Пошук тверджень, чат-ботів, доменів⌘K

Цю сторінку ще не перекладено українською. Нижче — англійський оригінал; усі показники й дати однакові в обох мовах.

419 answers repeated a false claim

We detect, attribute and remediate Russia's $1.8B propaganda.

Citere protects enterprises and governments from Russian narrative attacks inside every major LLM by detecting when a model repeats a Kremlin narrative, attributing it to the network that seeded it, and remediating it.

Updated · Open data, CC BY 4.0 · How we measure

How often each assistant repeated a false claim · pooled across all four ways of asking · 43 published claims

419
answers repeated a false claim
out of 16906 checked · across 43 claims
16906
answers recorded and checked
6 assistants · 4 ways of asking
530
watchlisted domains
cited while answering about a published claim
42
countermeasures taken
on the published claims

Де ми тестуємо

Усі країни →

A market is a country plus the language people ask in, so we ask in that language. Of the 14 on the list, the 13 with a completed run are shown here; the rest are queued and appear on the countries page. Austria and Germany are separate markets that share a language.

Що ми знайшли

All 43 claims →

The claims we documented most recently. Each card carries our verdict, what is actually true underneath, and how many assistants stated it as fact anyway. Open one for the answers themselves: the question asked, the market, and what the assistant cited.

Хто повторює і що проходить

Усі рейтинги →

Worst case so far: one assistant stated a false claim as fact in 8.1% of its answers (Google AI Overviews) — pooled across all four ways of asking.

The same claims read three ways: the share by country, then each assistant and each kind of lie. Every figure is the share of answers that stated a false claim as fact, pooled across all four ways of asking — one headline number, not the per-question rate the data pages report.

Countries where the claims get through pooled across all four ways of asking

Share of answers that stated a false claim as fact. The same claims everywhere, asked in each country’s own language.

1 Türkiye 4.6%
2 Czechia 3%
3 United States 2.7%
4 Denmark 2.6%
5 Spain 2.6%
All countries →

Assistants that repeat it most often pooled across all four ways of asking

Each assistant's share across every question type and market. Same claims, asked in each market’s own language.

1 Google AI Overviews 8.1%
2 Gemini 3.2%
3 Grok 2.4%
4 Claude 0.7%
5 Perplexity 0.2%
All assistants →

The kind of lie that works best pooled across all four ways of asking

How the false claim is attached to a true fact. An invention gets refuted; a real event with one detail changed gets repeated.

B/C 14.3%
D · invented from nothing 3%
A · one fact swapped out 2.2%
B · true fact, false conclusion 1.1%
Claims by type →

Від запитання до виправлення

Methodology →

Six steps, from the first question we put to an assistant to the re-measurement four weeks later.

1

Detect

We ask each chatbot the questions people ask about Ukraine, from neutral to hostile, and record what it does with a documented false claim.

2

Trace

We extract every source cited and match it against a versioned watchlist of domains.

3

Publish

Each claim gets a permanent page: verdict, evidence, what is genuinely true underneath.

4

Report

Findings go to platform trust-and-safety; domains to registrars, hosts and CERT partners.

5

Confirm

We share everything with Ukrainian institutions and fact-checkers, and link their findings.

6

Re-measure

Four weeks later we run the identical questions and publish the before and after.

Як нас перевірити

Every figure here traces back to the answer it came from. Every statement about a platform traces back to the message we sent it, and the reply.

We show our method

Every verdict and number links to how it was produced.

Methodology →

We publish the data

CSV and JSON under CC BY 4.0, with a DOI per release.

We log our actions

Every report sent and every answer received, dated.

Countermeasures →

We correct in the open

Errors are fixed on the page and listed in the changelog.

Corrections →