A register, updated 9 September 2026
Leaving the labs
People keep leaving the companies building frontier AI, and some of them say why. This page records who left and what they said, and how good the evidence is that the two are connected.
The register
84 records, 14 where the person said the concern was why they left, 40 that say nothing about why.
- Concern-linked14
- Said afterwards8
- Mixed9
- Group7
- Contested6
- Not established40
What these six words mean, and what is in the register
The subject is the race to build systems more capable than the people building them, and whether the companies running that race can be trusted to keep their word about it. That is a narrower question than AI ethics and a wider one than extinction risk. Nobody has to say the word extinction to belong here. Someone can lose confidence that their employer will keep its commitments, or support the safety work it promised, or tell the truth about what it is building. Any of those is about whether the move to very capable AI goes well.
The unit is one person and one transition. Someone who left OpenAI for Anthropic and then left Anthropic has two records and is still one person. The people who left OpenAI to found Anthropic have a record each and are one episode. Keep that distinction, because if you count rows and call each one a witness, a single disagreement looks like a movement.
Each record answers three questions separately, because they are three questions.
- What was it about. The stated reason, in the terms the person or the reporting used.
- How strongly is that tied to the exit. The account, below, running from a first-person statement down to nothing established.
- What kind of document says so. A signed declaration, a resignation thread, an interview, a tracker profile somebody else compiled.
- Concern-linked14
- The person said the concern was why they left, or reporting at the time said so. This is the strongest claim the register makes.
- Said afterwards8
- The warning is on the record and the departure is on the record, and nothing links them. Someone can leave for an ordinary reason and be right about the danger a year later.
- Mixed9
- The person gave a reason and it was not an accusation. Wanting to do the work somewhere else is a different claim from saying the work could not be done here.
- Group7
- The reason belongs to a group that left together, not to the person. A group leaving in one episode is one story, not one witness per row.
- Contested6
- The person and the employer describe the same event differently and nobody has settled it. Both versions are in the row.
- Not established40
- A departure that belongs in the picture, with no public account of why it happened. This is by far the largest group.
A line from a signed court declaration is not the same kind of evidence as a line an outlet copied from a thread that is no longer reachable, and a warning given a year after someone left is not an explanation of why they left. That is what the two copper lines at the foot of every open record are for.
84 of 84 records
| Details | |||||
|---|---|---|---|---|---|
| Paul Christiano | Elsewhere | Ecosystem | Not established | ||
Joined the Foundation board and Safety and Security Committee. Not a new employee resignation or a return to ordinary employment. Official statement specifies recusal arrangements.
| |||||
| Jacob Coxon (carries a quotation) | Anthropic | Out | Concern-linked | ||
Rejected the race toward self-improving AI and criticized both his recent employers.
Original X page inaccessible; short quotation corroborated in reporting. Do not treat his two employers as two independent witnesses.
| |||||
| René Mayrhofer | Google DeepMind | Not established | Not established | ||
Profile attributes resignation to opposition to military-AI arrangements. Reported notice May 18 and planned final day August 31. This is broader Google employment, not a verified DeepMind role. Completion needs confirmation.
| |||||
| Joe Benton | Anthropic | Out | Not established | ||
End of role and later warnings are documented separately; departure motive unresolved. Do not convert agreement with Coxon after leaving into proof of why Benton left.
| |||||
| Joshua Achiam (carries a quotation) | OpenAI | Not established | Mixed | ||
Announced an intention to pursue the mission from outside a frontier lab.
July departure is reported in later coverage. His explanation stressed mission work outside, not a specific grievance.
| |||||
| Johannes Heidecke | OpenAI | Not established | Not established | ||
Planned departure reported; no personal public explanation captured. OpenAI described integrating safety with research; completed departure date is not independently confirmed.
| |||||
| Alex Turner (carries a quotation) | Google DeepMind | Out | Concern-linked | ||
Said he could not remain after disagreements over the Pentagon arrangement and safeguards.
Google DeepMind, not OpenAI. His own post says no public information links Jeff Dean’s later departure to this controversy. The employer said Google defended its safeguards, including its position on mass surveillance and autonomous weapons without appropriate human oversight.
| |||||
| Aleksander Madry | OpenAI | Not established | Not established | ||
Departure reported; personal cause not established. His removal from Preparedness leadership in July 2024 was a separate role change, not this 2026 exit. The 2023 brief departure during the board crisis is not assessed here.
| |||||
| Caitlin Kalinowski (carries a quotation) | OpenAI | Not established | Concern-linked | ||
Objected to the process and safeguards surrounding the Pentagon agreement.
