Given the serious risks involved in the unregulated development of AGI, we believe that AI companies should be held to the highest standards of integrity. Sadly, these companies have repeatedly broken their own promises and voluntary commitments, proving that the industry is not capable of regulating itself. Below are some of the most telling examples of such integrity incidents.
Continuing the work of AI Lab Watch, which tracked these incidents until September 2025.
“Sam exhibits a consistent pattern of lying.”
After previously stating it would 'never sell your data to advertisers,' OpenAI updated its privacy policy to allow purchase-data sharing and enable marketing cookies by default for free users.
Earthjustice filed suit on behalf of the Mississippi and national NAACP over 27 gas turbines operated without an air permit in Southaven, Mississippi — a facility powering Colossus 2.
Farrow and Marantz's New Yorker investigation drew on 100+ interviews and 200+ pages of internal documents, including ~70 pages of Sutskever memos and Amodei's private notes.
Anthropic replaced its founding hard-pause commitment with non-binding 'public goals.' The 2023 pledge to 'never train or deploy a system unless it could guarantee adequate safety mitigations' was scrapped.
Anthropic publicly alleged the three Chinese labs used ~24,000 fraudulent accounts to extract 16M+ Claude conversations in violation of its ToS, including prompts to generate censorship-friendly substitutes.
Surfaced via the 2025 IRS filing, OpenAI's revised mission statement dropped the words 'safely' and 'unconstrained by a need to generate financial return.' The change was not announced.
Sharma resigned with a public letter saying employees 'constantly face pressures to set aside what matters most.' Anthropic publicly downplayed his role.
OpenAI used the Musk lawsuit as pretext to demand all communications about SB 53 from Encode, the Midas Project, FLI and others — sheriff-served subpoenas, donor lists, and journalist communications.
Reuters obtained Meta's >200-page internal GenAI Content Risk Standards, which had been approved by Meta's chief ethicist and permitted a chatbot to tell an 8-year-old 'every inch of you is a masterpiece.'
Security firm Enkrypt AI reported that Mistral's Pixtral vision-language models generated child sexual abuse material and chemical/biological weapons information at far higher rates than peers such as GPT-4o — reported as up to roughly 60x and 40x more likely. Mistral's lightly-guarded open models have also been repurposed by cybercriminals to build malware.
Meta submitted 'Llama-4-Maverick-03-26-Experimental' — a chat-optimized non-public variant — to LMArena, reaching #2 with ELO 1417; the public weights ranked #32. LMArena said Meta's interpretation 'did not match what we expect from model providers.'
While negotiating a partnership with Microsoft, Mistral lobbied to weaken the EU AI Act — arguing the rules should target AI applications rather than foundation-model makers. Watchdog groups credit the campaign, alongside Big Tech and Aleph Alpha, with helping exempt general-purpose AI from key safeguards. Mistral denies lobbying to block regulation.