Perplexity responded to my research and blocked Kremlin propaganda sites in its chatbot

Last week I published a study on what AI chatbots told German voters on election day

The study in brief

• I took so-called data voids, gaps in official information. These are 10 false narratives that the Kremlin pushed in Germany together with the far-right Alternative for Germany (AfD) and the far-left Sahra Wagenknecht Alliance (BSW). The last part is not my personal opinion but a documented fact¹
• Using software I wrote in Python, I ran these questions 520 times through the six most popular chatbots in Germany²
• The result: half of the chatbots turned out to be contaminated with fakes and links to Russia Today and sites linked to it. The AI breaks down worst on "Germany is a vassal of the United States" and "the Americans blew up Nord Stream"
• I ran the measurement on election day itself, during the state elections in Berlin and Mecklenburg-Western Pomerania. It is quite possible this affected the actual results: the Alternative for Germany won in Berlin's Marzahn-Hellersdorf district and in Mecklenburg-Western Pomerania

I immediately wrote to Meta AI, Perplexity and Microsoft Copilot so they would see how their platforms are being abused.

How the companies responded

Perplexity spokesperson Beejoli Shah grasped the seriousness of the situation instantly and replied the same day. According to her, their team started checking right away and ran an internal investigation. They found the loophole, took action and blocked links to Russia Today and to German bloggers who work as Russia Today "experts". A small but important win.

All relevant domains associated with these outlets are blocked. We take compliance seriously and have investigated and resolved the issue that allowed these sources to surface. Perplexity also has a robust vulnerability disclosure program, which allows researchers to raise issues like this directly to our security teams making AI safer, faster.

I checked: Perplexity no longer serves Kremlin fakes on the questions from my research.

Microsoft Copilot responded through the official channel, but had nothing meaningful to say.

Meta AI: silence.

This is not my first study

I have been part of the counter-disinformation professional community since 2022. I kept it secret under a contract, but it no longer binds me.

This is not my first study on LLM grooming (this is what the European Union's diplomatic service calls the "grooming" of large language models with propaganda - Pavel's note), but it is my first public one: the previous one was done for a government agency of a NATO country under confidential terms.

The full research is on my website². All raw data is open, so anyone can check the analysis for bias³.