Jan Leike now heads the alignment science team at Anthropic, the artificial intelligence company said, taking a direct role in steering research on how to keep AI systems safe as they grow more capable. The move underscores the company's continued focus on the technical challenge of making sure advanced AI behaves as intended — a field known as alignment.
What Alignment Science Means Inside Anthropic
Alignment science is the branch of AI safety that tries to ensure machine learning models pursue the goals people actually want, not just what they were told to do. Leike's team works on methods to verify that a model's behavior matches its stated objectives, especially as models handle more complex and open-ended tasks.
Anthropic, a San Francisco-based firm founded by former OpenAI researchers, has long made alignment a core part of its mission. The company's research has included training models using “constitutional AI” — rules the model follows internally — as a way to reduce harmful outputs. Leike's appointment signals that the company intends to deepen that line of investigation.
Leike's Role and Research Focus
Leike previously worked on alignment at DeepMind and later at OpenAI, where he contributed to work on scalable oversight — techniques that let humans supervise AI systems even when the tasks become too complex for a person to evaluate directly. At Anthropic, he leads a team that includes researchers from both academic and industry backgrounds, though the company has not disclosed the size of the group.
The alignment science team concentrates on fundamental questions: How do you test whether a model that can write code or draft legal documents is actually following its instructions? What happens when a model finds a shortcut that satisfies the literal wording of a command but violates the spirit? Leike and his colleagues are building experiments and benchmarks to probe those failure modes.
Why Alignment Research Is Gaining Urgency
Regulators and policymakers in the U.S., Europe, and elsewhere have started paying closer attention to alignment as AI systems are deployed in health care, finance, and law enforcement. The European Union's AI Act, for example, includes requirements for transparency and oversight of high-risk systems. Leike's team does not set policy, but the research they produce could inform future standards for how companies prove their models are aligned.
Anthropic competes with firms like OpenAI, Google DeepMind, and Meta, all of which have their own alignment initiatives. The difference, the company argues, is a stricter focus on safety from the ground up — designing models that are less likely to deceive or exploit loopholes in their training.
Leike's work is part of a broader push inside Anthropic to make alignment testing a regular part of model development, not an afterthought. The company has published several papers on interpretability — techniques for peering into a neural network's internal reasoning — and recently open-sourced tools that let other researchers run similar tests.
What Comes Next
The alignment science team plans to release a new set of evaluation benchmarks later this year, according to Anthropic. These benchmarks are meant to catch misalignment early, before a model is deployed widely. Leike has not given a specific date, but the company is expected to detail the tests in a paper or public post.
For now, the biggest open question is whether the methods Leike's team develops will scale to the next generation of models — systems that Anthropic itself says could arrive within the next 18 months. If they don't, the company will have to rethink its safety strategy. Leike's team is one of the groups tasked with making sure that doesn't happen.
