Anthropic’s automated researcher improves AI alignment without human help

Anthropic has published a paper that offers a concrete glimpse of self-improving AI. The paper, “Automated Researchers Can Reliably Mitigate Alignment Failures,” describes a system that autonomously improves a model’s performance on alignment benchmarks—those tests designed to catch misbehavior like bias or unsafe outputs. When given ten such benchmarks, the automated system improved performance on every single one, without degrading the model’s overall capabilities.

The approach mimics the workflow of a human researcher. Each automated system scans the relevant literature, proposes a method, and trains the model with it for 30 minutes. Over several iterations, it gradually pushes the benchmark scores higher, keeping what works and discarding what doesn’t. The result, the paper says, is evidence that “automated alignment post-training could become practical in the near term.”

A step toward recursive self-improvement

This is an early step toward recursive self-improvement—the idea that AI models can improve their own training, including the very alignment techniques that keep them safe. If that becomes reality, human AI researchers might eventually be out of a job. The paper doesn’t shy away from that comparison. It claims the best automated method “beats what experienced humans propose, on average within six hours,” and that human-guided research directions don’t lead to stronger performance.

There’s also a stark cost difference: the automated researcher costs roughly $4 per hour in API inference, versus the $150 per hour paid to human researchers.

Not a silver bullet

Still, the approach has clear limits. The automated system is only as good as the benchmarks it’s optimizing against. If those benchmarks don’t capture the real alignment goals, the fixes could be superficial. And maintaining and expanding the literature the system draws from still requires human oversight.

For developers, the takeaway is more subtle. This isn’t just about automating a research process—it’s about changing the economics of AI safety. If alignment research can be conducted at machine speed and scale, the bottleneck shifts from compute to the quality of the goals we set. The paper is a reminder that the next frontier isn’t just smarter models, but models that can make themselves safer without waiting for humans to catch up.