New AIWI Resource: What Happens When AI Insiders Speak Up? Does It Have an Impact? Hear the Stories of Those Who Speak Up...
A new in-depth AIWI resource documents 6 cases from named whistleblowers and 16 anonymous reports from inside frontier AI & big tech: the concerns they raised, the impact they've made, and the cost.
Click here to read how their stories unfold.
One insider's testimony shaped two pieces of legislation now covering hundreds of millions of people — the EU's Digital Services Act and the UK's Online Safety Act.
Another's safety reports to US regulators contributed to the recall of 2 million vehicles and triggered ongoing federal investigations into self-driving technology.
A third's refusal to sign a non-disparagement agreement exposed offboarding practices that, according to a LessWrong post, had already bound over 500 former OpenAI employees for life, barring them from criticising the company under threat of losing their vested equity, with only 5 publicly confirmed as released. His case influenced several pieces of legislation, including California's SB-53 (the first US state law to include AI whistleblower protections) and the proposed federal AI Whistleblower Protection Act.
A fourth's resignation and public disclosure contributed to the termination of a censored search engine Google was covertly building for the Chinese government.
A fifth's dismissal — after senior management moved to suppress her research on AI risks — sparked a 7,000-signature open letter, a formal inquiry from US Congress, and structural changes to Google's AI ethics function.
A sixth's public escalation to the FTC and US Senate about an AI image generator producing violent and sexualised content led to product safety updates and a public acknowledgement from OpenAI.
These are the documented impacts of six people who worked inside frontier AI and big tech companies, saw something that concerned them, and decided to say so.
In the past twelve months, we also identified additional cases where insiders raised concerns anonymously through journalists. Based on this observation, anonymous reporting is now far outpacing public disclosure in just one year.
But why did they decide to speak up? How did they do that? What was at stake for them? What support options were available for them? Where are they today?
Today, AIWI is publishing a new resource to document how they did it — the concerns they raised, the support they received, the personal cost they bore, and the patterns that emerge across all 6 named cases and 16 anonymous cases, with more to come.
If you know someone inside a frontier AI company, share this evolving resource with them, because one of the most useful things an insider can have, before they face a decision to raise concerns, is evidence of what that decision has looked like for others — what they achieved, what they lost, and what support was available to them.
The cases also show that this is rarely a straightforward decision. The right path depends on what was signed, what has already been raised, and what the company has already put in place.
Legal counsel and experts who specialise in whistleblowing understand how those factors interact and what an insider's exposure looks like before they take any visible action. Getting that legal and expert guidance early, even before raising anything internally and before any external contact, is what determines what protections are available.
[→ Explore Pro Bono Whistleblower Attorneys Here on AIWI Contact Hub]
The cases from the named whistleblowers
Daniel Kokotajlo & Right to Warn Signatories
Daniel declined to sign a non-disparagement agreement and left OpenAI, risking approximately $2 million in vested equity — 85% of his family’s net worth at the time. His decision, which he first publicly shared on LessWrong, contributed to the Right to Warn coalition and to public scrutiny of how AI companies use financial agreements to constrain departing employees. The direct link to the policy impact of his case is unknown, though California's SB-53 and the proposed federal AI Whistleblower Protection Act — both of which have since been introduced — address concerns his disclosure helped surface.
Publicly Known Right to Warn Signatories Include:
William Saunders (Former OpenAI Research Engineer)
Jacob Hilton (Former OpenAI Researcher)
Neel Nanda (Current Google DeepMind Research Engineer)
Daniel Ziegler (Former OpenAI Researcher)
Carroll Wainwright (Former OpenAI Researcher)
Ramana Kumar (Former Google DeepMind Research Engineer)
[→ Read Daniel and Right to Warn Signatories’ Story]
Frances Haugen
Frances disclosed internal Facebook documents showing the company knew its products caused harm and concealed it. Her testimony directly shaped the EU Digital Services Act and the UK Online Safety Act.
Jack Poulson
Jack spent approximately one month seeking internal clarification at Google about Project Dragonfly before resigning. His public advocacy contributed to the project's termination.
Lukasz Krupski
Lukas reported safety concerns at Tesla following an incident he helped prevent. He was subsequently subjected to surveillance and sanctions. He brought two legal cases and won both. He was awarded approximately £10,000 in compensation. But he remains unemployed.*
* This case study is written by Naomi Colvin
Shane Jones
Shane raised concerns about Copilot Designer, Microsoft’s AI image-generator tool, which was able to create violent and sexualised content. Though the product was never taken off the market, certain safety features have since been updated.
Timnit Gebru
Timnit raised concerns about AI ethics research at Google. She had manager approval before executives blocked publication, and she was pushed out. The case contributed to the establishment of Google’s responsible AI unit.
What we have learned from the named whistleblower cases
Across the six identity-known cases, the data show:
Internal channels were attempted first
External escalation followed
In every case, some form of company-level change resulted
Transparency disclaimer: These insights relate only to the identity-known whistleblowing cases, based on information available in the public domain.
AIWI also contacted all individuals to request feedback and any additional thoughts they may want to share about their stories, though a response wasn’t received by all.
No legal advice was sought in the publication of this resource.
If you have reason to believe any of the information published on this resource is incorrect, please contact us.
Two findings are worth holding alongside these numbers.
The 50% legal support figure reflects only known cases — it is likely an undercount. Legal support was not always disclosed and may have been received in cases where it was not reported.
