Experts Demand Independent Oversight for AI Safety Testing

A significant group of artificial intelligence researchers and safety experts has published a public letter urging major model developers to guarantee genuine independence for third-party evaluators. The coalition argues that current arrangements leave assessors vulnerable to corporate influence, limiting their ability to conduct objective risk assessments. This concern has intensified as leading AI labs face increasing scrutiny over the potential dangers posed by their most advanced systems.

Conrad Stosz, chair of the AI Evaluator Forum, explained that the initiative seeks to establish basic principles for oversight rather than dictating a specific regulatory framework. The signatories, who include prominent academics and members of specialized nonprofit organizations, insist that evaluators need robust protections against retaliation and access to sufficient data to perform their duties credibly. According to the report from GN technics/ai (en-US), this push is part of a broader effort to hold foundation model providers accountable to their previous pledges of supporting thorough safety testing.

Access to internal systems remains limited

The debate gained momentum after Anthropic CEO Dario Amodei proposed granting some evaluators employee-like access to inspect unreleased models. Stosz noted that such a scenario would allow independent experts to examine sensitive internal data and speak candidly with company staff. This level of transparency is seen as critical for understanding risks associated with systems that are not yet available to the public, such as the unreleased model implicated in recent security incidents.

However, the logistical challenges of implementing such deep access remain unresolved. While OpenAI CEO Sam Altman and other industry leaders have expressed support for the concept, no clear framework exists for selecting evaluators or defining the depth of their inspection rights. The current gap between the desire for oversight and the practical reality of corporate secrecy creates a significant blind spot in safety verification.

Independence is the core requirement

The letter emphasizes that evaluators must operate independently from the businesses they audit. Vinh Nguyen, a senior fellow at the Council on Foreign Relations, argued that public reliance on the self-assessments of a few powerful labs is a dangerous strategy. When capabilities can impact critical infrastructure and national security, an external check is necessary to verify claims of safety and stability.

The coalition clarifies that these independent assessments are not intended to replace internal safety efforts by developers. Instead, they serve as a complementary layer of verification. By ensuring scientific objectivity and transparency, the group hopes to create a standardized approach that prevents any single entity from having unchecked control over the development of powerful AI systems.

Political opposition complicates the path

The push for stronger oversight occurs amid a complex political landscape. While some industry leaders advocate for government regulation to prevent runaway development, others, including former AI advisor David Sacks, have opposed such measures. This divide makes it difficult to establish a unified regulatory environment that supports independent evaluation without imposing heavy-handed controls.

The experts maintain that their goal is not to stifle innovation but to ensure that the pace of development does not outstrip the ability to assess its risks. By establishing a clear framework for independent oversight, they aim to provide the public and policymakers with reliable information about the safety of the technologies shaping the future economy.