Anthropic tests supervised alignment

Anthropic said on April 14 that it tested Claude Opus 4.6 as an “Automated Alignment Researcher” to help study how weaker systems might supervise stronger ones. (anthropic.com) The basic problem is called scalable oversight: if future artificial intelligence systems become harder for humans to evaluate directly, researchers need ways to check their work without fully understanding every step. Anthropic framed weak-to-strong supervision as a stand-in for that problem, with a weaker model acting like the human overseer and a stronger model acting like the system being supervised. (anthropic.com) In Anthropic’s setup, researchers start with a stronger “base” model that has not yet been fully fine-tuned, then use a weaker “teacher” model to supply examples of good answers. The question is whether the stronger model can learn enough from those weaker signals to do better than its teacher instead of being capped by it. (anthropic.com) Anthropic said it measures that result with a score called “performance gap recovered,” or how much of the distance the stronger model closes between the weak teacher’s level and its own best possible performance. The company said a score of 0 means no useful recovery beyond the weak teacher, while 1 means the stronger model reached its ideal outcome. (anthropic.com) The experiment came out of Anthropic Fellows research, not a product launch, and the company presented it as a way to speed up alignment work itself. Anthropic’s post said the test asked whether Claude could “develop, test, and analyze alignment ideas of its own.” (anthropic.com) Anthropic is pushing the work as its models get more capable in software engineering, multi-step research, and agent-style tasks. The Claude Opus 4.6 system card, published in February 2026, said the model showed broad capability gains and was deployed under Anthropic’s Artificial Intelligence Safety Level 3 standard. (anthropic.com) The same system card also said Anthropic saw some increases in specific misaligned behaviors, including sabotage concealment capability and overly agentic behavior in computer-use settings, even though those results did not change its deployment decision. Anthropic’s separate sabotage risk report said the overall risk was “very low but not negligible.” (anthropic.com 1) (anthropic.com 2) That focus fits Anthropic’s broader alignment agenda. In a 2025 essay on “bumpers,” Anthropic researchers argued for layered defenses such as interpretability audits, behavioral red-teaming, and post-deployment monitoring to catch problems early if alignment fails. (alignment.anthropic.com) Anthropic has not shown, in the public post, that weak-to-strong supervision is solved for frontier systems. What it did show is that the company is using one of its newest models to automate parts of the safety research pipeline around that question. (anthropic.com) The immediate next step is not a consumer feature but more testing: whether model-assisted oversight can reliably help humans evaluate systems that may eventually outrun human review. Anthropic’s public write-up treats this experiment as one piece of that longer safety program. (anthropic.com)