On 3 November 2026, I had the pleasure of moderating a session at the Expert Roundtable on Digital Human Rights Tracking Tools and Databases, held at the Palais des Nations in Geneva and online. The roundtable brought together developers and users of these tools from across the field, under the theme “From AI experimentation to responsible integration.”
The day moved from how AI is already used across the field, to reliable everyday use, to interoperability. My session, “Ethical use of AI in human rights work,” asked the harder question: when should AI be used in this work, and when should it not?
Framing the question from standards. At ITU-T, ten Study Groups develop technical standards by consensus, and more than 700 of them now relate to AI. In 2025, with OHCHR, we surveyed the Study Groups: 84% of respondents see human rights as relevant to their work, but 40% report low or no familiarity with what that means in practice. A comparison of eight organisations, including ISO, IEC, IEEE, IETF, ETSI and the Council of Europe, pointed to the same lesson: integration is feasible when it is framed in technical and procedural terms, not legal ones.
This is the logic behind the draft guideline for addressing human rights in ITU-T standard development, a Member State contribution now under discussion. It is light-touch and optional: 15 screening questions, asked when new work is proposed, when a draft is stable and before approval. Would affected users be able to understand when and how the technology affects them? Are there mechanisms to detect, report and address errors, misuse or harm? The same questions apply to AI in human rights monitoring: when to use it, when not to, who is affected, and what happens when something goes wrong.
Ethics has to become a set of operational choices made early, not a check on a finished tool.
AI prepares, people decide. The short updates covered very different tools but converged on one design principle. In one case, AI cross-references recommendations with legislation, budgets and development goals and drafts a diagnosis, but it does not decide what compliance means; nothing becomes validated information until the responsible institution has verified and signed it. In another, a model sorts incoming reports and drafts case briefs, while researchers check every case against several sources and review each brief before use. The reasoning is institutional: a recommendation that sits in a database changes nothing; what changes practice is an institution discussing it and taking it as its own.
Other contributions widened the lens. Semantic search can help users who do not know the mechanisms or the right keywords. Three questions were put to any team building these tools. What does the design do to us? Friendly, confident chat assistants invite trust, and even experts tend to defer to them; a human in the loop is only a safeguard if that human can and does check. Whose values does the tool reflect? Models trained mostly on Western, English-language data may sound aligned with human rights and still miss what discrimination looks like in a given community. And should we automate this at all? AI can reduce exposure to distressing material, but reading a case carefully or listening to a survivor is also an act of care and witnessing.
A challenge to the room. One participant asked whether we were meeting our own standard of do no harm. Using AI to draft treaty body lists of issues risks undermining the independence of the experts whose role is to frame those questions, and AI-drafted state reports risk the same lack of ownership, and the same weak implementation, long seen with reports drafted by external consultants. Concerns were also raised about the values of the companies behind the models, biased outputs, military uses of AI and its environmental cost. A recurring point was that human rights, not ethics alone, must be the floor, because these obligations are legally binding.
Others extended the argument rather than opposing it. Much of this work exposes staff to secondary trauma, and some automation, such as knowing what an image shows before opening it, would protect them. Yet some communities choose not to automate, out of respect for the people whose cases they document, so the choice belongs to those doing the work. Reading, as one participant put it, is also a way of hearing. Several contributions stressed the many steps between a searchable database and AI-written text: a research layer can speed up analysis without any text generation, which should remain a last resort. Making data, not only tools, interoperable, through data contracts that state what quality is guaranteed, could build trust without recentralising. And civil society, like some public institutions, can build its own tools rather than only consume those of large technology companies.
Three points stayed with me from the session.
Human oversight has to be designed, not declared. The most convincing tools in the room were those where human review is a step with a name on it: a focal point’s signature, a researcher’s approval, a link to every source.
The question is less whether to use AI than at which step. Searching, sorting and flagging can reduce workloads and exposure to distressing material. Drafting, deciding and witnessing are different, especially where independence and ownership are what make an outcome legitimate.
Values are also a question of infrastructure. Whose model, trained on what data, governed by whom, and where sensitive information goes are not technical details for practitioners working with survivors or in restrictive settings. As I wrote after a recent Human Rights Council side event, those who depend on systems built elsewhere risk inheriting assumptions they did not shape.
These are, in substance, the questions the ITU-T screening approach asks of standards. Seeing practitioners put them to their own tools, early and in operational terms, was the most encouraging part of the afternoon.
The roundtable closed with a session to agree shared messages and practical proposals for the RightsX Summit on 18-19 November 2026 and the Road to the Geneva AI Summit on 21-22 June 2027, to be captured in an outcome note validated by participants. If one idea from the afternoon should travel with them, it is that AI can take on the volume, while the judgment and the care stay with people.
My thanks to the Geneva Human Rights Hub and all the co-organisers for the invitation, and to the speakers and participants, in the room and online, for a frank and generous exchange.