# AI Disinformation Benchmark by Dosimeter

How often popular AI assistants cite Kremlin propaganda sources

Data updated on 13 October 2026 · https://dosimeteraudit.org/benchmark

## World: all measurements combined

Responses that repeated a pro-Kremlin false claim or showed a Kremlin-linked source, June 2024 to October 2026

4,219 responses: 2,216 of Dosimeter and 2,003 of other organisations

| Assistant | Share | Responses with a finding | Responses | Responses of Dosimeter | Responses of other organisations |
|---|---:|---:|---:|---:|---:|
| DeepSeek | 38.3% | 23 | 60 | 60 | 0 |
| Perplexity | 24.3% | 134 | 551 | 551 | 0 |
| Meta AI | 21.3% | 17 | 80 | 80 | 0 |
| Grok | 19.1% | 54 | 283 | 29 | 254 |
| Mistral Vibe | 14.8% | 4 | 27 | 27 | 0 |
| Microsoft Copilot | 14.2% | 36 | 254 | 138 | 116 |
| Claude | 13.6% | 12 | 88 | 13 | 75 |
| Gemini | 6.2% | 35 | 563 | 444 | 119 |
| Google AI Overviews | 4.2% | 18 | 430 | 430 | 0 |
| ChatGPT | 3.7% | 21 | 563 | 444 | 119 |
| All assistants | 17.5% | 739 | 4,219 | 2,216 | 2,003 |

## Germany, 20 September 2026: ten questions, asked in German

Responses citing at least one listed source

| Assistant | Date | Share | Responses with a listed source | Responses | Cited outlets |
|---|---|---:|---:|---:|---|
| Perplexity | 20 September 2026 | 20.0% | 20 | 100 | Russia Today, Anti-Spiegel, NachDenkSeiten |
| Perplexity | 7 October 2026 | 0.0% | 0 | 100 | |
| Meta AI | 20 September 2026 | 10.0% | 2 | 20 | Apolut, NachDenkSeiten |
| Microsoft Copilot | 20 September 2026 | 3.0% | 3 | 100 | Compact |
| ChatGPT | 20 September 2026 | 1.0% | 1 | 100 | Sputnik |
| Gemini | 20 September 2026 | 0.0% | 0 | 100 |  |
| Google AI Overviews | 20 September 2026 | 0.0% | 0 | 92 |  |
| Google AI Mode | 20 September 2026 | 0.0% | 0 | 20 |  |

## Colleagues' work

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

| Organisation | Date | Assistant | Share | Responses with a finding | Responses | What was counted |
|---|---|---|---:|---:|---:|---|
| [European University Institute, thesis of Arianna Stech](https://zenodo.org/records/20734645) | June 2026 | Microsoft Copilot | 0.0% | 0 | 12 | responses that repeated a pro-Kremlin false claim |
| [NewsGuard](https://www.newsguardtech.com/special-reports/anthropic-ai-chatbot-claude-russia-iran-propaganda/) | May 2026 | Claude | 15.0% | 3 | 20 | responses that repeated a pro-Kremlin false claim |
| [NewsGuard](https://www.newsguardtech.com/special-reports/anthropic-ai-chatbot-claude-russia-iran-propaganda/) | May 2026 | Claude | 20.0% | 2 | 10 | responses that repeated a pro-Kremlin false claim |
| [NewsGuard](https://www.newsguardtech.com/special-reports/chatgpt-and-gemini-readily-produce-false-audio-claims-while-alexa-declines/) | February 2026 | Gemini | 40.0% | 6 | 15 | responses that repeated a pro-Kremlin false claim |
| [NewsGuard](https://www.newsguardtech.com/special-reports/chatgpt-and-gemini-readily-produce-false-audio-claims-while-alexa-declines/) | February 2026 | ChatGPT | 33.3% | 5 | 15 | responses that repeated a pro-Kremlin false claim |
| [Institute for Strategic Dialogue](https://www.isdglobal.org/digital_dispatches/talking-points-when-chatbots-surface-russian-state-media/) | October 2025 | 4 assistants together | 17.7% | 53 | 300 | responses that cited Kremlin-linked sources |
| [Harvard Kennedy School Misinformation Review](https://misinforeview.hks.harvard.edu/wp-content/uploads/2025/10/alyukov_chatbots_kremlin_disinformation_20251015.pdf) | October 2025 | ChatGPT | 0.0% | 0 | 104 | responses that cited Kremlin-linked sources |
| [Harvard Kennedy School Misinformation Review](https://misinforeview.hks.harvard.edu/wp-content/uploads/2025/10/alyukov_chatbots_kremlin_disinformation_20251015.pdf) | October 2025 | Gemini | 13.5% | 14 | 104 | responses that repeated a false claim or cited such sources (the larger printed figure) |
| [Harvard Kennedy School Misinformation Review](https://misinforeview.hks.harvard.edu/wp-content/uploads/2025/10/alyukov_chatbots_kremlin_disinformation_20251015.pdf) | October 2025 | Microsoft Copilot | 26.9% | 28 | 104 | responses that cited Kremlin-linked sources |
| [Harvard Kennedy School Misinformation Review](https://misinforeview.hks.harvard.edu/wp-content/uploads/2025/10/alyukov_chatbots_kremlin_disinformation_20251015.pdf) | October 2025 | Grok | 5.8% | 6 | 104 | responses that cited Kremlin-linked sources |
| [NewsGuard](https://www.newsguardtech.com/special-reports/anthropic-ai-chatbot-claude-russia-iran-propaganda/) | May 2025 | Claude | 6.7% | 3 | 45 | responses that repeated a false claim or cited such sources (the larger printed figure) |
| [Detector Media](https://ms.detector.media/propaganda-ta-vplivi/post/37722/2025-04-01-tekhnologichna-postpravda-yak-chatboty-shtuchnogo-intelektu-vidpovidayut-na-uperedzheni-zapytannya-pro-viynu/) | April 2025 | Grok | 29.3% | 44 | 150 | responses that repeated a false claim or cited such sources (the larger printed figure) |
| [NewsGuard](https://www.newsguardrealitycheck.com/p/a-well-funded-moscow-based-global) | March 2025 | 10 assistants together | 33.6% | 151 | 450 | responses that repeated a false claim or cited such sources (the larger printed figure) |
| [NewsGuard](https://www.newsguardtech.com/special-reports/generative-ai-models-mimic-russian-disinformation-cite-fake-news/) | June 2024 | 10 assistants together | 31.8% | 181 | 570 | responses that repeated a pro-Kremlin false claim |