Reuters corrected the role from robotics head to hardware leader. The employer said OpenAI said its agreement preserved red lines concerning domestic surveillance and autonomous weapons.
| |||||
| Mrinank Sharma (carries a quotation) | Anthropic | Out | Mixed | ||
Letter describes values tensions, completed work, and a change in personal direction.
Not an unqualified allegation that Anthropic uniquely disregards safety. Avoid calling him chief of all AI safety.
| |||||
| Zoë Hitzig | OpenAI | Not established | Concern-linked | ||
Linked her resignation to ChatGPT advertising and concerns about monetizing sensitive interactions. NYT essay: “OpenAI Is Making the Mistakes Facebook Made. I Quit.” Original essay and exact resignation date still need capture; no original-text quotation supplied.
| |||||
| Hieu Pham | OpenAI | Not established | Not established | ||
Reportedly left citing burnout. Do not equate the word existential with extinction, or proximity in time with a causal resignation account.
| |||||
| Harsh Mehta | Anthropic | Not established | Not established | ||
Announced a new direction; no concern-linked departure reason established. Tracker correction warns against inferring a safety motive from timing alongside other departures.
| |||||
| Behnam Neyshabur | Anthropic | Not established | Not established | ||
Announced a new direction; no concern-linked departure reason established. Tracker correction warns against inferring a safety motive from timing alongside other departures.
| |||||
| Ryan Beiermeister | OpenAI | Not established | Contested | ||
Termination followed objections to adult-mode plans; causality is not established. Reported company reason was a discrimination allegation, not refusal to develop AI. Do not assert proven retaliation. The employer said She denied discrimination; OpenAI denied a link to issues she raised.
| |||||
| Jacob Coxon | OpenAI | Another lab | Not established | ||
Later criticism does not itself establish why he made this earlier move. The supplied Grok text gives both July 2026 and a four-month Anthropic tenure ending in September. Those cannot both be right. The earlier move is not counted as a separate protest or independent witness.
| |||||
| Andrea Vallone | OpenAI | Another lab | Not established | ||
Left without a public criticism in the checked report. No individual AGI-race or leadership-integrity cause verified; no inference from mental-health lawsuits.
| |||||
| Tom Cunningham | OpenAI | Out | Concern-linked | ||
WIRED reported a departing message about tension between independent research and advocacy. He declined comment to the outlet. Treat the message as reported, not a public first-person resignation account. The employer said OpenAI defended the rigor and expanding scope of its research.
| |||||
| Devin Kim | xAI | Out | Contested | ||
Lawsuit alleges safety-related retaliation. Complaint allegations are not a court finding. Full complaint and any subsequent response need verification. The employer said No response established in the reviewed account.
| |||||
| Rosie Campbell (carries a quotation) | OpenAI | Not established | Concern-linked | ||
Describes team dissolution and concerns about whether the safety approach is sufficient.
Mixed organizational and substantive reasons; do not turn the essay’s publication date into an exact last working day.
| |||||
| Richard Ngo | OpenAI | Out | Concern-linked | ||
Says the main catalyst was Miles’ team dissolving, with less research freedom if he stayed. His own retrospective cautions against giving him much credit for this exit. The supplied discussion does not establish the original departure date.
| |||||
| Lilian Weng | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest.
| |||||
| Steven Adler | OpenAI | Out | Said afterwards | ||
Warning is documented; causal departure explanation remains unestablished here. January 2025 statement came after the November departure. Do not infer motive only from chronology.
| |||||
| Miles Brundage | OpenAI | Out | Mixed | ||
Sought greater impact and independence outside the company. AGI Readiness team was dissolved. External-impact rationale should not automatically be relabeled a protest. The employer said OpenAI publicly supported his next work and offered resources without publication restrictions.
| |||||
| Mira Murati (carries a quotation) | OpenAI | Another lab | Said afterwards | ||
Public exit emphasized personal exploration; separate reporting describes earlier AGI-leadership distrust.
Do not turn reported 2023 remarks into a September 2024 resignation quote. The later company is not presented as an independent nonprofit. The employer said Altman characterized the departures as independent and amicable in contemporaneous AP coverage.
| |||||
| John Schulman | OpenAI | Another lab | Mixed | ||
Moved to focus on hands-on alignment research. Explicitly denied leaving for lack of alignment support at OpenAI.
| |||||
| Jonathan Uesato | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest.
| |||||
| Jan Hendrik Kirchner | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest.
| |||||
| Harri Edwards | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest.
| |||||
| Girish Sastry | OpenAI | Out | Not established | ||
No personal departure explanation captured. Amicus participation proves expressed governance concerns later, not the reason for this departure.
| |||||
| Yuri Burda | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest.
| |||||
| Steven Bills | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest.
| |||||
| Jeffrey (Jeff) Wu | OpenAI | Another lab | Said afterwards | ||
Individual departure motive not captured. Do not infer the exact sequence date from the April 2025 signatory list.
| |||||
| Collin Burns | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest. OpenAI Files says mid-2021, conflicting with his documented presence in May 2024; use the historical 2024 tracker with a visible qualification.
| |||||
| Todor Markov (carries a quotation) | OpenAI | Another lab | Concern-linked | ||
His signed declaration explicitly connects departure to lost trust after the nondisparagement controversy.
