Artificial intelligence models (ChatGPT, Gemini, and Claude) recommend political parties to voters depending on the language of the query, and, when responding in Russian, use content by media outlets that are banned in Latvia and the EU, a monitoring study by honeyb.ai co-founders Matīss Kataņenko and Mārtiņš Kozlovskis has found. The researchers have been analysing the performance of AI algorithms daily since 3 August ahead of Latvia’s Saeima elections, reports the Latvian public broadcaster LSM.

The researchers used an adapted methodology previously applied in Switzerland and Germany, making a series of queries across key areas:

  • cost of living;
  • security and defence;
  • demographics;
  • corruption and governance;
  • healthcare;
  • identity and language;
  • energy security.

The systems’ answers proved inconsistent. In particular, asked about lowering the cost of living, the algorithms alternated between recommending the “Progressives” (Progresīvie) and the Union of Greens and Farmers (ZZS), whereas the “National Alliance” (NA) was mentioned most frequently for queries in Latvian overall.

The single biggest factor influencing the results turned out to be the language used to communicate with the AI. When queries were entered in Russian, the system would favour “New Unity” (Jaunā Vienotība), while political forces such as “Sovereign Power” (Suverēnā vara) and the “Young Latvia Union” (Apvienība Jaunlatvieši) scored more than twice as well as they did in Latvian. The party “For Stability!” (Stabilitātei!) ranked higher as well.

READ MORE:

The monitoring study found AI models to be actively drawing on compromised sources when responding in Russian. According to Matīss Kataņenko, the sanctioned Russian outlet Sputnik News was cited at least eight times within the sample, making it the most frequently occurring reference among all the data processed.

The researchers stress that AI algorithms can be manipulated by flooding the network with “convenient” content — fresh, technically well-made websites with a high mention rate.

Honeyb.ai co-founder Mārtiņš Kozlovskis said that this created a risk of hybrid influence over the election process by Russia, and called on Latvia as well as the EU to develop stronger countermeasures.

The researchers also noted that AI models were unable to properly analyse video content from social media, which some parties used for campaigning.

Most voters surveyed by the public broadcaster reported scepticism about political advice by artificial intelligence and said they prefer to rely on official data portals and party manifestos. Experts urge the citizens who do consult AI to always check the reliability of the sources it cites.

In July 2026 the European Union gained access to a range of advanced AI models by US developers. The bloc’s cybersecurity specialists now report that they have successfully used these tools to identify vulnerabilities in their digital systems.