Evaluating Moral Competence in AI Systems

This position continues recent work from Violet Gordon within our Data x Direction program: developing ways to actually examine AI morality, rather than take it on faith from how a system's outputs read.

What we are looking for is the question you would pursue and how you would go about it.

Julia Haas and colleagues at Google DeepMind argue, in a 2026 Nature roadmap, that evaluating AI morality has to move past moral performance — whether a system's outputs look right — and toward moral competence: whether those outputs are produced because of morally relevant considerations.

The distinction matters because these systems are trained on enormous amounts of human writing already full of moral reasoning. Fluent, appropriate-sounding moral output is exactly what you would expect whether or not anything underneath is reasoning about it.

This is an open call for people interested in that question: what would count as evidence of moral competence in an AI system, and how would you go about finding it. We are not looking for a specific method or a predetermined answer. A proposal could be empirical, interpretive, or conceptual — drawing on interpretability and mechanistic analysis, instruments from moral psychology, evaluation and benchmark design, or philosophical work on what would count as evidence either way. Bring the approach that fits your training and interest, as long as it engages the distinction directly.

Graduate students, postgraduate researchers, early career researchers, and undergraduates prepared to work at that level. Backgrounds in cognitive science, philosophy, machine learning or interpretability, or moral psychology are all relevant, and no single one is required.

A proposal of 500 to 1000 words covering four things:

1. The question and the claim. What you would work on, and what a claim about it would look like that you could actually be wrong about.
2. Your approach. What you would do, and what you would try first to find out whether it is tractable at all.
3. A timeline for the term, from mid-September to the first week of December, in three or four phases rather than week by week, with a midpoint you could be held to.
4. What you want from the term, including whether you would want to work closely with Violet Gordon or run more independently.

Also submit a CV. A letter of recommendation is optional and there is a slot for one. Not including a letter will not count against your proposal.

These positions are unpaid. We do not offer stipends.

We will work with your university to arrange experiential learning or internship credit where your program allows it, and we welcome applicants who bring their own funding or fellowship support. If either applies to you, say so in your proposal and we will work out the arrangement.

We read for whether the claim is genuinely arguable, whether the scope is achievable within the term, whether the approach shows real familiarity with the material, and whether your own interest in the question is legible.

Every proposal is read in full by a person. Nothing in this process runs through keyword matching, automated ranking, or AI screening of any kind, and there are no recorded interviews without a person present. If we interview you, you will be talking with someone who can answer your questions.