Anthropic and OpenAI push to restrict external AI evaluators' publications over trade‑secret fears
ai-safety anthropic openai
| Source: Techmeme | Original article
Anthropic and OpenAI are seeking extensive control over external AI evaluators' publications, citing concerns that trade secrets could be exposed.
Anthropic and OpenAI are seeking to tighten control over what independent safety evaluators can publish, according to sources cited by Reuters. The two labs have asked that external reviewers agree to pre‑approval of any public disclosures, arguing that unrestricted reporting could expose proprietary techniques and trade secrets. The move would embed evaluators within the companies’ own research environments, giving the labs a say over the scope and timing of any findings.
The proposal has sparked a debate within the AI safety community. Proponents say close access to the labs’ inner workings could yield more thorough risk assessments and help prevent catastrophic outcomes. Critics warn that such gate‑keeping threatens the independence essential for credible oversight, potentially limiting transparency and public scrutiny. The tension reflects a broader industry struggle to balance commercial confidentiality with the need for open, verifiable safety testing.
What follows will hinge on how the negotiations unfold. Regulators in Europe and North America are watching closely, as any formal arrangement could set a precedent for future oversight frameworks. Stakeholders will be looking for concrete terms on publication rights, mechanisms for dispute resolution, and whether third‑party watchdogs can retain the ability to report concerns without prior approval. The next weeks may also see reactions from academic researchers and civil‑society groups, who could push for legislative safeguards to preserve evaluator independence. How the two companies reconcile their security concerns with broader demands for openness will shape the evolving governance model for advanced AI systems.
Sources
Back to AIPULSEN