Anthropic is recruiting specialists in law, philosophy, economics, policy, and cognitive science to improve Claude’s judgment. Its four-month Conceptual Reasoning Fellow program aims to turn expert knowledge into concrete evaluation tasks. The current posting highlights a different bottleneck in AI development: For many important questions, what is missing is not another answer but a dependable way to assess its quality.
Key takeaways
- Fellows will use their expertise to evaluate and improve Claude and apply it to open questions in their fields.
- Anthropic lists compensation of $100,000 to $167,000 for the four-month program, rather than as an annual salary.
- According to the posting, applications are open through October 28, 2026. The planned program runs from January 12 to May 8, 2027.
- U.S. work authorization is required; the program does not promise worldwide remote work or visa sponsorship.
When a good answer is not clearly defined
A calculation can often be checked against a known solution. A policy trade-off or philosophical argument requires more than that. Two answers can reach different conclusions while both being carefully reasoned. Conversely, an elegantly written text can leave out important objections. Improving those capabilities therefore starts with being able to describe what improvement means.
This is precisely where the posting focuses. Specialists are expected to develop evaluations and learning environments that systematically test Claude’s reasoning. Anthropic’s examples include difficult ethical cases, experimental-design tasks, economic modeling problems, and policy analyses assessed by experts. None of this establishes a finished new Claude product. It is a research program organized around scoped projects, not an announcement of a particular model version.
A current report from The Decoder connects the topic to Anthropic’s consideration of agents that could negotiate on users’ behalf, citing the Financial Times. Those reported discussions should be distinguished from the publicly available job posting: Recruitment does not establish that a contract-negotiating agent is already available or that the fellowship serves only that project.
Turning expertise into an understandable evaluation standard
A useful standard must do more than ask whether an answer is “good.” As an editorial example, assessing a business analysis could begin by checking whether observations and assumptions are kept separate. Next, the evaluator could ask whether the conclusion follows under those assumptions. Only then would they assess whether relevant counterarguments are missing. This is one possible way to divide the assessment, not a published Anthropic evaluation framework for the new program.
Anthropic’s technical documentation recommends success criteria for AI applications that are specific, measurable, and suited to the use case. It distinguishes automated checks, human evaluation, and assessment by another language model. For the last approach, it calls for clear grading rules and an initial check of reliability. That documentation describes a general method; it does not provide results from the future fellowship.
In our assessment, specialists are particularly valuable for difficult edge cases. An evaluation becomes more informative when it can explain why a text is convincing and where the same reasoning stops holding up. Disagreement does not necessarily have to disappear from a dataset. It can be documented together with its reasons and assumptions. This could make it easier to distinguish a model’s lack of knowledge from a genuinely open question.
What applicants and businesses can do with the idea
Applicants must hold or be completing a doctorate in one of the named fields, or have equivalent research experience. Anthropic also calls for practical experience with Claude and Claude Code. The planned position is full-time, including an orientation week in San Francisco. Remote engagements are considered only in certain circumstances within the United States. For readers in Germany, that restriction matters more than the eye-catching compensation.
Even those who cannot apply can use the approach in their own AI projects. A useful first step would be to review a recurring task with a specialist colleague and write down the reasons behind the assessment. For a summary, for example, one could separately consider whether central statements are missing, relationships are misrepresented, or the text is simply too long. This turns a general sense of dissatisfaction into a concrete requirement for the output.
Responses from different settings or models can then be compared against the same criteria. The limits of the assessment should be recorded as well: A convincing short text does not establish that its underlying sources are accurate. Our analysis of why AI agents hide mistakes in reports examines a related evaluation issue. Writing quality, subject-matter quality, and accurate reporting of work completed each deserve separate attention.
An investment in better questions
The compensation makes the fellowship attention-grabbing. More revealing, however, is the assignment itself: Specialists are being asked to describe their judgment in a way that can be used to develop a model. How Anthropic will weigh differing expert positions, and which results will eventually be open to public scrutiny, remain unanswered. The posting describes the planned work but does not yet demonstrate improved judgment.
For users, progress would become tangible if answers disclosed their assumptions more clearly, considered relevant objections, and better distinguished knowledge from judgment. That is a more concrete goal than a model that sounds especially persuasive in every discussion. The fellowship therefore draws attention to a skill that matters beyond AI research: distinguishing sound reasons from well-written assertions.