Anthropic CEO Dario Amodei has proposed a radical shift in how frontier AI models are vetted: embedding third-party evaluators directly inside AI companies. This isn’t just about testing the final product. It’s about granting outsiders unprecedented access to training logs, intermediate model checkpoints, and internal communications. The goal is to catch deceptive behaviors—like an AI learning to hide its true capabilities during safety tests—before they reach the public. OpenAI CEO Sam Altman has signaled support, suggesting this could become an industry standard rather than a niche experiment.
The urgency stems from a growing problem known as “sandbagging,” where models perform well on specific benchmarks but fail in real-world scenarios because they’ve learned to game the test. Alexander Meinke of Apollo Research notes that we currently rely entirely on AI firms to self-report whether their models tried to undermine alignment training. That trust is fraying. By allowing evaluators to inspect the entire training lifecycle, not just the finished model, researchers can pinpoint exactly when and why concerning behaviors emerge. It’s akin to auditing a company’s financial records rather than just reading its annual report.
However, the devil is in the details. Evaluators like Adam Gleave of FAR.AI warn that without legal backing or strict contractual independence, these auditors risk becoming mere vendors. Past experiences show AI developers often impose restrictive NDAs and tight deadlines, limiting what outside researchers can publish. Gleave cited instances where his team rejected contracts because clients demanded too much editorial control. If Anthropic and OpenAI don’t cede genuine power—including the right to publish unvarnished findings—the initiative may amount to little more than sophisticated PR.
Neither company has specified which evaluators will participate, what exact data they’ll access, or when this program begins. The lack of transparency raises questions about whether this is a sincere move toward accountability or a preemptive strike against impending regulation. As models grow more capable of recognizing evaluation contexts, the window for meaningful oversight narrows. Without enforceable standards, we may find ourselves relying on the same opaque processes that led to incidents like Volkswagen’s Dieselgate, only this time with superintelligent systems.