Signed April 2025 declaration explains his 2024 exit and the nondisparagement/equity controversy. This is testimony, not a judicial finding; current employer then was Anthropic. The employer said Separate company-response source is needed for his specific account.
| |||||
| Ashyana-Jasmine Kachra | OpenAI | Not established | Not established | ||
Departure motive not established. Tracker lists ambiguous motives. Role and timing do not establish a safety protest.
| |||||
| Carroll Wainwright | OpenAI | Not established | Concern-linked | ||
Existing tracker attributes the exit to diminished confidence in OpenAI’s mission and structure. Former employment and Right to Warn signature verified. Individual causal explanation remains secondary-only; do not use tracker prose as a verbatim personal quote. Legal filing uses Carrol.
| |||||
| Jan Leike (carries a quotation) | OpenAI | Another lab | Concern-linked | ||
Described disagreement over priorities and inadequate support for safety research.
May 17 statement said yesterday was his last day. Reports of the decision and the final day differ. The employer said Sam Altman publicly acknowledged that more safety work remained.
| |||||
| Ilya Sutskever | OpenAI | Founded a lab | Said afterwards | ||
Announced a new personal project; no direct safety-protest cause established here. Karen Hao reports his concerns in 2023. That is not his May 2024 resignation explanation; do not label a confirmed protest. He declined comment in the cited report.
| |||||
| Gretchen Krueger | OpenAI | Out | Concern-linked | ||
Public thread reportedly connected her exit to safety and accountability concerns. Relevant governance/integrity departure even though the statement also discusses other social harms. Reproduced statements do not independently verify every alleged internal failure.
| |||||
| Sherry Lachman | OpenAI | Not established | Not established | ||
Departure reported; no individual safety-motivated cause established. Track separately from technical Superalignment departures. AI Lab Watch treats these as non-safety executive exits.
| |||||
| Chris Clark | OpenAI | Not established | Not established | ||
Departure reported; no individual safety-motivated cause established. Track separately from technical Superalignment departures. AI Lab Watch treats these as non-safety executive exits.
| |||||
| Pavel Izmailov | OpenAI | Not established | Contested | ||
No individual safety-motivated cause verified in this edition. Fired in April 2024, before the May Superalignment dissolution. Do not attribute Aschenbrenner’s account to Izmailov or describe the firing as following that dissolution. The employer said Contemporaneous reporting cited alleged leaks; individual response not captured.
| |||||
| Leopold Aschenbrenner | OpenAI | Not established | Contested | ||
Alleged that security concerns contributed to his firing; the reported company rationale involved information sharing. Contested account. Security advocacy, leak allegations, and later political views are distinct questions. The employer said Company-side reporting described alleged information leaks; he disputed that characterization.
| |||||
| Daniel Kokotajlo | OpenAI | Out | Concern-linked | ||
Lost confidence in responsible AGI decision-making. Refused exit nondisparagement terms and expected an equity loss. Later evidence supports retention/restoration; do not say he ultimately lost millions. The employer said OpenAI walked back the equity-linked nondisparagement policy. The cited interview includes its defense of its safety approach.
| |||||
| Cullen O’Keefe | OpenAI | Not established | Not established | ||
AI Lab Watch places this outside its safety/integrity-caused list. The tracker explicitly treats this as a normal-reason exit. Do not infer hostility from the surrounding departure wave.
| |||||
| Ryan Lowe | OpenAI | Not established | Not established | ||
AI Lab Watch places this outside its safety/integrity-caused list. The tracker explicitly treats this as a normal-reason exit. Do not infer hostility from the surrounding departure wave.
| |||||
| William Saunders (carries a quotation) | OpenAI | Not established | Concern-linked | ||
Said concerns raised inside OpenAI were not adequately addressed; later advocated external accountability.