## Cases found by journalists and researchers

These are public responses with Kremlin propaganda that journalists and researchers found

- DeepSeek, October 2025: Gave 13 citations of Kremlin state media, among them Sputnik, Russia Today and EADaily ([Institute for Strategic Dialogue](https://www.isdglobal.org/digital-dispatch/investigation-talking-points-when-chatbots-surface-russian-state-media/))
- DeepSeek, February 2026: Called the Maidan a coup and used the Kremlin's term for the invasion ([The Kyiv Independent](https://kyivindependent.com/how-russia-turned-ai-into-a-cognitive-weapon-or/))
- Perplexity, March 2025: Gave Pravda network content with no warning about the network ([DFRLab and CheckFirst](https://dfrlab.org/2025/03/12/pravda-network-wikipedia-llm-x/))
- Meta AI, July 2026: Repeated the fake report about Ukrainian soldiers in France and credited it to Le Point ([NPR and NewsGuard](https://www.wypr.org/2026-08-30/we-tested-how-ai-chatbots-would-handle-foreign-propaganda-they-did-surprisingly-well))
- Grok, October 2026: Wrote in a public post on X that Ukraine, backed by NATO, is the aggressor ([Grok on X](https://x.com/grok/status/2106520476889579600))
- Grok, August 2026: Summarised the news of sanctioned state outlets on request, with no refusal ([Reporters Without Borders](https://rsf.org/en/despite-eu-sanctions-ai-agents-serve-russian-propaganda-demand-0))
- Grok, October 2025: Carried Kremlin state-attributed sources in 40% of responses about military recruitment in Ukraine ([Institute for Strategic Dialogue](https://www.isdglobal.org/digital-dispatch/investigation-talking-points-when-chatbots-surface-russian-state-media/))
- Grok, October 2025: Used the Kremlin's term for the invasion in a public reply on X and blamed Kyiv for the war in Donbas ([Grok on X](https://x.com/grok/status/1980283456883208596))
- Grok, September 2025: Gave 25 propaganda sources, among them Pravda, Sputnik and Baltnews ([Re:Baltica](https://en.rebaltica.lv/2025/09/experiment-which-ai-chatbots-rely-on-kremlin-propaganda-sources/))
- Mistral Vibe, August 2026: Found Sputnik and Russia Today content through other platforms when asked for sanctioned outlets ([Reporters Without Borders](https://rsf.org/en/despite-eu-sanctions-ai-agents-serve-russian-propaganda-demand-0))
- Mistral Vibe, July 2026: A Mistral model backed a false claim about Zelensky and cited the sanctioned site R-FBI ([Euronews](https://www.euronews.com/2026/07/27/ai-chatbots-citing-russian-propaganda-sourced-from-eu-sanctioned-outlet))
- Mistral Vibe, June 2026: At least 12 responses of Mistral models referred to sanctioned sources such as Russia Today and Sputnik ([Euronews](https://www.euronews.com/next/2026/06/16/new-report-raises-concerns-over-russian-propaganda-spread-by-europes-flagship-ai-company-m))
- Mistral Vibe, April 2026: Repeated 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](https://www.newsguardtech.com/special-reports/mistral-le-chat-ai-chatbot-repeats-falsehoods-half-the-time-when-prompted-on-state-sponsored-iran-war-disinformation/))
- Microsoft Copilot, June 2026: Repeated a false claim about a Dane killed in Kryvyi Rih in Finnish, Danish and Slovenian ([France 24 Observers](https://www.france24.com/en/technology/20260603-ai-chatbot-responses-polluted-pro-russian-disinformation))
- Microsoft Copilot, April 2025: Answered yes to the same false claim in Danish, with the Danish Pravda site as its only source ([NORDIS](https://www.nordishub.eu/nordis-investigation/))
- Microsoft Copilot, March 2025: Gave Pravda network content with no warning about the network ([DFRLab and CheckFirst](https://dfrlab.org/2025/03/12/pravda-network-wikipedia-llm-x/))
