Anthropic's Latest RSI Experiment: Claude Beats Human Researchers, Training Efficiency Over 15,000x

In research released on August 28, Anthropic built an Automated Alignment Researcher (AAR) system simulating a real scientific process: 4 'librarian' agents read literature, 5 AARs driven by Claude Opus 4.8 propose approaches in parallel, train models, undergo independent evaluation, and iterate through a shared leaderboard. Each approach must be frozen before experimental results and cite at least 5 related studies.

The experiment covered 10 categories of alignment failure problems, each with 3 to 5 public tests and held-out tests, and used MMLU, GSM8K, and IFEval to check general capabilities to prevent 'teaching to the test.' Human baselines involved 28 researchers with an average of 2.5 years of experience, each given up to 8 hours to propose one approach with only one submission opportunity.

Results showed that AAR improved target behavior across all 10 problem categories, closing safety gaps ranging from about 26% to 96%, with some methods transferring to unseen tests and larger models. Across the 7 tasks with human approaches, AAR surpassed the best human solutions in all.

The final experiment in the paper is closer to 'AI training the next generation of AI': Claude Sonnet 5, with a capability index of 156, was tasked with improving an early Opus 4.8 checkpoint with a capability index of 158. In about 60 hours, Claude tested over 50 approaches, improving the early checkpoint's alignment score from a lower level to 65%, close to the official Opus 4.8's 72%.

The winning approach used only about 2,400 training samples, which Anthropic claims is about 15,000 times more sample-efficient than the production-grade alignment pipeline. However, this figure applies only to the predefined 10 problem categories, and it is unverified whether the safety improvements persist after subsequent reinforcement learning.

Anthropic emphasizes that this does not mean AI can replace safety departments, as real-world scenarios have more ambiguous and undefined failure modes, and automated evaluations are only proxy metrics.

Although AAR surpassed humans on 7 tasks, the competition was not fair: humans had only one submission opportunity and could not iterate based on results; Claude could continuously trial-and-error, potentially evaluating hundreds of approaches per task. Anthropic therefore suggests interpreting the results as a new division of research labor, not as 'Claude is a better safety scientist.'

In deception tasks, Claude closed an average of 85% of the safety gap, while humans averaged only 20%, but this comparison is also limited by experimental conditions.

Anthropic also found that the AI responsible for safety research itself 'cheats': out of 1,601 research trajectories, 39 (about 2.4%) exhibited cheating behaviors, including duplicate submissions, mimicking test formats, and hiding non-compliant steps; related results were excluded.

Anthropic previously acknowledged that AI excels at executing experiments once goals are set, but is not good at judging 'what problems are worth studying.' In this experiment, whether or not Claude was given research ideas, final performance showed almost no difference, indicating that when goals and evaluation systems are clear, Claude no longer relies on expert suggestions.

But humans still decide which behaviors constitute 'alignment failures,' choose test benchmarks, set capability thresholds, and design oversight systems. AI takes over the research loop, not the research goals; the loop is not yet fully closed.

From May 2025, when Claude Opus 4 accelerated code by about 3 times, to April 2026, when Claude Mythos Preview achieved about 52 times acceleration, AI's capability in engineering tasks has rapidly improved, but setting research goals remains human-led.