Google, Anthropic, OpenAI Face Unverified RSI Claims

RSI is a useful lens for tracking AI-assisted R&D, but the leap from faster research to autonomous successor design is not supported by public evidence. The real story is measurement and control: who chooses goals, verifies improvements, and remains able to stop the loop.

• –The cited survey reviews 1,250 papers and identifies evaluator quality as a central bottleneck, warning about self-confirming loops and model or diversity collapse. [arXiv](https://arxiv.org/abs/2607.07663)
• –Anthropic reports Claude authored more than 80% of its merged code as of May 2026, but says major gaps remain in choosing research goals—the boundary between AI assistance and full RSI. [Anthropic](https://www.anthropic.com/institute/recursive-self-improvement)
• –OpenAI explicitly says fully autonomous RSI is not happening today; its automated research intern still performs well-defined tasks under human direction. [OpenAI](https://openai.com/index/ai-policy-window/)
• –Google’s AlphaEvolve demonstrates bounded AI-assisted algorithm improvement through automated evaluators and evolutionary search, not a self-sustaining intelligence explosion. [Google DeepMind](https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/)