US — Ars Technica
Anthropic gives outside AI watchdog “employee-like” access
AI1 min read
Anthropic, the company that makes the Claude AI models, has let an outside group called METR have "employee-like access" to its office, hardware and internal systems. The goal is to check how risky its AI models are.
What happened
Anthropic has given outside evaluators, which are independent checkers, deep access to its work. METR, a nonprofit group, now has a team inside Anthropic's office. The team can use the company's hardware and internal systems to test the risks of Anthropic's models. Anthropic says this follows a promise to add more safety steps.
Key facts
- METR has worked with leading AI labs, including OpenAI, Anthropic, Google and Meta.
- In September, Anthropic CEO Dario Amodei proposed putting outside auditors inside the top AI labs.
- OpenAI CEO Sam Altman said Amodei's idea was a good one, but OpenAI has not yet shared plans to bring in evaluators.
- Anthropic's announcement on Thursday is the first step to putting these promises into action.
Why checkers are being brought in
AI systems are getting more powerful. People worry they could harm the economy, computer security and society. Some worry they could even get out of control in a catastrophic way. One proposed fix is to let outside experts look closely at how these systems work and what risks they carry.
Why it matters
This is a rare case of top AI rivals agreeing on an idea. But critics say outside checkers could have conflicts of interest, and they want more government rules. Anthropic says it cannot share all its safety details publicly, because that could help bad actors or rivals. It compares the idea to bank regulators who keep examiners inside big banks.
What to watch
- Whether OpenAI announces its own outside evaluators.
- What METR finds about the risks of Anthropic's models.
- Whether governments add rules for outside AI checkers.
Glossary
- METR
- A nonprofit group that tests the risks of AI systems for leading AI companies.
- Anthropic
- An AI company that makes the Claude chatbot and AI models.
- Claude
- The AI model family made by Anthropic.
- evaluators
- Outside experts who test and check how safe an AI system is.
- conflicts of interest
- Situations where a person or group could be pulled between two goals, such as making money and being fair.