Anthropic seeks PhDs to train and test Claude in 2027

Anthropic has opened applications for two four-month fellowships focused on Claude’s weaknesses in specialized fields. Running from January to May 2027, the programs will embed PhD-level researchers within Anthropic to create training tasks, expose reasoning failures, and design evaluations for future models.

The STEM Fellowship targets mathematicians, physicists, life scientists, and engineers. The Conceptual Reasoning Fellowship covers philosophy, cognitive science, economics, policy, and law, where expert judgment often matters more than a single verifiable answer.

Turning expertise into training infrastructure

Fellows will build the materials Anthropic uses to train and assess Claude, working on practical model-development infrastructure instead of pursuing independent academic projects. The expected outputs include:

• Datasets: Carefully constructed examples that teach or test domain-specific reasoning.
• Evaluations: Hard-to-game tests that reveal whether a model can solve a problem reliably.
• Simulations: Controlled settings for tasks such as planning experiments, interpreting data, and reasoning about mechanisms.
• Failure analysis: Systematic investigations of cases where Claude produces confident, plausible, and incorrect answers.
• Training environments: Sandboxed tasks with outcomes or expert rubrics that provide feedback during reinforcement learning.

The conceptual reasoning track places particular weight on evaluation design for questions that require nuanced judgment. Anthropic lists several representative projects:

• A philosopher constructing difficult cases in ethics or formal epistemology.
• A cognitive scientist designing experimental-design and model-comparison tasks.
• An economist writing modeling and identification problems with known answers.
• A policy scholar drafting regulatory and cost-benefit analysis tasks with expert scoring rubrics.
• A legal scholar building datasets related to AI policy and law.

Reinforcement-learning environments turn an expert’s standards into repeatable feedback. A model attempts a task, receives a score or structured critique, and uses that signal during training. Creating reliable feedback becomes harder in fields such as law or ethics, where several answers may be defensible and the quality of the reasoning determines the result.

Four months, full time

Anthropic will pay each fellow between $100,000 and $167,000 for the four-month term, equivalent to an annualized rate of $300,000 to $500,000, plus benefits. Participants will work through an employer of record under fixed-term contracts.

The Claude Code requirement signals that academic credentials alone will not qualify an applicant. Fellows must already know how to work with Anthropic’s coding agent and use model behavior as an object of structured investigation.

A second cohort with a wider brief

Anthropic previously ran a STEM fellowship for more than 20 researchers in mathematics, physics, life sciences, and engineering from June through September 2026. Problems created by that cohort now contribute to Claude’s training and to an internal science evaluation, according to the company.

The 2027 program extends the model into disciplines where evaluation depends on defensible assumptions, explicit rubrics, and expert interpretation. Those fields create a different technical challenge: Anthropic must capture the reasoning behind a judgment while preserving room for multiple valid conclusions.

The path toward an “AI scientist”

Anthropic frames the STEM fellowship as part of its effort to build an AI scientist capable of sustained reasoning and experimental judgment. The company cites open mathematical conjectures and drug design for neglected diseases as examples of problems that require models to plan across many steps, interpret evidence, revise hypotheses, and recognize uncertainty.

Embedding domain experts with model teams lets Anthropic convert specialist knowledge into datasets, evaluators, simulations, and training signals while those systems are being designed. For researchers, the fellowships provide direct access to model-development workflows. The role also requires a full-time, four-month commitment, a mandatory week in San Francisco, and existing U.S. work authorization.