Training AI models with other AI models has become a shared goal across the newer labs. A researcher in Anthropic's fellows program has now given an early look at what that might mean in practice.

On Friday the company published a paper titled "Automated Researchers Can Reliably Mitigate Alignment Failures," detailing how AI systems could reliably improve a model's performance on a set of alignment benchmarks.

The result: given ten benchmarks for specific misaligned behaviours, the automated systems improved performance on every single one — without degrading overall performance.

How it works

Led by Anthropic fellow Chen Yueh-Han, the system replicates much of the traditional approach to research. Each automated system:

  • Searches the available literature,
  • Proposes a method,
  • Trains the model using that method for 30 minutes,
  • Gradually raises the benchmark over several iterations.

Effective methods are preserved while ineffective ones are discarded, allowing the system to operate quickly and at scale.

In the paper's own words: "Overall, these results provide early evidence that automated alignment post-training could become practical in the near term."

Toward recursive self-improvement

The paper is a step toward recursive self-improvement, which many see as the next significant move in AI progress. The chain of reasoning: if models can improve their own alignment training, it is plausible they could improve training practices more broadly — at which point human AI researchers might soon become obsolete.

The paper does not shy away from this, explicitly comparing the Automated Alignment Researcher to its human equivalent: "The best AAR method beats what experienced humans propose, on average within six hours. Human guided research directions do not lead to stronger performance."

For anyone still unconvinced, there is a cost comparison: "An AAR costs roughly $4 per hour in API inference against the $150 per hour we pay our human researchers."

Why alignment first

It is no accident that this experiment happened in alignment. Alignment is one of the areas densest in measurable benchmarks: how prone a model is to deception, sycophancy or harmful output can be reduced to a number. That is precisely what an automated researcher needs — a score, and a target to raise it against.

It is also why the approach does not generalise easily. A similar system can be built anywhere the metric is clear; where the metric is contested, an automated researcher runs fast in the wrong direction.

The limits

The paper lists its own limitations, and they are not minor.

The automated system works only insofar as the benchmarks reflect the actual alignment goals. The system passes the exam it is given; whether that is the right exam is a separate question, and humans still answer it.

Beyond that, establishing and maintaining those benchmarks is serious work in itself — as is maintaining and expanding the literature the automated researchers draw from.

So the picture is not "human researchers are finished" but a shift in where the weight of the work sits: from producing methods to defining the target. A human still decides what is being measured; the machine now finds the best method against that measure.