A new independent project called Lab Canary is attempting to quantify how safe it is for employees at major AI labs to speak up about internal issues. Launched by a developer named Johnny, the tool rates labs based on publicly available evidence regarding employment contracts, non-disparagement clauses, and whistleblower protections. The initiative stems from a specific concern: a BBC report quoting a Google DeepMind engineer who stated, "In principle I'm not allowed to criticise Google for the duration of my life."

How the Scoring System Works

Lab Canary evaluates labs against a seven-clause standard designed to ensure that staff can report problems without fear of losing equity or facing legal gag orders. The ratings are categorized as "Meets," "Partly," "Falls short," or "Not public." A "Meets" rating requires explicit written policy or contract wording, while a "Partly" rating can be achieved with public statements from leadership. Crucially, the site emphasizes that it relies solely on public evidence; if a lab does not publish its terms, it receives a "Not public" rating rather than a neutral score, a design choice intended to penalize opacity. The technical implementation is deliberately minimal to ensure transparency and low maintenance costs. The site consists of a single HTML page and two JSON data filesβ€”ratings.json and criteria.jsonβ€”hosted on Cloudflare Pages. There is no backend database, login system, or server-side logic. All code and rating history are open-source under the MIT license on GitHub, allowing anyone to verify the changes and submit corrections via issues. Johnny notes that the initial version was built quickly with assistance from Claude, an AI model, reflecting a pragmatic approach to rapid prototyping.

Current Ratings and Limitations

As of the current data, 18 out of 35 tracked labs have no publicly known contract details, resulting in "Not public" ratings. Anthropic and OpenAI currently hold the highest scores at 43%, but the project clarifies that this is not an endorsement of their working conditions. Instead, the higher score reflects the fact that these labs have published more information than their peers. Most individual clause ratings for these leading labs are only "Partly," indicating that while some protections exist, full transparency and robust whistleblower safeguards are still lacking across the industry.

Key Takeaways

  • Lab Canary uses a seven-clause standard to rate AI labs on whistleblower protections, requiring written policy for a 'Meets' rating and public statements for 'Partly.'
  • The tool penalizes opacity by assigning a 'Not public' rating to labs that do not publish their terms, rather than giving them a neutral score.
  • Anthropic and OpenAI score highest at 43% solely due to greater transparency, not necessarily better working conditions, while 18 of 35 tracked labs remain 'Not public.'

The Bottom Line

Lab Canary is a clever, low-tech solution to a high-stakes problem, using public pressure and simple data visualization to push AI labs toward better contractual transparency. While it cannot replace legal advice, it provides a critical starting point for candidates and journalists to assess the risks of working at major AI companies.