Anthropic Demonstrates AI Agents That Outperform Humans in Alignment Research

Led by Anthropic fellow Chen Yueh-Han, a newly published paper details how Automated Alignment Researchers (AARs) can improve a model's performance on alignment benchmarks. [2] The automated systems search available literature, propose methods, and train models for 30 minutes per iteration. [2] Across 10 measured alignment failures, such as deception and sycophancy, the strongest AAR methods significantly reduced failures without degrading overall performance. [1][2] Furthermore, these methods successfully generalized to held-out benchmarks, multi-turn behavioral audits, and models up to 4.7 times larger than the targeted model. [1]

To establish a human baseline, 28 experienced researchers were given up to eight hours to develop methods for the same alignment benchmarks. [1] The researchers found that the best AAR methods outperformed the ideas proposed by the experienced humans. [1][2] Providing human ideas as an initial research direction for the automated systems did not improve performance. [1] In terms of expenses, an AAR costs approximately $4 per hour in API inference, compared to the $150 per hour paid to human researchers. [2]

This development represents a measurable step toward recursive self-improvement, where models enhance their own training practices. By successfully mitigating failures like jailbreaks and deception autonomously, the AAR framework suggests that AI systems could handle increasingly complex post-training tasks without human oversight. The cost disparity—$4 per hour for an AAR versus $150 for a human expert—presents a compelling economic argument for labs to scale automated alignment workflows. Furthermore, the fact that human-guided directions failed to improve AAR performance indicates that AI research methodologies may already be diverging from human intuitions. What the sources don't address: How these automated researchers perform when tasked with identifying entirely novel alignment failures that are not already measurable by public benchmarks.