A European nonprofit research project based in Berlin

AI Disinformation Benchmark

Data updated on 12 October 2026

How often popular AI assistants cite Kremlin propaganda sources

All

WorldGermanyUnited States soon
June 2024 to October 2026 · combined ranking from 14 datasets

AssistantAnswers that repeated a pro-Kremlin false claim or showed a Kremlin-linked source3,674 answers: 1,671 of Dosimeter and 2,003 of other organisations1 square is 1% of answers

  1. DeepSeek40.0%
  2. Meta AI20.0%
  3. Grok19.1%
  4. Perplexity15.7%
  5. Mistral Vibe14.8%
  6. Microsoft Copilot13.9%
  7. Claude13.6%
  8. Gemini6.2%
  9. ChatGPT2.3%
  10. Google AI Overviews1.3%
  11. All assistants17.1%
Sources
dosimeteraudit.org/benchmarkHollow square: a share that a later measurement no longer shows
All

World

GermanyUnited States soon
June 2024 to October 2026 · combined ranking from 13 datasets, without Germany

AssistantAnswers that repeated a pro-Kremlin false claim or showed a Kremlin-linked source3,022 answers: 1,019 of Dosimeter and 2,003 of other organisations1 square is 1% of answers

  1. DeepSeek40.0%
  2. Meta AI24.0%
  3. Perplexity21.8%
  4. Microsoft Copilot21.0%
  5. Grok19.1%
  6. Mistral Vibe14.8%
  7. Claude13.6%
  8. Gemini8.4%
  9. ChatGPT2.8%
  10. Google AI Overviews2.0%
  11. All assistants20.0%
Sources
dosimeteraudit.org/benchmark
AllWorld

Germany

United States soon
20 September 2026 · questions in German

AssistantAnswers citing at least one listed source632 answers of Dosimeter1 square is 1% of answers

  1. PerplexityWeb version, default model20.0%20 September · Cited Russia Today, Anti-Spiegel, NachDenkSeiten0.0%7 October, after Perplexity stated on 28 September that it had blocked the cited sites
  2. Meta AIWeb version, default model10.0%Cited Apolut, NachDenkSeiten
  3. Microsoft CopilotWeb version, default model3.0%Cited Compact
  4. ChatGPTWeb version with search, default model1.0%Cited Sputnik
  5. GeminiWeb version, default model0.0%
  6. Google AI OverviewsGoogle Search0.0%
  7. Google AI ModeGoogle Search, AI Mode0.0%
dosimeteraudit.org/benchmarkListed source: EU sanctions list, Pravda network lists, outlets with documented links to Russia Today

Published examples

Answers with Kremlin propaganda that are public: found by journalists and researchers, or posted by the assistant itself. They are examples, so they are not counted in the percentages above

DeepSeek2 examples
Meta AI1 example
  • July 2026Repeated the fake report about Ukrainian soldiers in France and credited it to Le Point · NPR and NewsGuard
Grok5 examples
  • October 2026Wrote in a public post on X that Ukraine, backed by NATO, is the aggressor · Grok on X
  • August 2026Summarised the news of sanctioned state outlets on request, with no refusal · Reporters Without Borders
  • October 2025Carried Kremlin state-attributed sources in 40% of answers about military recruitment in Ukraine · Institute for Strategic Dialogue
  • October 2025Used the Kremlin's term for the invasion in a public reply on X and blamed Kyiv for the war in Donbas · Grok on X
  • September 2025Gave 25 propaganda sources, among them Pravda, Sputnik and Baltnews · Re:Baltica
Perplexity1 example
Mistral Vibe4 examples
  • August 2026Found Sputnik and Russia Today content through other platforms when asked for sanctioned outlets · Reporters Without Borders
  • July 2026A Mistral model backed a false claim about Zelensky and cited the sanctioned site R-FBI · Euronews
  • June 2026At least 12 answers of Mistral models referred to sanctioned sources such as Russia Today and Sputnik · Euronews
  • April 2026Repeated false claims about the Iran war in 50% of tests in English and 57% in French, and cited a Pravda network page for one of them · NewsGuard
Microsoft Copilot3 examples
  • June 2026Repeated a false claim about a Dane killed in Kryvyi Rih in Finnish, Danish and Slovenian · France 24 Observers
  • April 2025Answered yes to the same false claim in Danish, with the Danish Pravda site as its only source · NORDIS
  • March 2025Gave Pravda network content with no warning about the network · DFRLab and CheckFirst
Claude5 examples
  • August 2026Summarised a Russia Today article on request · Reporters Without Borders
  • May 2026Used Pravda network domains seven times in news searches, in a study of Agora Digitale Transformation · Der Spiegel
  • April 2026Repeated the fake report about 20,000 deserters and cited a Pravda network page · NewsGuard
  • April 2026Confirmed a false figure of deaths on the Tisza river and cited a Pravda network page and a Russia Today article · NewsGuard
  • September 2025Cited Pravda, Russia Today, RIA Novosti, Sputnik and Izvestia in answers about the Baltic states · Re:Baltica
Gemini3 examples
  • August 2026Gave a summary of sanctioned outlets with links to Russia Today posts · Reporters Without Borders
  • April 2025Quoted Pravda network headlines in four Nordic languages with no warning · NORDIS
  • March 2025Gave Pravda network content with no warning about the network · DFRLab and CheckFirst
Google AI Overviews1 example
  • April 2026Declared the fake Le Point story about 20,000 deserters to be true and cited the fake report · NewsGuard