Later interview is not the resignation date. Signed exit terms; subsequent assurances are not identical to guaranteed liquidity.
| |||||
| Scott Aaronson | OpenAI | Out | Said afterwards | ||
End of OpenAI stint; later AGI-safety criticism does not prove why it ended. The directory gives mid-2024. Whether this was simply an expected end to a visiting period is not established here; not a verified protest.
| |||||
| Anish Tondwalkar | OpenAI | Not established | Not established | ||
Departure motive not established. No individual AGI-risk view or safety-related departure cause verified. Kept outside default view pending stronger evidence.
| |||||
| Tasha McCauley | OpenAI | Not established | Contested | ||
Board membership ended in the 2023 governance crisis; subsequent public accounts challenge self-governance. Nonemployee governance event, not an ordinary employee resignation. Competing accounts of the crisis must be retained. The employer said The succeeding board cited its review as not finding safety or deception to be the cause and defended Altman’s cooperation; former directors disputed the adequacy of self-governance.
| |||||
| Helen Toner | OpenAI | Not established | Contested | ||
Board membership ended in the 2023 governance crisis; subsequent public accounts challenge self-governance. Nonemployee governance event, not an ordinary employee resignation. Competing accounts of the crisis must be retained. The employer said The succeeding board cited its review as not finding safety or deception to be the cause and defended Altman’s cooperation; former directors disputed the adequacy of self-governance.
| |||||
| Jade Leung | OpenAI | Out | Mixed | ||
Moved to help build a public institution for assessing frontier AI risk. Public-interest institutional move, not a documented allegation that OpenAI was uniquely unsafe. AISI later renamed AI Security Institute.
| |||||
| El Mahdi El Mhamdi | Google DeepMind | Not established | Not established | ||
Reportedly left after conflict over publishing work on model vulnerabilities. Original NYT reporting is linked by the index; fuller primary account and Google response remain to be captured.
| |||||
| Jacob Hilton | OpenAI | Out | Said afterwards | ||
No personal departure explanation captured. ARC leadership is verified. Do not call Hilton a co-founder on this evidence.
| |||||
| Geoffrey Hinton | Google DeepMind | Out | Mixed | ||
Left to discuss AI risks without weighing effects on Google. Explicitly distinguished this from criticism of Google’s conduct at the time. Do not backdate later forecasts or later criticism to 2023.
| |||||
| Jeffrey Ladish | Anthropic | Out | Mixed | ||
Described leaving to do research useful for public control of AI. Employment vs consultancy distinction remains important; excluded from strict employee counts.
| |||||
| Beth Barnes | OpenAI | Out | Mixed | ||
Sought an external setting for evaluation work and more freedom to address uncomfortable questions. METR lineage runs through ARC Evals. Do not imply METR was already a separate organization at the moment of departure.
| |||||
| Tom Henighan | OpenAI | Another lab | Not established | ||
Early OpenAI-to-Anthropic transition; individual reason unverified. Early staff / cohort member, not a verified co-founder designation.
| |||||
| Paul Christiano | OpenAI | Out | Not established | ||
Departure and subsequent ARC founding are documented; a protest motive is not established. Not all safety institution-building implies a rupture with the former employer. Later governance return recorded separately.
| |||||
| Nicholas Joseph | OpenAI | Another lab | Not established | ||
Early OpenAI-to-Anthropic transition; individual reason unverified. Early staff / cohort member, not a verified co-founder designation.
| |||||
| Daniel Ziegler | OpenAI | Not established | Not established | ||
No personal departure explanation captured. The letter confirms former OpenAI employment, not motive, exact date, or all later affiliations.
| |||||
| Amanda Askell | OpenAI | Another lab | Not established | ||
Early OpenAI-to-Anthropic transition; individual reason unverified. Early staff / cohort member, not a verified co-founder designation.
| |||||
| Tom Brown | OpenAI | Founded a lab | Group | ||
Joined the Anthropic founding split; personal causal account not established here. Founding / career affiliation is distinct from individual dissent. Several roles and years are directory-reported; exact employment terms are not assumed.
| |||||
| Sam McCandlish | OpenAI | Founded a lab | Group | ||
Joined the Anthropic founding split; personal causal account not established here. Founding / career affiliation is distinct from individual dissent. Several roles and years are directory-reported; exact employment terms are not assumed.
| |||||
| Richard Ngo | Google DeepMind | Out | Concern-linked | ||
Left to pursue conceptual research at Cambridge. Distinct from his later OpenAI departure. He says this earlier choice involved less accumulated wealth and prestige.
| |||||
| Nisan Stiennon | OpenAI | Not established | Said afterwards | ||
Departure cause unestablished; later objected to surrendering nonprofit control. The 2025 quotation is reproduced by OpenAI Files from CNBC; original CNBC access failed. It is not a 2020 resignation statement.
| |||||
| Jack Clark | OpenAI | Founded a lab | Group | ||
Joined the Anthropic founding split; personal causal account not established here. Founding / career affiliation is distinct from individual dissent. Several roles and years are directory-reported; exact employment terms are not assumed.
| |||||
| Dario Amodei | OpenAI | Founded a lab | Mixed | ||
Later interview describes loss of trust and a different vision for developing AI. Founder-level account; do not mechanically attribute the same motive to every colleague who joined the new company.
