are public speakers; families of the deceased are estate-mediated.

Amazon Flex algorithmic firings

Independent-contractor delivery drivers were terminated by automated email after their composite app rating dropped below a threshold, as documented in Bloomberg’s investigation by Spencer Soper, published 28 June 2021, to which Amazon responded on the record (Bloomberg; AI Incident Database). Stephen Normandin — a named driver, age 63, an Army veteran with four years on the platform — is the most-quoted of about 15 drivers Bloomberg interviewed. Per that reporting, the contractor classification denies drivers most labor protections, and appeals require a $200 arbitration fee that is widely reported as ineffective.

The Through-Line

The through-line these witnesses share is not ideology, heroism, or a single technology. It is that each of them, at a cost, refused to let an institution’s account of reality stand in for their own — and the record shows what that refusal cost, in the person’s own words or in a sworn or settled document, and no further.

Every adverse claim about a named third party — Altman, OpenAI, Google, Meta, Boeing, the Post Office, Fujitsu, the Detroit Police — is carried as the witness’s testimony, the reporter’s finding, or the tribunal’s ruling, never as an assertion. Where an account is contested — Aschenbrenner’s firing, Gebru’s fired-versus-resigned — both sides are cited and neither is adjudicated. The dead are not editorialized: Balaji’s death is stated once, the Chief Medical Examiner’s ruling and the family’s dispute are cited side by side, cause is adjudicated by neither, and no person is named in connection with the death. The Horizon 13-suicide question is quoted in Sir Wyn Williams’s exact measured language and left there. Sworn and settled sourcing — Senate testimony, a settled case with admitted error, a statutory inquiry’s findings — outranks the rest; podcast and social-media statements are used as the speakers’ own words. And where a witness’s own litigation later failed, as Woodruff’s did at the trial level, the record says so; the point is the pattern, not a scoreboard.

Get updates on the Evil Robots series

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