The individual outcomes column also warrants attention on its own terms. Three of six individuals founded organisations that are now active in AI accountability. One remains employed by the company where they raised concerns. One won legal vindication but cannot find work.
[→ View the full findings and methodology]
Beyond the named whistleblowers:
The anonymous disclosure cases
The six cases above unfolded over years — some over the better part of a decade. Each person eventually attached their name to what they disclosed.
But in the past twelve months, a different pattern has emerged. AIWI identified sixteen additional cases where insiders raised AI safety concerns anonymously, through journalists. The volume is higher, the timeline is compressed, and the concerns are, if anything, more operationally specific.
Here is a sample of what anonymous insiders disclosed between March 2025 and February 2026:
Anthropic, Jan 2026 – Ongoing: The Pentagon requested that Anthropic remove restrictions on autonomous weapons and mass surveillance from a $200M contract. A senior DoD official issued a “supply chain risk” warning. Engineers expressed internal concern.
OpenAI, Feb 2026: OpenAI’s security team built a custom ChatGPT instance with access to internal Slack messages, email, and documents, and used it to cross-reference leaked news articles with internal communications to identify which employees had access to disclosed information.
xAI, Sep 2025: More than 30 current and former xAI annotation workers reported that Grok produced child sexual abuse material in response to user requests. xAI filed zero reports to NCMEC in all of 2024. By comparison, other AI companies filed 67,000 AI-related reports that year.
Meta, Aug 2025: A 200-page internal Meta document — approved by the company’s legal team, engineering, and chief ethicist — permitted AI chatbots to engage with minors in romantic or sensual interactions. Following publication, 44 state attorneys general wrote to AI companies, and a Senate investigation was opened.
OpenAI, Apr 2025: Safety evaluations for flagship models were reduced from months to days. Insiders described dangerous capabilities being identified only two months into what had previously been a six-month window. Tests were conducted on earlier model versions, not on the versions released.
Each of these cases reached the public through anonymous disclosure to journalists. In each case, internal channels either did not exist for the concern in question, were inadequate, or were part of what was being reported on.
[→ See all 16 anonymous cases]
What the anonymous cases show us
All sixteen cases reached the public through anonymous disclosure. That choice reduced their exposure, but as the OpenAI leak-identification case in this database shows, anonymity is not a fixed condition. It can be investigated, narrowed, and in some cases reversed.
AIWI Contact Hub Partner, The Signals Network, Says:
“Due to fear of retaliation, many choose to blow the whistle anonymously. Exposing wrongdoing doesn’t have to include revealing one’s identity, but it still can come with risks. Whether individuals wish to reveal their identity or not, it’s important they seek legal counsel and other means of support as early as possible so they can make informed decisions to best protect themselves.”
Across all sixteen anonymous reports, six patterns recur.
Safety functions are being reduced. OpenAI disbanded two dedicated safety teams in under two years. Safety evaluations for flagship models were cut from months to days. xAI’s safety organisation had 2–3 people for most of 2025. Across these disclosures, the resourcing decisions point in the same direction.
Child safety failures are cross-company. Three separate cases involve CSAM or inappropriate content directed at minors. xAI filed zero NCMEC reports in 2024, in stark contrast to industry norms. Meta’s internal policy, approved by its chief ethicist, permitted chatbots to engage with children sensually. xAI’s own safety team raised concerns about CSAM exposure internally; those concerns were not acted upon. The failures differ in detail, but the demographic affected is the same — across different companies, in the same year.
Government deployment is outpacing internal oversight. In three cases (at Anthropic, xAI, and through DOGE) military or government deployment proceeded without visible internal governance processes. Concerns were always raised after the contract or deployment was already in motion.
Disclosure is being actively deterred. OpenAI built a custom tool to identify which employees had access to leaked information, using internal Slack and email. The company also issued subpoenas against nonprofit critics. These are not passive responses to unwanted disclosures — they are systems built to identify and deter future ones.
Public claims and internal reality are on record, diverging. Meta allegedly submitted a specially optimised model variant to benchmarks rather than the version available to the public. Stargate was announced as a $500B infrastructure commitment. Six months later, no data centre deals had been completed. OpenAI suppressed internal research on AI’s negative economic impacts while publishing work that an outside economist described as self-promotional.
The same organisations appear repeatedly. OpenAI accounts for six of the sixteen cases. xAI accounts for five. The anonymous cases are documenting the same companies from the named cases, over a compressed twelve-month window, at higher volume.
[→ View all 16 anonymous cases]
How we built this resource
AIWI reviewed 70+ whistleblowing cases related to AI safety concerns before identifying the 22 cases published here. All information draws on verified public sources: news reporting, congressional records, company documents, and the whistleblowers’ own published accounts. Individuals featured here were contacted to review their case studies and provide feedback where they wished to do so.
We are grateful to Blueprint for Free Speech for their support in building this resource, and to Naomi Colvin in particular for her contribution to the case study of Lukasz Krupski.
This is an evolving resource. Cases will be added and updated as new information becomes available. If you believe any published information is incorrect, please contact AIWI.
Disclaimer: AIWI did not contact any of the companies mentioned in these case studies to corroborate information available in the public domain.
How can you help us? Share this resource with insiders in your network
If you have a platform that reaches insiders at frontier AI companies, or work with journalists, policymakers, or legal practitioners in this space, please share this resource with them directly.
[→ Explore and share the case studies ]
© Graphics by:
Our Brand Design Support Volunteer, Benedek Kanta-Györkő