Anthropic gives third-party auditors ’employee-like access’ in AI industry first
Original reporting by ZDNet
Anthropic's new safety initiative involves granting "employee-like access" to third-party evaluators from the nonprofit METR. On Thursday, the AI lab announced it would embed six METR evaluators directly within its operations, providing them with hardware, office space, and access to internal tools on par with its own risk assessment teams. This unprecedented move aims to foster greater transparency and accountability for frontier AI models, allowing independent experts to scrutinize systems and processes that are typically guarded closely. Anthropic cites a desire to bridge the understanding gap between companies and external evaluators, allowing METR to publish its findings without company censorship, while still protecting commercially sensitive or potentially malicious information.
Regulatory pressure This comes amidst growing calls for mandatory third-party review of AI models, a significant shift from the current landscape of voluntary agreements between labs and governments. Recent incidents, such as Anthropic's Claude generating a false homicide tip, have intensified scrutiny, prompting regulators like the FTC to subpoena leading AI companies. The government has also begun requiring AI companies to disclose safety incidents and security breaches, moving away from previous self-regulation. While legislation like the FRONTIER Act proposes independent verification, Anthropic’s move could be seen as either a proactive step towards greater safety or a strategic maneuver in anticipation of stricter mandates. The initiative highlights an evolving industry grappling with how best to ensure the safe development of increasingly powerful AI.
Anthropic's pioneering decision to grant METR evaluators unprecedented "employee-like access" marks a pivotal moment in the ongoing debate surrounding AI safety and accountability. This strategic move, aligning with CEO Dario Amodei’s earlier proposals, seeks to bridge the transparency gap between AI developers and external auditors, addressing concerns that commercially sensitive or security-critical information cannot always be publicly disclosed. It represents a significant step beyond the voluntary pledges that have largely characterized the industry's self-regulatory efforts thus far, responding to escalating calls for more robust oversight.
A new precedent
The implications of Anthropic's initiative extend far beyond a single company. In a climate of intensified regulatory scrutiny — evidenced by the bipartisan FRONTIER Act, recent FTC subpoenas, and the government's new mandate for incident disclosure — this embedding strategy could well set a powerful precedent. If successful, it may become a de facto standard, influencing not only future legislation but also the operational practices of other frontier AI labs facing similar pressures. The industry is navigating a delicate balance between rapid innovation and responsible development, and the independent, deep-dive evaluation offered by METR could be crucial in fostering public trust and ensuring model safety. This approach signals a potential shift towards more integrated and transparent third-party oversight, a critical evolution as powerful AI models become more pervasive and their societal impact, both positive and negative, grows increasingly profound. How other leading AI developers respond to this new model of accountability will shape the regulatory landscape for years to come.
Frequently asked questions
- What is Anthropic doing to ensure the safety of its AI models through external evaluation?
- Anthropic is collaborating with METR, a third-party nonprofit, to evaluate its frontier AI models. METR evaluators receive "employee-like access" to internal tools, systems, and office space, enabling deep scrutiny of model risks. METR will publish its findings independently, aiming to provide unbiased assessments of model safety and alignment, closing the information gap between the company and external experts. This initiative promotes accountability and transparency in AI development.
- Why is third-party evaluation of advanced AI models becoming increasingly important for safety?
- Third-party evaluation provides independent oversight of advanced AI models, addressing concerns about safety and potential misuse. External evaluators, often from non-profits, gain deep access to internal systems to identify risks, biases, and vulnerabilities that companies might miss or be reluctant to disclose. This process aims to increase accountability, build public trust, and proactively mitigate harm from powerful AI systems before widespread deployment, ensuring public welfare.
- What new government regulations or requirements are being considered for AI companies?
- Governments are increasing pressure on AI companies regarding safety and transparency. The FRONTIER Act, a proposed bipartisan bill, suggests mandatory third-party reviews of AI development processes. Additionally, recent US government mandates require AI companies to immediately disclose safety incidents and security breaches involving their models. These measures aim to shift the industry from voluntary agreements towards binding regulations, ensuring greater accountability and protection against potential harms from advanced AI.