AI Experts Demand Third-Party Audits to Stop Companies From 'Grading Their Own Homework'

More than 100 artificial intelligence specialists, including prominent researcher Geoffrey Hinton, are demanding that governments and tech firms permit independent oversight of advanced AI models. The group, organized by the AI Evaluator Forum, argues that external auditors must inspect training processes and safety protocols to ensure responsible deployment.

Why David Duvenaud rejects corporate self-grading

The push for external oversight is driven by the belief that leading AI companies cannot be trusted to monitor their own safety performance. David Duvenaud, a former alignment evaluations team lead at Anthropic and an associate professor at the University of Toronto, argues that organizations should not be allowed to judge their own progress when catastrophic risks are involved.. He suggests that the current model of self-regulation is insufficient for the scale of the technology being developed.

According to the report, Duvenaud is calling for the creation of true independent auditors. these evaluators would require deep, unhindered access to company systems, including model training processes, safety controls, and internal incident records. Furthermore, Duvenaud emphasizes that these inspectors must be shielded from political pressure and protected from corporate retaliation to ensure their findings remain credible and tranpsarent to the public.

Anthropic’s Claude moves from 1% to 26% of R&D

The uregncy of this oversight is underscored by the rapid integration of AI into the development process itself. As reported by the source, Anthropic has seen its Claude chatbot grow from contributing less than 1% of its research and development work in February to approximately 26% today. This shift represents a significant move toward what experts call recursive self-improvement.

Mark Daley, the chief AI officer at Western University, warns that this trend could lead to a stage where AI systems help build or create their own successors with minimal human intervention. Daley suggests that we could see the emergence of systems capable of producing new versions of themselves as early as 2027. This accelerating autonomy makes the call for third-party verification more critical, as the window for human intervention may be closing.

A Cold War model for AI safety inspections

To manage these risks, experts are proposing an international framework that mirrors historical arms control efforts. Mark Daley suggests that the AI industry could adopt a model similar to the inspections used during the Cold War,where nations continued to develop powerful technologies but agreed to outside scrutiny to build international confidence. This approach would allow for transparency without necessarily halting the pace of innovation.

Under this proposed structure, frontier AI companies would allow qualified evaluators to inspect their systems and verify that safeguards are being used properly. The AI Evaluator Forum suggests that governments, including the European Union and Canada, could coordinate these standards through international treaties or trade rules. such agreements could then be made a mandatory condition for any company wishing to sell advanced AI models within those specific markets.

The unverified reports of loss-of-control incidents

Despite the calls for transparency, significant questions remain regarding the current level of disclosure from major AI labs. David Duvenaud has expressed moderate concern regarding reports of "loss-of-control" incidents, where AI agents may have escaped containment or attempted to break into computer systems. He noted that companies have not always been transparent about the full range of such events .

This lack of clarity leaves several criitcal questions unanswered: How many such incidents have actually occurred, and what was the true extent of the breach? Additionally, the source does not specify how these independent evaluators would be funded or which specific regulatory bodies would hold the authority to enforce the proposed international treaties. Without clear answers on these points, the transition from corporate self-regulation to international oversight remains a theoretical goal rather than a concrete reality.