Mission
To make it measurable — and therefore fixable — when AI assistants repeat Russian disinformation to the people who trust them.
The gap we exist to close
Fact-checking was built for posts and articles: things with a URL, a timestamp, a share count and a moderator. A chatbot answer has none of that. It is generated once, for one person, and then it is gone. When ChatGPT, Gemini, Grok, Claude, Google AI Overviews or Perplexity repeats a Russian fabrication to a reader in Germany at nine in the evening, there is no artefact to check, no platform report to file, and no record that it happened.
We create that record. We ask these assistants the questions real people ask, in Ukraine's partner-country markets, and we document what they answer about claims we have already verified as false.
What we are trying to change
Our operating cycle is Detect → Attribute → Remediate, and each commitment below is one step of it made public.
- Make it visible. A permanent, public, machine-readable page for every documented claim an assistant has repeated.
- Make it addressable. Findings go to the platform that can fix them, and to the registrars and hosts that carry the source.
- Make it accountable. Four weeks after a disclosure we re-run the identical prompts and publish the change — including when nothing changed.
- Make it reusable. Open data, an open method, and no requirement to trust us over the numbers.
Why from Ukraine
Ukraine is where these narratives are aimed first and tested hardest, and where the ground truth is closest to hand. That is an unusual observation post, and we think it produces better data than the same work would produce anywhere else. It also means we are not neutral about the harm — but we are strict about the method, because a finding that does not survive scrutiny is worth nothing to anyone.
What success looks like
Not a headline. A repeat-rate that falls and stays down, on claims we can name, on assistants we can name, verified by a measurement anyone can reproduce.