Anthropic seeks greater transparency in AI model development
Anthropic has called for public reporting on AI development, flagging the risk of models accelerating their own improvement.
AI frontier lab Anthropic has made a public push for greater transparency in how advanced AI models are developed, warning that systems are growing more powerful and increasingly automating parts of their own construction.
In a blog post, the company said that as the world weighs slowing the pace of frontier AI development, the public needs more information about what is happening inside labs. It set out a set of measurement tools covering critical aspects of development, including the degree to which AI is building its next version rather than being built by humans, the ability to oversee and intervene in actions taken by AI agents, and the resources that power the creation of more capable models.
Anthropic flagged a specific concern around what it calls recursive self-improvement — a model fully autonomously building its successor. While using AI to build future AI models lets labs in democratic countries develop capabilities faster and run more safety testing before release, the company said models accelerating their own development could make it harder for humans to understand or control such systems. It argued that sharing these metrics is important to gauge how close the world is to that point.
The company also pointed to its policy proposal, the Advanced AI Framework, which lays out rules for how any lab could release safe models. These include transparency obligations that governments could require, such as risk reports. Anthropic described the proposed measurements and policies together as a starting point for monitoring the pace of AI development from outside the labs.
The call follows remarks by CEO Dario Amodei urging caution over the speed of AI progress. Speaking to CBS News Sunday Morning, Amodei said he had not fully appreciated what the pace of change would feel like, while adding that there was no need to panic or shut everything down. He said the technology needs safeguards that can keep up with the rate at which it is advancing.
Anthropic said it hoped to model transparency by releasing such measurements and would continue to do so, adding that the gap between what frontier labs know and what the public knows should be minimised as societies decide how to use the information.