| |||||
| Daniela Amodei | OpenAI | Founded a lab | Group | ||
Joined the Anthropic founding split; personal causal account not established here. Founding / career affiliation is distinct from individual dissent. Several roles and years are directory-reported; exact employment terms are not assumed.
| |||||
| Chris Olah | OpenAI | Founded a lab | Group | ||
Joined the Anthropic founding split; personal causal account not established here. Founding / career affiliation is distinct from individual dissent. Several roles and years are directory-reported; exact employment terms are not assumed.
| |||||
| Benjamin Mann | OpenAI | Founded a lab | Group | ||
Joined the Anthropic founding split; personal causal account not established here. Founding / career affiliation is distinct from individual dissent. Several roles and years are directory-reported; exact employment terms are not assumed.
| |||||
| Page Hedley | OpenAI | Out | Said afterwards | ||
No personal departure explanation captured. Intervening CSET, Longview and Forecasting Research Institute roles mean this is not necessarily a direct OpenAI-to-Guidelight move.
| |||||
| Geoffrey Irving | OpenAI | Another lab | Not established | ||
No personal explanation for this transition captured. This career-history edge should not be counted as a documented safety protest.
| |||||
| Yonadav (Yo) Shavit | OpenAI | Ecosystem | Not established | ||
Foundation affiliation documented; individual transfer rationale not established. Keep in the institutional layer, not protest counts. The supplied Grok draft claims a May 26, 2026 start and a resignation quotation without an original post URL. Those remain unverified leads; increased outspokenness was not measured.
| |||||
| Ramana Kumar | Google DeepMind | Not established | Not established | ||
No personal departure explanation captured. Do not count a letter signature as a safety-motivated resignation.
| |||||
| Neil Chowdhury | OpenAI | Not established | Not established | ||
Departure motive not established by the amicus filing. Signatory list establishes former affiliation, not an individual account.
| |||||
| Neel Nanda | Anthropic | Another lab | Not established | ||
No personal departure explanation captured. Not a September 2026 DeepMind resignation. His current-employee commentary in the supplied discussion is a separate fact.
| |||||
| Mati Roy | OpenAI | Out | Not established | ||
Former affiliation stated in personal biography; exit reasons not supplied. Continued work on model-training/evaluation environments; do not describe this as exiting all capabilities development.
| |||||
| Mati Roy | xAI | Out | Not established | ||
Former affiliation stated in personal biography; exit reasons not supplied. Continued work on model-training/evaluation environments; do not describe this as exiting all capabilities development.
| |||||
| Jeffrey (Jeff) Wu | Anthropic | Out | Not established | ||
Current biography records independent work, not a reason for departure. Distinct event from OpenAI departure, not another independent witness.
| |||||
| Jared Kaplan | OpenAI | Founded a lab | Group | ||
Joined the Anthropic founding split; personal causal account not established here. Founding / career affiliation is distinct from individual dissent. Several roles and years are directory-reported; exact employment terms are not assumed.
| |||||
| Geoffrey Irving | Google DeepMind | Out | Not established | ||
Career history records later AISI leadership; no protest explanation captured. His later AISI departure is a different event, with personal/family reasons described on his website.
| |||||
| David Duvenaud | Anthropic | Out | Not established | ||
No personal departure explanation captured. A fixed-term academic stint is not inherently a protest resignation.
| |||||
What the shape of it shows
Of the 84 records, 31 are from 2024. That one year covers the collapse of OpenAI’s superalignment team, the resignations that followed, the fight over the exit agreements, and the Right to Warn letter. No other year comes close, including 2026.
OpenAI accounts for 65 of the records. Some of that is a fact about OpenAI. A good deal of it is a fact about the three trackers this was built from, which were compiled to cover OpenAI and cover nobody else as closely. Anthropic accounts for 9 and Google DeepMind for 7. I do not think those two numbers are right.
There are 9 people who left one lab to start another, and 12 who left one lab for an existing one. Both groups are filed apart from the people who left the industry, because a departure that ends at a competing frontier lab changes who someone works for and not what is being built. Almost all of the 2020 bar above is the founding of Anthropic. In September 2026 an Anthropic researcher who had come there from OpenAI resigned and said that neither company was acting responsibly. Those two events are six years apart in the same list.
The account, against where they went
| The account | Out | Another lab | Founded a lab | Ecosystem | Not established | All |
|---|---|---|---|---|---|---|
| Concern-linked | 7 | 2 | 5 | 14 | ||
| Said afterwards | 4 | 2 | 1 | 1 | 8 | |
| Mixed | 6 | 1 | 1 | 1 | 9 | |
| Group | 7 | 7 | ||||
| Contested | 1 | 5 | 6 | |||
| Not established | 8 | 7 | 2 | 23 | 40 |
Every number is a link into the register, filtered to that slice. Reading down the first column: of the 14 people who said the concern was why they left, seven left frontier AI, two moved to another lab, and for five the destination is not established.
