Anthropic Says AI-Assisted Research May Near Its RSI Threshold

Anthropic has said in a revised S-1/A filing that it may imminently meet its internal Responsible Scaling Policy threshold for “dramatically accelerated progress attributable to AI.” The company has not designated any current model, including Claude, as exceeding that threshold, and says it continues to assess the underlying data.

The disclosure concerns the use of AI systems in Anthropic’s own research and development. The company warns that AI-assisted R&D may advance model capabilities faster than it can develop, validate and implement corresponding safety, alignment and security measures.

Anthropic frames accelerated progress as a safety issue

Recursive self-improvement, or RSI, describes a situation in which AI materially improves the process of building more capable AI. Anthropic says its Responsible Scaling Policy addresses the risks associated with approaching this possibility, particularly if the pace of progress sharpens suddenly.

Its filing adds that the company has previously failed to identify a significant acceleration in its development rate until after it occurred. Internal estimates and expectations for future models, rather than the capabilities of current models, are the basis for its view that the policy threshold may be near.

Anthropic’s financial and IPO context has also drawn close attention, with Anthropic’s near-trillion-dollar pre-IPO valuation placing the company near a $1 trillion valuation ahead of an IPO, while the new filing sets out operational risks alongside growth expectations.

Human supervision remains Anthropic’s stated position

Anthropic says it does not anticipate safely using Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if doing so becomes technically possible. It is considering updates to its Responsible Scaling Policy intended to keep humans in control of frontier-model development.

The company’s language is notable because it pairs a warning about faster technical progress with an acknowledgement that safety controls may lag. It does not state that a current model has reached RSI, nor does it claim that autonomous frontier-model development is safe today.

What the filing means for organizations

For businesses using or evaluating frontier AI, the filing reinforces the value of retaining accountable human oversight as models and their development cycles evolve. Governance processes should be able to review changes in capability, security and risk controls rather than assuming that current safeguards will remain sufficient.