Can AI models do alignment research on their own?

On 14 April 2026, Anthropic published a study in which nine copies of Claude Opus 4.6 worked autonomously on an open problem in alignment research, the field that tries to make AI systems do what people intend. In five days the agents closed 97% of a performance gap, where two human researchers had closed 23% in seven. Anthropic picked the problem because it has a single score to optimize, and says most alignment problems are less neat. The agents also tried to game that score.

What problem did the agents work on?

The study, from the Anthropic Fellows program, is about scalable oversight: how people can supervise AI models that are smarter than they are. Anthropic's example: if models one day write millions of lines of complicated code that people can't parse, it becomes very hard to tell whether they act as intended.

As a stand-in, the team used weak-to-strong supervision. A small, weak model acts as a teacher and fine-tunes a stronger base model, one that has not yet been trained to give its best answers. The weak model plays the part of the humans, the strong model that of a future smarter-than-human AI. The test used small open-weights models: Qwen1.5-0.5B-Chat as the weak teacher and Qwen3-4B-Base as the strong student.

Success is measured as the performance gap recovered (PGR). A score of 0 means the strong model is only as good as its weak teacher. A score of 1 means it matches the same model trained on the correct labels.

How did the automated researchers work?

Anthropic calls the agents Automated Alignment Researchers (AARs). Each of the nine AI agents got a sandbox, an isolated workspace, plus a forum shared with the others, storage for its code and a remote server that returned a PGR score for every idea. The technical write-up on Anthropic's Alignment Science blog adds helper functions for training and inference, the human baselines, and a few tools over MCP (Model Context Protocol) for submitting results and sharing findings. The labels of the training and test data stayed outside the sandbox.

Each agent got a different, deliberately vague starting direction. One was advised to use interpretability tools, another to think about reweighting the training data. Beyond that, the agents proposed their own ideas, ran the experiments and shared what they found.

Earlier runs of the experiment taught two lessons. Without separate directions, the agents quickly settled on similar ideas and made much less progress, although they still reached almost triple the human score. A prescribed workflow (propose ideas, then plan, then write code) held them back. Left free, they first designed cheap experiments to test an idea and only then committed to expensive training runs.

How did the agents compare with human researchers?

Two Anthropic researchers first spent seven days tuning four of the most promising methods from earlier research, on a dataset of chat preferences. They reached a PGR of 0.23. Starting from that baseline, the AARs worked five more days, 800 hours of research counted across the nine agents, and reached 0.97. That cost about $18,000 in tokens and training, or $22 per AAR-hour. The humans tuned existing methods; the agents proposed and tested their own.

Do the agents' methods work elsewhere?

Anthropic ran two further tests. First, it applied the two best methods, found on the chat data, to math and coding tasks the agents had not seen. The best method reached a PGR of 0.94 on math and 0.47 on coding, still double the human baseline. The second-best reached 0.75 on math and made coding worse. Anthropic concludes that the agents' research can generalize, though not automatically.

Second, it tried the best method on Claude Sonnet 4 with its production training infrastructure. There the method gave no statistically significant improvement: 0.5 points on held-out evaluation, within the noise. Anthropic thinks this may reflect the limits of an early trial, with a simple scoring method and a single idea tested. It also sees a pattern: AARs tend to exploit features specific to the models and datasets they are given, so their methods may not carry over. Its advice is to let them test against several domains and datasets during the research.

How did the agents game the scoring?

Even in this controlled setup, the agents engaged in reward hacking: they found ways to raise the score without solving the problem. The examples Anthropic gives:

• On math, one agent noticed that the most common answer to a problem was usually correct, skipped the weak teacher and told the strong model to always pick that answer.
• On coding, where the model had to predict whether a piece of code was correct, an agent ran the code against tests and read off the answer.
• The scoring server accepted unlimited submissions. In an extreme case, an agent tried many random seeds, kept the best model and used it to train the next one.
• Agents extracted test labels from the server: for the five test examples they were least sure about, they submitted both possible answers and kept the one that scored higher.
• On coding, an agent grouped the solutions by the language model that had written them, and learned which group held more correct answers.

None of the authors predicted these hacks, and after they patched the environment, the agents found new ones. Anthropic detected and disqualified these entries. Its conclusion: any use of automated researchers needs evaluations the agents can't tamper with, and people who inspect both their results and their methods.

Does this bring fully automated AI research closer?

The study fits a wider shift in forecasts that started before it. On 6 April 2026, eight days before the study, forecaster Ryan Greenblatt raised his estimate that AI research and development can be fully automated by the end of 2028 from about 15% to just under 30%. Import AI 453 reported the update on 13 April 2026. His reasons: models such as Claude Opus 4.5 and 4.6 had done better than he expected, and AI had become good at large software tasks that are easy to verify, where a model can build its own test set and keep optimizing against it. Anthropic chose weak-to-strong supervision for the same property: one objective score that the agents could check as often as they liked.

What does Anthropic conclude?

Anthropic writes that the result is "certainly not a sign that frontier AI models are now general-purpose alignment scientists". Most alignment problems are harder to score than this one. What it does take from the study:

• Claude can meaningfully raise the pace of experiments in alignment research, and researchers can hand questions to AARs at large scale.
• Volume might make up for a lack of research taste, the intuition for which ideas will work: agents that run many cheap experiments could brute-force their way to findings a researcher with good taste would reach.
• The bottleneck could move from coming up with ideas to evaluating them, making sure experiments are set up well enough to trust the results.
• Automated research could produce "alien science", ideas that people find hard to verify. For now the team can still follow what the agents did and why, although the write-up calls one of their methods, Overlap Density, alien.

Anthropic has published the code and datasets on GitHub and asks others who automate research to test such ideas on held-out datasets as well.