10 of the records carry a direct quotation, and that is all of them. If you want people inside these companies in their own words, filter the register to the ones with a quotation and press Open all showing. Each one is printed with the date it was said and with the kind of document it came out of.
Built by people in this register
This is the strongest argument for keeping a register like this, and it is mine. I made it in September 2026, in a comment on Kabir Kumar’s Dear God, Please Don’t Resign In Protest, which argues that someone who objects should stop working and make the company fire them rather than quit. I do not find that convincing. Many of the organizations now doing safety work from outside exist because someone left a lab and used the credibility they took with them to raise money.
AI Futures Project
Daniel KokotajloLeads / co-founder
AVERI
Miles BrundageExecutive director
Guidelight
Steven Adler; Page HedleyCo-founders
ARC Evals → METR
Beth BarnesFounder / director
ARC
Paul ChristianoFounder
ARC
Jacob HiltonVice president of research
UK AISI → Resolution
Geoffrey IrvingFormer AISI chief scientist; Resolution co-founder / chief scientist
Palisade Research
Jeffrey LadishLeadership / external research
Evitable
Gretchen KruegerVolunteer support in cited description
Center for AI Safety
Devin KimPresident appointed June 2026
Anthropic
Dario AmodeiCo-founder / executive
Safe Superintelligence
Ilya SutskeverCo-founder
OpenAI Foundation
Yonadav (Yo) ShavitAffiliation verified; role unresolved
Guidelight
Girish SastryResearcher
AVERI
Jeffrey (Jeff) Wu; David DuvenaudAdvisers
UK AISI
Jade LeungCTO after OpenAI
That list is the case for leaving. The organizations on it do most of the independent evaluation and forecasting work that exists today, and almost none of them existed before. It is also the problem. Someone who founds an organization after leaving a lab starts with credibility a first-time founder does not have, and can raise money that a person with the same ideas and no lab on their record cannot. That is a fact about how funding works, not a fact about who was right.
What a resignation is evidence of
A resignation is strong evidence about conditions inside a company. Someone who was in the room decided that what they saw was bad enough to leave over. That is the reason to keep this register.
It is much weaker evidence about the person. Almost nobody hears about a resignation unless the person had a job that people already paid attention to. You get the audience of a former frontier lab researcher by having been a frontier lab researcher. So the people best placed to be believed when they warn about this are the people who took the job.
Jan Kulveit put this to Richard Ngo directly, in a comment on What just happened? Pragmatism and Pessimization. His first point is that nobody gets credit for a step they did not take. His research group worked out several capability ideas early and decided not to write them up. The version of the group that had published them would have been larger, richer and better known. His second point is harder. Daniel Kokotajlo is the standard example of someone who behaved well. What he did was join OpenAI, leave it in a way that cost him, and then use the standing of a former OpenAI researcher to warn people. A more careful person who never joined has no standing, and nobody names them as an example.
John Wentworth’s reply is that the argument is missing a mood: he could make ten times his income on a capabilities job, and what he wants much more than the money is to not die. Kulveit answers that this is not the failure mode. Almost nobody thinks they are trading the world for a salary. What happens is that the incentive quietly moves what people believe, and the smarter the person, the better the justification.
Ngo agrees that people should be more skeptical of him for making the argument from a position of prestige and money he got at OpenAI, and says most of the value of having been there was learning to handle those dynamics by being subjected to them, which he does not think is reliably good for anyone else.
I do not think this cuts against the register. It cuts against one use of it. Reading a list of departures as a ranking of integrity gets the causation backwards, because the list can only contain people who were inside. What it is good for is the other thing: evidence about what the people with the most information believe, and how often the institutions running this lose the people they hired to watch them.
The people who never joined
The register cannot see them, so I will say it here. There are people who turned down the offer, people who declined to build the skills that would have made the offer come, and people who spent a decade on unglamorous oversight, evaluation and policy work that no lab was ever going to bid for. None of them produces a resignation thread. None of them ends up in a database like this one, and that absence is the single largest thing wrong with it.
I am one of them, which is why I notice. I was invited to apply to two roles at Anthropic and said no both times, and over the years I deliberately did not build the skill set the labs pay the most for. Part of that was that the labs are going to do that work whether or not I help. Part of it was that I did not want to be attractive to them, because I wanted to stay able to say what I think about them in public. I have written the same thing in my own funding application, so it is not a position I invented for this page.
