An agentic tool for research data management: making it safe

Slides and accompanying notes for a short talk given at the Verification, Testing and Agents workshop, School of Computer Science, University of Sheffield, on 5 November 2026 organised by Professor Phil McMinn and featuring a talk from Professor Mark Harman, a Research Scientist at Meta London. The Data Director is an AI assistant, specified in the Research Data Alliance's Data Director Agentic AI Blueprint, that helps researchers make their data FAIR: choosing a repository, selecting metadata standards, drafting metadata and documentation, and preparing a deposit. The talk describes how a pilot reference implementation of the Blueprint tries to make such an assistant safe. The AI is not trusted to report on itself: a separate harness checks every answer against the trusted registries it cites (such as FAIRsharing and re3data), withholds any answer it cannot account for, and records every run. Declining and referring a question to a person are treated as correct outcomes, not failures. Research data stay on the researcher's computer; only a summary is sent. The talk is explicit about the limits of these checks: they show that an answer is accounted for, not that it is right, which still needs human judgement. It closes with open questions on checking an assistant's reasoning, testing requirements written in plain English, evaluating without known right answers, and indirect prompt injection. The notes are written for data stewards and the wider research community and assume no technical background. Use of generative AI. Claude Opus 5.5, a large language model made by Anthropic, was used to help prepare the slides and these notes. It drafted text and slide content from the author's outline and from the Data Director project's documentation and source code, and it checked figures and links against the repository. The author decided what to include, reviewed and edited all of the content, and is responsible for it.