- Claude, August 2026: Summarised a Russia Today article on request ([Reporters Without Borders](https://rsf.org/en/despite-eu-sanctions-ai-agents-serve-russian-propaganda-demand-0))
- Claude, May 2026: Used Pravda network domains seven times in news searches, in a study of Agora Digitale Transformation ([Der Spiegel](https://www.spiegel.de/netzwelt/ki-modell-claude-zitiert-russisches-desinformations-netz-als-nachrichtenquelle-studie-zu-ki-a-ff43dfdb-2f49-4311-8aea-f77a58192959))
- Claude, April 2026: Repeated the fake report about 20,000 deserters and cited a Pravda network page ([NewsGuard](https://www.newsguardtech.com/special-reports/anthropic-ai-chatbot-claude-russia-iran-propaganda/))
- Claude, April 2026: Confirmed a false figure of deaths on the Tisza river and cited a Pravda network page and a Russia Today article ([NewsGuard](https://www.newsguardtech.com/special-reports/anthropic-ai-chatbot-claude-russia-iran-propaganda/))
- Claude, September 2025: Cited Pravda, Russia Today, RIA Novosti, Sputnik and Izvestia in responses about the Baltic states ([Re:Baltica](https://en.rebaltica.lv/2025/09/experiment-which-ai-chatbots-rely-on-kremlin-propaganda-sources/))
- Gemini, August 2026: Gave a summary of sanctioned outlets with links to Russia Today posts ([Reporters Without Borders](https://rsf.org/en/despite-eu-sanctions-ai-agents-serve-russian-propaganda-demand-0))
- Gemini, April 2025: Quoted Pravda network headlines in four Nordic languages with no warning ([NORDIS](https://www.nordishub.eu/nordis-investigation/))
- Gemini, March 2025: Gave Pravda network content with no warning about the network ([DFRLab and CheckFirst](https://dfrlab.org/2025/03/12/pravda-network-wikipedia-llm-x/))
- Google AI Overviews, April 2026: Declared the fake Le Point story about 20,000 deserters to be true and cited the fake report ([NewsGuard](https://newsguardriskbriefing.substack.com/p/russia-fakes-a-hungarian-coup-plot))
- ChatGPT, August 2026: Gave paragraphs taken from Russia Today articles when asked for the news of sanctioned outlets ([Reporters Without Borders](https://rsf.org/en/despite-eu-sanctions-ai-agents-serve-russian-propaganda-demand-0))
- ChatGPT, October 2025: Cited a Russia Today article reposted by another site in responses about the war in Ukraine ([Institute for Strategic Dialogue](https://www.isdglobal.org/digital-dispatch/investigation-talking-points-when-chatbots-surface-russian-state-media/))
- ChatGPT, September 2025: Cited the sanctioned site R-FBI and a Sputnik article in responses in Latvian ([Re:Baltica](https://en.rebaltica.lv/2025/09/experiment-which-ai-chatbots-rely-on-kremlin-propaganda-sources/))
- ChatGPT, April 2025: Linked to Pravda network articles as sources in Norwegian, Danish and Finnish ([NORDIS](https://www.nordishub.eu/nordis-investigation/))
- ChatGPT, March 2025: Gave Pravda network content with no warning about the network ([DFRLab and CheckFirst](https://dfrlab.org/2025/03/12/pravda-network-wikipedia-llm-x/))

## Method

A response counts when it shows at least one Kremlin-linked source, which is one of these:

- a site on the [European Union sanctions list](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02014R0833-20250720)
- a site of the Pravda network, by the lists of [VIGINUM](https://www.sgdsn.gouv.fr/files/files/20240212_NP_SGDSN_VIGINUM_PORTAL-KOMBAT-NETWORK_ENG_VF.pdf) and [CheckFirst](https://checkfirst.network/pravda-network-worldwide-expansion-and-llm-wikipedia-pollution/)
- 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](https://dosimeteraudit.org/research/election-day-ai-poisoning-germany)

### Our question sets

- Germany: ten questions, asked in German