I am not claiming that was the better choice. Kulveit’s point applies to me too, and the honest version is that I do not know what I gave up or what I would have done differently on the inside. What I am claiming is narrower: that a decision not to help is a real decision, that it is invisible by construction, and that a register of exits will systematically undercount the people who were never in the building.
Who is not in here
Five kinds of absence, in roughly the order they matter.
- Quiet exits. Someone who signed something on the way out, or who simply did not want a public fight, is invisible here. There is no way to estimate how many that is, and the fact that the 2024 records exist at all is partly because one cohort refused to sign.
- People still inside who agree. They are the largest group and they are not a group this can count.
- Everyone outside the English-language, mostly American record. The trackers this was assembled from watch a handful of American companies. Departures from labs elsewhere are essentially absent.
- The people who never joined. Above.
- Departures over near-term AI ethics. Bias, copyright, labour practice and research freedom are real subjects and none of them is this one. A departure belongs here when the checked account connects it to building systems more capable than the people building them, and most do not.
How this was built
The rows were assembled from three public trackers and then attached to primary statements and reporting where those exist. The trackers are AI Lab Watch’s OpenAI departures page, which is a very good record of 2024 and says on its own site that maintenance stopped in September 2025, the OpenAI Files former employee directory, and Ethical AI Departures, which covers a deliberately wider subject than this page does. None of their totals is this page’s total and none of them should be quoted as one.
In 15 records the only source is a roster. Those rows say so, and they name the roster rather than hiding behind it. Nothing in this register has been independently re-reported, and a claim about a named person that rests on one secondary profile should be treated as a lead rather than as a finding. If you are about to quote one of these in something that matters, open the source first.
The assembly used a language model, which is worth saying out loud on a page about trusting institutions. The value it added was coverage: reading three rosters, matching the names across them, and noticing which ones had no record anywhere. The thing it is bad at is exactly the thing this register is for, which is refusing to turn a plausible reading into a fact. That is why the evidence columns are on every row and why the corrections below were kept instead of quietly resolved.
Where the sources disagreed
Each of these is a place where a convenient reading was available and was not taken. They are here because a register that shows only its conclusions is asking to be trusted rather than checked.
The 12 places two sources disagreed
Scope
The first pass mixed near-term AI ethics departures into an AGI-focused register.
Near-term-only cases were moved out and named in the exclusions rather than dropped.
AI Lab Watch freshness
Website says maintenance stopped September 2025.
Historical source only; not a current live tracker.
Ethical tracker total
The tracker holds 69 profiles, including context-only entries and unadjudicated allegations.
That total is not relabelled as a count of AGI-safety resignations.
Coxon earlier transition
July 2026 departure and four months at Anthropic before September are inconsistent.
Remove exact month pending reconciliation; retain one person / two events.
Shavit exact date
The only source for the exact day and the quotation was an AI-generated draft with no link to an original post.
The transition is kept. The exact day and the quotation are not.
Collin Burns date
OpenAI Files says mid-2021; he is still described as at OpenAI in May 2024.
Use 2024 historical tracker with source qualification.
Pavel Izmailov chronology
April dismissal precedes the May team dissolution.
The April firing is not explained as following the May dissolution, and Leopold Aschenbrenner’s stated motive is not borrowed for it.
Kokotajlo equity
Threatened or expected loss is not a permanent realized forfeiture.
Retain risk-to-equity distinction from first edition.
Cohort motives
Sharing an employer / departure month does not establish common motives.
A founding split and a team cohort share one episode, and each person keeps their own evidence fields.
Mira / Ilya chronology
Reported 2023 AGI-leadership concern is not a 2024 resignation statement.
Include AGI relevance, but code the causal exit link as unconfirmed.
Role removal vs departure
Madry’s 2024 role removal is separate from his reported May 2026 exit.
Do not count the role removal as a completed company exit.
Hieu Pham wording
Existential-threat language may refer to displacement; burnout was the reported reason.
Boundary case rather than a verified extinction-risk resignation.
Every source
There are 67, with what kind of document each one is and whether it was actually read. An index lead means the record was found in a directory that was not opened past its listing.
All 67 sources
- AI Lab Watch — historical 2024 departure trackerSecondary tracker. Read.
- Anthropic founding announcementOrganization announcement. Read.
- Associated Press — Jacob Coxon departureReporting. Relevant report/search material inspected.
- Ilya Sutskever departure — Associated PressReporting. Read.
- Associated Press — Mira Murati departure announcementReporting. Search/report summary inspected.
- AVERI teamOrganization biographies. Read.
- Jacob Coxon departure report — Business InsiderReporting. Read.
- William Saunders explains resignation — Business InsiderInterview. Read.
- Turner resignation over Pentagon contract — Business InsiderReporting. Read.
- Mrinank Sharma’s exit letter — Business InsiderLetter reproduced in reporting. Read.
