Politicians on AI

We collect documents from official government sources, such as parliamentary records, ministry press pages and institutional APIs, and split them into utterances by a single speaker. Those are filtered down to statements that mention AI, using keyword lists in each of the 11 languages our sources are written in. Three LLM judges then score and tag every surviving utterance independently, and it enters the dataset only where two of the three agree that it is substantive, belongs to one speaker, and that the speaker is a public official. See the method page for the full pipeline, including the judge prompts and the human spot-check results, or the repository for the code.

Quotes are fetched from official government websites, for example parliamentary records. For a full list, see the method page. Abridged quotes are checked against the source programmatically.

Countries are included based on volume of AI-related discourse and availability of official quote data. Some sources had to be excluded due to copyright. Contact us if you would like to see a country or intergovernmental organization added.

This project was made by Lennart Finke at ETH Zurich together with Markov Grey at CeSIA. You can reach out with feedback and questions at lennart@finke.dev.

Yes. Tags are assigned via a majority vote of three LLM-based judges, based on definitions from existing taxonomies of AI risk (from the MIT AI Risk Repository) and AI regulation (from the Existential Risk Observatory’s AI Governance and Regulatory Archive), as well as some overarching keywords we defined. For details and definitions of the tags, see the method page and the tag definitions.

No. We did however spot-check agreement with human ratings. See the method page for details. When a tag seems inaccurate, it could be about the larger utterance the short quote is drawn from; check the source text link to be sure. Speaker names and profile details are not all human-checked either.

Not for comparisons between jurisdictions. We aim to include the most salient sources for each country, but it could be that we missed some important AI discussion for some jurisdictions. Our data should be mostly representative over time within one jurisdiction, however; if within the EU the share of quotes about public investment increases from 2023 to 2026, we recommend interpreting that as a real effect.

Some of the data sources publish with a delay, which can produce an illusion of dropoff of number of quotes in recent months. This is especially true for the United States data (such as the Congress and House hearings), with some delays as long as a year.

Yes. The code for this page and the code for data collection are both open-source.

Yes, see the Downloads section below.