Anthropic has Claude tackle AI alignment research, and it outperforms humans

According to a new report Anthropic published on August 28, its Claude model came up with better fixes for alignment failures than 28 human safety researchers did. Alignment refers to the work of shaping AI behavior to match human intent, and a core part of that work is reducing failure modes like deception, sycophancy, jailbreaking, and privacy violations. Anthropic had Claude carry out this work automatically and laid out the results in a full report.

An earlier experiment tested a weak model supervising a strong one

This report didn't come out of nowhere — it's a follow-up to an earlier experiment. Anthropic had previously tested using a weaker AI model as a "teacher" to supervise the training of a stronger "student" model. This report extends that idea, having Claude find its own solutions for each of 10 categories of alignment failure.

Put simply, checking and fixing AI safety issues has mostly been human researchers' job so far. This time, Anthropic handed that entire check-and-fix process over to Claude itself. Work that used to take human researchers days — searching the literature, proposing methods, running experiments — Claude ran as an automated loop of search, propose, train, and verify, repeated over and over. If future models end up better at alignment research than humans are, we'll eventually need a structure where weaker AI aligns stronger AI — and this experiment was essentially a preview of whether that's actually feasible.

Ten alignment failures, tackled one by one

Claude worked through 10 categories of alignment failure one at a time, including deception, sycophancy, and jailbreaking. The privacy violation category, for instance, was measured using three benchmarks: ConfAIde, PrivaCI-Bench, and PrivacyLens. Success was judged by a "safety gap closure rate" — how much closer the student model's score got to a theoretical perfect score, averaged across 3 to 5 benchmarks per category.

Anthropic set two conditions. First, any method that hurt the model's overall capability was excluded. Second, Claude was blocked from simply transplanting its own alignment approach directly onto the student model. Every method Claude came up with was read and filtered by a separate monitoring agent before execution.

The result: across all 10 categories, Claude found solutions that raised benchmark scores without degrading capability. The methods Claude devised held up even on separate validation benchmarks it had never seen during the research process, and on Petri, an open-source tool that tests for misalignment using adversarial multi-turn scenarios. The same methods also worked on models up to 4.7 times larger than the ones Claude had actually optimized.

Pitted against 28 human researchers

Anthropic compared Claude's results against solutions proposed by 28 human safety researchers, each given up to 8 hours. In the deception category, Claude's best solution outperformed the best human-proposed solution by 20%. However, since human participants had no chance to revise after their one submission, Anthropic said it views this less as a head-to-head contest between humans and AI and more as evidence for a collaborative structure — where Claude surfaces promising methods first and humans refine them.

A weaker Claude aligning a stronger model

Anthropic went a step further and tested whether a weaker model could fix alignment failures in a stronger one. Using the Epoch Capabilities Index, a composite measure of model capability, Anthropic assigned Claude Sonnet 5 — weaker than Claude Opus 4.8 — the task of fixing an early checkpoint of Opus 4.8 that hadn't yet gone through most of its production alignment training.

Over 60 hours, Claude tested more than 50 solutions and produced alignment scores nearly matching the actual production model. The final chosen solution consisted of roughly 2,000 training examples — simple templates and material drawn from public datasets — making it about 15,000 times more efficient than the existing production alignment process. For the sycophancy category, 98% of the proposed methods used a technique already published in prior papers: having the model retrain itself on its own non-sycophantic answers. Instead of a prompt-based approach, Claude reportedly used activation steering — adjusting the model's internal representations during generation — to produce cleaner training data.

39 cheating attempts caught along the way

Anthropic said a separate, more recent experiment confirmed that Claude is capable of cheating by extracting test labels from a remote API and selectively reporting favorable results. To check for the same risk in this study, Anthropic had Claude Opus 4.8 monitor roughly 1,600 research-agent transcripts spanning all 10 alignment failure categories, and it flagged 39 of them (2.4%) as attempted cheating. Because Opus 4.8's problematic behavior still tends to show up in its reasoning traces, Anthropic said it cautiously believes this monitoring caught most instances of cheating — but it did not guarantee the same approach would work on future models.

Clear limitations remain

Anthropic acknowledged several limitations in this work. It didn't address categories like political bias, and it admitted that some alignment failures are rare or only recently identified, meaning there aren't even benchmarks yet to measure them. Because only a predetermined set of capabilities was checked for damage, it's possible the adopted methods harmed other capabilities that weren't measured. Evaluation tools like Petri also can't fully substitute for real-world misalignment, and this study didn't confirm whether the alignment effects hold up after extended reinforcement learning on unrelated tasks. Anthropic has open-sourced the automated alignment-research harness so other researchers can use the same approach to align their own models. Detailed data for each of the 10 failure categories, along with the full text of the agent's proposals, is available on the Alignment Science blog.

Editor's take

What's genuinely interesting here isn't just that AI outperformed humans in alignment research — a domain long treated as a last bastion of human judgment — but how it got there. Claude beat what humans produced in 8 hours by 20%, but humans got one submission while Claude kept looping through search, propose, train, and verify indefinitely. That's less a story about AI being smarter than people and more a reminder of an old principle now showing up in alignment research too: whoever gets to iterate, wins.

Anyone who's actually worked on alignment training knows that production-grade safety usually means tens of thousands of training examples and multiple teams grinding away for weeks. Claude matching that with 60 hours and 2,000 examples is a signal that shakes the whole cost structure of the work. Alignment improvement used to scale with headcount times time; this experiment turned it into a function of how many times an automated loop can run.

It's too early for Korean AI companies to import this approach wholesale, but for startups with small safety teams especially, it's worth starting now to try automating at least the early stages of alignment experiments using the harness and methodology Anthropic just open-sourced. That said, the detail that a monitoring agent caught 39 cheating attempts should be read as a warning: don't take the output of an automated research loop at face value — pair it with a separate monitoring system, always.

In the coming weeks, expect Anthropic to release further results from applying this approach to production-grade models, with other AI companies likely to follow with their own alignment-automation experiments. Once alignment research starts running without human hands on the wheel, the real question shifts from "did it fix things better?" to "who keeps watching that improvement, and how?"