An AI system built inside Anthropic beat experienced human researchers at fixing an AI model’s alignment problems, and it did the job for a fraction of the cost. That’s the headline from a new paper Anthropic published Friday, surfaced by TechCrunch AI, titled “Automated Researchers Can Reliably Mitigate Alignment Failures.” This is significant because it’s one of the first concrete looks at self-improving AI moving from theory into something that actually runs.
The work came out of Anthropic’s fellows program, led by fellow Chen Yueh-Han. According to TechCrunch AI, the team handed an automated system 10 benchmarks, each targeting a specific misaligned behavior, and asked it to improve the model on every one.
What the researchers actually did
The automated system copies how a human researcher works, just faster and at scale. Each loop runs through the same steps:
- Searches the existing research literature
- Proposes a method to fix the target behavior
- Trains the model on that method for 30 minutes
- Keeps what works, throws out what doesn’t, and repeats
Over several iterations, benchmark scores climb. Because the system discards dead ends automatically, it can churn through far more attempts than any human team could manage in the same window.
The results
The numbers are the part worth sitting with. The system, which the paper calls the Automated Alignment Researcher (AAR), improved performance on all 10 benchmarks without dragging down the model’s overall performance. That last detail matters. Fixing one bad behavior often breaks something else, and the system avoided that trap.
Then it went head to head with people. Here’s how the comparison shook out, per the paper:
- Quality: “The best AAR method beats what experienced humans propose, on average within six hours.” The paper adds that “human guided research directions do not lead to stronger performance.”
- Cost: An AAR runs “roughly $4 per hour in API inference against the $150 per hour we pay our human researchers.”
So it’s faster, cheaper, and in this narrow task, better. “Overall, these results provide early evidence that automated alignment post-training could become practical in the near term,” the paper reads.
Why this matters
What stands out here is the direction, not just the result. This is a step toward recursive self-improvement, the idea that models could improve their own training, then improve how they improve, and so on. Many in the field see that as the next big jump in AI progress.
The paper doesn’t dodge the uncomfortable implication. If models can sharpen their own alignment training today, it’s plausible they handle broader training work tomorrow. At that point the role of the human AI researcher starts to shrink. Anthropic put the human-versus-machine comparison right in the paper, cost line included, so they clearly want people thinking about it.
The limitations
The researchers are upfront about where this breaks down, and practitioners should hold onto these caveats:
- The system only works as well as the benchmarks it targets. If a benchmark doesn’t capture the real alignment goal, the AAR optimizes for the wrong thing.
- Building and maintaining those benchmarks is hard, ongoing work.
- The automated researcher draws on existing literature, so that body of research has to be kept current and expanded too.
In other words, the AI can run the research loop, but humans still define the target and feed the pipeline. The judgment about what “good” means hasn’t been automated. Not yet.
What you can do with this
If you build or fine-tune models, the practical takeaway is about your evaluation setup. This paper is a reminder that automated post-training is only as trustworthy as your benchmarks. Before you hand any loop the keys, ask whether your evals actually measure the behavior you care about, because an automated optimizer will exploit every gap you leave.
And if you follow the industry strategically, watch this thread. An early result from a fellows program is not a shipped product, but it’s a signal of where Anthropic is pointing. The economics alone, $4 against $150 an hour, tell you why labs are racing here. You can read the full paper and Anthropic’s framing at the original TechCrunch AI report.