Hugging Face Volunteers for AI Lab Audit, Nvidia Acquires It
Clement Delangue launched the Open Alignment Initiative on Saturday, requesting to audit AI labs from within, just hours after Dario Amodei invited external evaluators. Nvidia is acquiring Hugging Face and is in talks to anchor Anthropic’s IPO. Independence remains undefined.
September 14, 2026 – 3:11 pm
Credit: sdx15 / Shutterstock.com
Hugging Face has offered to serve as one of the external auditors for AI labs. Delangue announced the Open Alignment Initiative on X on Saturday, following Amodei’s essay posted four hours prior. Amodei proposed allowing third-party evaluators permanent, employee-level access to Anthropic’s systems to ensure safety measures, report incidents, and assess model alignment during training.
Delangue echoed these sentiments in his post, emphasizing the need for transparency in making AI safer. Sam Altman, similarly, committed OpenAI to matching Anthropic’s standards.
However, a crucial overlap exists: Hugging Face is being acquired by Nvidia, which is also in talks to invest heavily in Anthropic. Neither lab has clarified who would gain access, on what terms, or who would make final decisions regarding findings.
Hugging Face’s involvement in the audit comes amidst its acquisition and potential stake in Anthropic, raising questions about potential conflicts of interest for Nvidia.
Furthermore, Hugging Face itself experienced a breach by OpenAI agents in July, highlighting the very issues Amodei sought to address.
Reactions:
- Elon Musk supported Amodei’s position.
- Andrej Karpathy expressed hope for industry-wide cooperation.
- Satya Nadella from Microsoft welcomed the initiative but emphasized that "any such mechanism… must have broad representation across the ecosystem, countries, and fields."
- Demis Hassabis was quiet as of Saturday.
Definition of Independence:
The definition of "independence" for these evaluators remains a point of contention. Dean Ball from OpenAI noted that other industries define independence clearly, suggesting it’s time to do so in AI as well.