- Dario Amodei on leaving OpenAI — Business InsiderInterview reporting. Read.
- Business Insider — OpenAI safety-leadership departuresReporting. Relevant article text read.
- Former OpenAI employees’ amicus brief and Markov declarationPrimary legal filing. Read.
- Gretchen Krueger — existing tracker profileSecondary discovery profile. Read.
- Carroll Wainwright — existing tracker profileSecondary discovery profile. Read.
- John Schulman — existing tracker profileSecondary discovery profile. Read.
- Leopold Aschenbrenner — existing tracker profileSecondary discovery profile. Read.
- Pavel Izmailov — existing tracker profileSecondary discovery profile. Index lead only.
- Devin Kim — existing tracker profileSecondary discovery profile. Read.
- El Mahdi El Mhamdi — existing tracker profileSecondary discovery profile. Read.
- René Mayrhofer — existing tracker profileSecondary discovery profile. Read.
- Ethical AI Departures — 69-profile rosterSecondary tracker. Read.
- Harsh Mehta — tracker profile and correctionSecondary tracker. Read.
- Behnam Neyshabur — tracker profile and correctionSecondary tracker. Read.
- Hieu Pham — tracker profileSecondary tracker. Read.
- Yonadav Shavit scholarly profileSelf-maintained biography. Search result inspected.
- Jan Leike departure and safety priorities — GuardianReporting. Read.
- Steven Adler’s post-departure warning — GuardianReporting. Read.
- Guidelight — Page HedleyOrganization biography. Read.
- Guidelight — Girish SastryOrganization biography. Read.
- Jacob Hilton personal websiteFirst-person biography. Read.
- Ngo / Kulveit discussion, supplied copyUser-supplied first-person discussion. Read supplied full text.
- Kokotajlo equity resolution, checked with Daniel — Paul CrowleyWitness report. Read.
- Senior AI departures including Hitzig — MarketWatchReporting. Search summary inspected.
- Mati Roy — personal biographyFirst-person biography. Read.
- Geoffrey Irving personal websiteFirst-person biography. Read.
- Paul Christiano joins the OpenAI Foundation boardOrganization announcement. Read.
- OpenAI Foundation resilience programOrganization announcement. Read.
- Caitlin Kalinowski resigns — Reuters (corrected role)Reporting. Read.
- Reuters — former directors on self-governance, and OpenAI board responseReporting. Read.
- A Right to Warn about Advanced Artificial IntelligenceCollective statement. Read.
- Leaving OpenAI — Rosie CampbellFirst-person account. Read.
- Turner / Coxon and security consultant background — SF ChronicleReporting. Search summary inspected.
- Scott Aaronson — Sad times for AI safetyFirst-person essay. Read.
- Situational Awareness — Leopold AschenbrennerFirst-person site. Read.
- Schulman move to Anthropic — TechCrunchReporting. Read.
- Devin Kim retaliation lawsuit — TechCrunchReporting. Read.
- Beiermeister and OpenAI responses — TechCrunchReporting. Read.
- Karen Hao — reported 2023 concerns of Sutskever and MuratiOriginal reporting. Read.
- OpenAI Files — employee testimonies and former-employee directorySecondary compilation. Read.
- Miles Brundage leaves OpenAI — The VergeReporting. Read.
- Two former OpenAI employees on whistleblower protections — TIMEInterview. Read.
- Beth Barnes and external AI evaluations — TIMEInterview / reporting. Read.
- OpenAI controversy timeline — TIMEReporting. Read.
- Jade Leung — TIME100 AIInterview / reporting. Search summary inspected.
- Hitzig resignation and New York Times essay — Times of IndiaReporting. Search summary inspected.
- Transformer — OpenAI is haemorrhaging safety talentOriginal reporting. Read.
- Why I left Google DeepMind — Alex TurnerFirst-person account. Read.
- Anthropic researcher quits — Wall Street JournalReporting. Partial / paywall.
- Ryan Beiermeister termination — Wall Street JournalReporting. Partial / paywall.
- Jeffrey Ladish profile — Washington PostInterview / reporting. Read.
- Geoffrey Hinton biography and his May 1 clarificationSecondary biography. Read.
- Jared Kaplan — biography and cited career sourcesSecondary biography. Career section inspected.
- OpenAI economic research tensions — WIREDReporting. Read.
- OpenAI’s chief futurist is leaving — WIREDReporting. Read.
- Coxon resignation statementFirst-person statement. Link only; access blocked.
- Leike resignation threadFirst-person statement. Link only; statement quoted in reporting.
To propose an addition or a correction, send the person, the employer, the role, the date at the precision actually known, the original statement if there is one, and what connects it to this subject. A quotation needs its date and its context to travel with it. A concern someone raised later does not become the reason they left.