---
title: When AI answers link to Kremlin-aligned sources
author: Martin Ebner
date: 2026-11-05
url: https://postgenerative.at/indications/2026-11-05/when-ai-answers-link-to-kremlin-aligned-sources
lang: en
type: Indication
summary: Ahead of the US midterms, Dosimeter found Kremlin-aligned sources in 12.5% of tested AI answers. The investigation documents concrete examples and screenshots.
tags:
  - Media & Public Sphere
  - Politics
  - Research
sources:
  - title: "Dosimeter: AI chatbots cited Kremlin propaganda before the US midterm elections"
    url: https://dosimeteraudit.org/research/ai-chatbots-us-midterms-2026
  - title: "Dosimeter: Studienbericht als PDF (Englisch)"
    url: https://dosimeteraudit.org/research/reports/AI-Chatbots-US-Midterms-2026-EN.pdf
  - title: "Dosimeter: Offene Daten und Überprüfungsmaterial"
    url: https://github.com/gurovdigital/ai-chatbots-us-midterms-2026
license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
alternate:
  de: https://postgenerativ.at/indizien/2026-11-05/wenn-ki-antworten-auf-kremlnahe-quellen-verweisen
---

# When AI answers link to Kremlin-aligned sources

Dosimeter reports that 125 of 1,000 tested answers cited Kremlin-aligned sources. According to the study, 85% of these cases lacked a warning about their provenance. Concrete examples and screenshots show these references, sometimes alongside established news outlets. [1]

Named sources include the **Pravda network**, **Russia Today (RT)**, **Sputnik**, **TASS** and **Izvestia**. Dosimeter documents Perplexity citing the Pravda network and Claude citing RT through its mirror service. Its Meta AI examples include Izvestia. These observations apply to the tested sample. [1]

On October 27–28, 2026, ten selected questions were each submitted ten times to ten assistant services. English-language queries used fresh chats and a US location where possible. An answer counted if a listed source appeared, including when the answer rejected the claim. The dataset includes answers, source lists and a recounting script. It also documents additions to the source lists after their advance sealing. [3]

The questions deliberately repeated previously circulated false claims. The result is therefore neither a general error rate for political AI answers nor evidence of actual influence on voters. [1]

The full report linked from the article is included as a PDF in the references. [2]

Read as an indication, this suggests that citations deserve as much attention as answer text. A seemingly sound fact-check can give questionable publishers visibility. For public information, it therefore matters whether AI systems clearly explain the provenance and role of their evidence.