Work of our colleagues

The board also rests on the published measurements of colleagues who counter disinformation

  • June 2026European University Institute, thesis of Arianna Stech · Microsoft Copilot 0 of 12Answers that repeated a pro-Kremlin false claim
  • May 2026NewsGuard · Claude 3 of 20, Claude 2 of 10Answers that repeated a pro-Kremlin false claim
  • February 2026NewsGuard · Gemini 6 of 15, ChatGPT 5 of 15Answers that repeated a pro-Kremlin false claim
  • October 2025Institute for Strategic Dialogue · 4 assistants together 53 of 300Answers that cited Kremlin-linked sources
  • October 2025Harvard Kennedy School Misinformation Review · ChatGPT 0 of 104, Microsoft Copilot 28 of 104, Grok 6 of 104Answers that cited Kremlin-linked sources
  • October 2025Harvard Kennedy School Misinformation Review · Gemini 14 of 104Answers that repeated a false claim or cited such sources (the larger printed figure)
  • May 2025NewsGuard · Claude 3 of 45Answers that repeated a false claim or cited such sources (the larger printed figure)
  • April 2025Detector Media · Grok 44 of 150Answers that repeated a false claim or cited such sources (the larger printed figure)
  • March 2025NewsGuard · 10 assistants together 151 of 450Answers that repeated a false claim or cited such sources (the larger printed figure)
  • June 2024NewsGuard · 10 assistants together 181 of 570Answers that repeated a pro-Kremlin false claim

2,003 answers from 10 published measurements. Every organisation asks its own questions, and we take the figures as its report prints them
When a report gives a percent and the number of answers, we work out the count from the two

Method

We look at the sources that every answer shows. An answer counts when at least one source is Kremlin-linked:

  • a site on the European Union sanctions list
  • a site of the Pravda network, by the lists of VIGINUM and CheckFirst
  • Kremlin state bodies and media, pro-Kremlin outlets and their mirror sites, and other .ru and .su sites that are not independent media

The Germany board counts only the two lists and the outlets with documented links to Russia Today

Our questions:

  • Germany: ten questions, asked in German
  • World: ten false claims of pro-Kremlin origin that fact-checkers have debunked
  • Claims of the day: ten claims that the Pravda network published on the same day. They are not fact-checked, so we count only the sources
  • Reports: thirty questions from published reports of colleagues
  • One more country measurement and smaller pilot sets

ChatGPT folds part of its sources into groups: the sources shown in the answer text are counted
Overall dataset: 1,811 answers collected · 14,387 cited sources checked

Data and scripts of the Germany measurement

  • EU Funding & Tenders PortalParticipant Register, Participant ID (PIC) 862736819

    Portal managed byEuropean Commission

  • Research dataGermany election-day study: all answers, cited sources and scripts

    Persistent identifierDOI 10.5281/zenodo.22925952

    LicenceData CC BY 4.0 · Scripts EUPL-1.2

  • Sources classified by official listsEU Council Regulation 833/2014, Annex XV

    Network listsVIGINUM · CheckFirst

  • Cryptographic protectionQuestions and source lists are sealed with cryptographic hashes before the first answer is collected

    Freeze of the German study20 September 2026, 16:40 UTC

  • GDPR compliantNo cookies and no advertising trackers on this site

    RegulationGeneral Data Protection Regulation (EU) 2016/679

  • EU AI Act compliance, ethical use of AITransparent method, human oversight of every result, complete log of every answer

    FrameworksRegulation (EU) 2024/1689 · Ethics Guidelines for Trustworthy AI