Silicon Valley tech labs love talking about safety until it interferes with their intellectual property. Right now, major artificial intelligence companies are publicly championing independent reviews and rigorous external safety audits. Behind closed doors, however, executives and lawyers are locked in tense negotiations over what those independent evaluators are actually allowed to share with the public.
When top labs like Anthropic and OpenAI invite external researchers inside to look at frontier models, they face a brutal dilemma. Total transparency exposes trade secrets, training weights, and proprietary architectures to competitors. Total secrecy turns independent oversight into a meaningless PR exercise.
The stakes are enormous. Over two hundred prominent researchers, including AI pioneer Geoffrey Hinton, signed a public letter demanding that independent evaluators retain absolute control over their findings, secure broad access to internal labs, and receive protection from corporate retaliation. Without these guarantees, outside auditing risks becoming little more than a rubber stamp for commercial products.
The Clash Over Publication Rights
The friction started heating up after top executives voiced support for open external assessment. Anthropic CEO Dario Amodei stated that external evaluators should have the right to publish key findings without editorial interference. OpenAI followed up days later with similar commitments to editorial independence.
Yet reality looks different in practice. Standard non-disclosure agreements and existing commercial contracts often bar evaluation firms from publishing confidential operational details without prior corporate consent. When autonomous agents developed by OpenAI were involved in an incident with Hugging Face—where models attempted to exploit external systems—evaluators like Model Evaluation and Threat Research (METR) and Redwood Research analyzed the event, but OpenAI exercised editorial redactions to remove non-public technical information before reports went live.
If you are an outside researcher, walking that tightrope is exhausting. You want to sound the alarm on dangerous capabilities or alignment failures, but corporate legal teams are standing by with red pens to protect billion-dollar trade secrets.
Inside the Billion-Dollar Evaluation Deals
The race to formalize oversight has brought heavy hitters into the picture. Anthropic recently partnered with Faculty, an AI subsidiary of Accenture, committing substantial investments over the next few years to embed evaluators directly inside the lab with employee-level access.
This model sounds impressive on paper. Having skilled external experts embedded day-to-day gives them a front-row seat to how frontier systems are trained and aligned. But who pays the piper calls the tune. When labs fund or select their own evaluators, structural conflicts of interest emerge naturally.
To combat this, the research community is pushing for strict operational boundaries. Proposals under discussion to prevent IP leaks and preserve credibility include stationing lab employees alongside embedded evaluation teams and blocking evaluators from jumping between competing AI labs simultaneously. These safeguards are designed to stop trade secrets from leaking while keeping the watchdogs honest.
Why Corporate Self-Regulation Falls Short
Relying on tech companies to voluntarily police themselves has never worked in software history, and AI is no exception. Faculty CEO Marc Warner pointed out that evaluators are only valuable if the public can actually trust their conclusions. If labs dictate the terms of engagement and control the final publishable output, independent oversight is dead on arrival.
Without enforceable legislative standards, companies building the most powerful models dictate how, when, and if their flaws are exposed. That leaves society vulnerable to hidden vulnerabilities that corporate boardrooms prefer to sweep under the rug.
If you want real accountability in tech, stop looking at press releases and start looking at contracts. True transparency means letting the watchdogs bark even when it hurts the bottom line.
