Sources: Anthropic and OpenAI have pushed for substantial control over what external AI evaluators can publish, out of concern that trade secrets could leak

If you are developing or evaluating advanced AI, you must now expect tighter controls over what can be published from third-party assessments.

Anthropic and OpenAI are pushing for substantial control over the publication rights of external AI evaluators. The generative AI leaders are concerned that the findings from these independent assessments could inadvertently reveal sensitive trade secrets. This move comes as the companies engage third parties to evaluate their advanced AI models, a common practice to ensure safety and identify potential risks before public release. The exact nature of the controls sought by Anthropic and OpenAI is not fully detailed, but it suggests a desire to review or restrict information that evaluators might share publicly about their models' capabilities, limitations, and underlying technologies. The implications for transparency in AI safety research are significant, as it may limit the scope of public understanding regarding these critical technologies.

The push by Anthropic and OpenAI for greater control over external evaluator publications signals a growing tension between the need for AI safety transparency and the protection of proprietary technology. As these models become increasingly complex and costly to develop, companies are becoming more guarded about revealing details that could be exploited by competitors or bad actors. This trend may lead to a more opaque AI development ecosystem, where independent researchers have limited access to the full picture of a model's workings and potential risks.

This situation could create a challenging environment for the AI safety community, potentially hindering collaborative efforts and slowing down the collective understanding of AI risks. It raises questions about who ultimately benefits from such controls: the companies seeking to safeguard their intellectual property, or a public that may be less informed about the technologies shaping its future. Watch for the development of new frameworks for external AI evaluation that balance proprietary concerns with the public interest in safety and transparency.