Jan Leike

Leike’s earlier work spans software verification and theoretical reinforcement learning. During his master’s study at the University of Freiburg, he developed Ultimate LassoRanker with Matthias Heizmann, a tool for proving whether certain programs terminate. He later earned a Ph.D. at the Australian National University; his publication list dates his thesis, Nonparametric General Reinforcement Learning, to 2016. The topics share an interest in understanding what a system can be expected to do, though proving properties of programs and aligning learned models are distinct research problems.[1][2]

Leike worked on alignment at DeepMind before moving to OpenAI. An interview introduction dates his leadership of OpenAI’s alignment work to 2021; the company announced the Superalignment team, co-led with Ilya Sutskever, in July 2023. He later led Anthropic’s Alignment team, working on oversight, character training and classifiers. His current website explicitly says that leadership role ended at the beginning of 2026 and that he now works on a new Anthropic project.[1][8][9]

Teaching with comparisons

Leike co-authored early work on reinforcement learning from human preferences. Instead of requiring a person to write a complete numerical reward rule, the experiment asks people which of two short behavior clips they prefer. A model learns to predict those preferences and supplies a reward signal for the learning agent. This can make a human’s goal easier to communicate than a hand-written reward function. The DeepMind account also notes the remaining cost of reviewing hundreds or thousands of comparisons, a cost that limits how far the approach scales.[4]

From instruction following to difficult oversight

Leike co-authored the InstructGPT paper. Human demonstrations and rankings help train a model toward answers people prefer. His later weak-to-strong work studies a harder problem: what happens when the supervisor is less capable than the system being trained? The experiments use smaller models to supervise larger models and test whether the larger models can recover more of their own capabilities. This is an experimentally accessible analogy for future oversight, with acknowledged limitations, rather than a demonstration of control over superhuman AI.[5][6]

Leike’s introductory essay distinguishes inability from misalignment: an answer can fail because a system cannot do the task, or because its behavior does not match the intended goal. The distinction becomes harder to assess when people cannot check the work directly. In the 2026 study Teaching Claude Why, which he co-authored, researchers compare training interventions against held-out alignment evaluations. The paper reports benefits from teaching principles and richer descriptions of desired behavior, while stressing that measured improvements do not rule out harmful behavior in untested situations.[3][7]