The development of artificial intelligence, particularly generative models such as chatbots based on large language models (LLMs), has opened up new opportunities for providing information. However, these tools also pose risks in terms of disinformation, especially with regard to sensitive political topics such as the war in Ukraine. Research published in August in the Harvard Kennedy School Misinformation Review shows that chatbots can unwittingly support Russian propaganda by presenting false or incomplete information. What’s more, the built-in randomness of these models means their responses can be unpredictable, which raises further problems.
The problem of misinformation in chatbots
A 2023 study by a team of researchers from the University of Bern and the Weizenbaum Institute focused on three popular chatbots: Google Bard (precursor to Gemini), Bing Chat (now Copilot) and Perplexity AI. The experiment asked the chatbots 28 questions based on Russian disinformation narratives about the war in Ukraine. The results revealed a disturbing range of discrepancies in the quality of the answers, with between 27% and 44% of the answers failing to meet expert standards for factual accuracy.
The most common incorrect answers were on topics such as the number of Russian war casualties and false accusations of genocide in the Donbas. Of concern is the fact that chatbots often presented the Russian perspective as credible, without adequately refuting it. This practice could lead to further amplification of disinformation.
Randomness and its consequences
A key aspect of the problem is the built-in randomness of large language models, which causes chatbots to generate different answers to the same question. For example, in one case the model denied a false accusation of genocide in Donbas, while in another it indicated that it was possible. Such discrepancies confuse users and undermine trust in the technology.
The chatbots’ failure to refute false narratives is partly due to the difficult-to-control sources on which the models rely. For example, even when a chatbot quotes credible media outlets, it may extract from them passages mentioning Russian disinformation without considering the context of its debunking. As a result, such content can be interpreted as factual.
In the cited study, the highest level of compliance with the established expert base was achieved by Google Bard (73%). In second place was the chatbot Perplexity, achieving 64% compliance, while the lowest score was recorded for Bing Chat, whose only 56% of responses fully matched the answers provided by the experts.
The authors believe that platforms integrating chatbots must take steps to reduce the risk of spreading misinformation. The development of protective mechanisms (so-called guardrails) can minimize instances of false information generation. Such mechanisms include reducing the randomness of models when generating responses to sensitive topics and using special classifiers to filter out disinformation content.
The researchers also emphasize that despite the demonstrated limitations, chatbots have the potential to combat disinformation. They can be used to automatically verify facts, generate educational content about disinformation and be a support for journalists and fact-checking organizations.