Anthropic AI Safety Fellow
interview guide
Prepare for implementation assessments, unfamiliar model code and a discussion of your safety research ideas, with guidance on how the fellowship varies by cohort.
What to expect at Anthropic
Anthropic's fellowship interviews test both implementation skill and research judgement. Accounts include progressive database work, NumPy debugging and tasks in an inference codebase. A research discussion asks you to turn a broad safety interest into something you could investigate, while practical rounds require you to understand and work within the supplied environment.
Assessment content, round order and tool permissions change across cohorts. The guide identifies which format each account describes rather than presenting them as one fixed process.
Making a research idea testable
Choose a safety question you could investigate, then explain the observation that makes it worth studying. Narrow the idea until you can describe an experiment, the evidence it would produce and what you would conclude from different outcomes. A broad interest in alignment is useful motivation, but it does not yet tell the interviewer what work you would do.
Rehearse a short discussion in which someone challenges your interpretation. Be ready to explain what else could account for the result and which follow-up would help distinguish the explanations. Include an outcome that would make you change your mind, so the proposal has a purpose beyond confirming your starting belief.
Inside the full guide
7 sectionsFellowship routes by cohort
How implementation, debugging, research and take-home formats differ between cohorts.
Application and preparation materials
Research interests, proposed work and checking the supplied technical environment.
Implementation assessments
Progressive tasks, dense specifications and the limits of what is known about each format.
NumPy debugging and codebase reading
Array shapes, incorrect results and finding the relevant path through unfamiliar code.
Research discussion and LLM implementation
Preparing a testable safety idea and explaining an inference calculation.
Tools and progression
Cohort-specific permissions and why individual scores are not admission thresholds.
What to practise first
Balancing implementation, debugging and research rehearsal for your invitation.