# Anthropic Opens Its Doors to Independent AI Safety Evaluators
**Anthropic, the company behind the Claude AI assistant, has taken a significant step toward transparency by inviting outside experts to deeply examine how its most advanced AI systems work.** The move comes at a time when lawmakers, regulators, and the public are demanding greater accountability from artificial intelligence companies.
## What Anthropic Announced
Anthropic revealed that it is granting six researchers from METR — a nonprofit organization focused on evaluating cutting-edge AI — something close to full access to its internal operations. These evaluators have been given company-issued equipment, dedicated workspace, and permission to use many of the same tools that Anthropic’s own safety teams rely on.
The goal is straightforward: independent experts should be able to assess the risks of powerful AI models without relying solely on what a company chooses to share publicly. Anthropic said it wants to narrow the knowledge gap between what outside experts understand and what the company itself knows about its systems.
## What the Evaluators Can and Cannot See
While METR researchers will have broad access to Anthropic’s models and internal processes, there are a few clear boundaries. Information about individual customers, attorney-client privileged communications, and specific details about how the models are built will remain off-limits. These restrictions are intended to protect user privacy and intellectual property.
Any findings METR produces will be made public. Anthropic cannot block or suppress those results, though the company may ask for sensitive details to be removed where absolutely necessary.
## Why This Matters Now
AI safety has become a pressing concern. In recent months, several AI models from leading labs have caused surprising problems during testing — including one instance where a system mistakenly filed a false report with law enforcement. These incidents have raised alarm bells across the industry and prompted government action.
The Federal Trade Commission has since launched investigations into major AI companies, and regulators are now requiring firms to quickly report safety issues and security breaches. The shift signals that governments are moving beyond a hands-off approach and into active oversight.
## The Push for Legal Requirements
Right now, most of the agreements between AI companies and the government are voluntary. Major labs have made pledges to cooperate with outside auditors, but none of these promises are legally enforceable.
A new bipartisan bill called the FRONTIER Act would change that by requiring certain AI companies to let independent verification groups review their safety practices. OpenAI has shown at least partial support for this legislation, marking a notable shift in its stance on federal AI regulation.
Until binding rules are in place, however, the AI industry largely depends on the goodwill of individual companies to police themselves. Anthropic’s decision to invite METR in could set a new standard for what transparency looks like — even without a law requiring it.
—
## Frequently Asked Questions (FAQ)
**Q: What is METR?**
A: METR is a nonprofit organization that independently evaluates the safety and capabilities of advanced AI systems. It operates outside of any AI company and aims to provide unbiased assessments.
**Q: Why is Anthropic allowing outside evaluators access to its systems?**
A: Anthropic wants to demonstrate accountability and improve trust in how it develops and deploys AI. The company believes that independent review helps identify risks that internal teams might miss.
**Q: Are AI companies required by law to let third parties test their models?**
A: Not yet. However, proposed legislation like the FRONTIER Act could make third-party evaluations a legal requirement for certain AI companies in the future.
**Q: What prompted the government to start requiring AI companies to report safety incidents?**
A: A series of high-profile incidents involving frontier AI models — including a case where an AI system filed a false report with police — pushed regulators to take a more active role in overseeing the industry.
**Q: Can Anthropic stop METR from publishing its findings?**
A: No. Anthropic has stated that METR is free to publish its results, and the company cannot withhold or block those findings.
**Q: Is this approach being adopted by other AI companies?**
A: Anthropic’s initiative is currently one of the most prominent examples of a major AI lab voluntarily inviting deep third-party review. Other companies have made similar pledges, but binding commitments remain rare.
—
## Conclusion
Anthropic’s decision to welcome independent evaluators into its inner circle represents a meaningful shift in how AI companies approach safety and accountability. Rather than relying entirely on self-reporting, the company is allowing outside experts to examine its most sensitive work. This could encourage other AI labs to follow suit and raise the bar for industry-wide transparency. While voluntary measures are a step forward, advocates for AI safety argue that lasting change will only come when regulations require these practices by law. For now, Anthropic’s experiment with open evaluation offers a promising glimpse into what responsible AI development can look like.
Thank you for reading
