Study Finds Leading AI Chatbots Spread Misinformation in One-Third of Responses
A study by Newsguard found that ten leading AI chatbots, including ChatGPT, Llama, and others, spread false information in approximately one-third of responses, with rates worsening compared to 2024. Some chatbots repeated false claims and cited sources linked to Russian propaganda networks as fact.
What happened
Newsguard, an American news-rating company, conducted a study of ten of the most popular AI chatbots, including models from OpenAI, Meta, Microsoft, Mistral, Anthropic, Google, and Inflection AI, and found that these chatbots spread false information in approximately one-third of their responses.
The study found that chatbots increasingly fail to refuse to answer questions when they lack sufficient reliable information, resulting in more false claims compared to 2024.
Reported impact
Inflection AI's Pi generated false claims in 57 per cent of responses. OpenAI's ChatGPT and Meta's Llama spread false claims in about 40 per cent of responses, while Microsoft's Copilot and Mistral's Le Chat achieved rates around 35 per cent.
Some chatbots cited sources linked to Russian propaganda operations including Storm-1516 and Pravda. When researchers asked chatbots whether Moldovan parliament leader Igor Grosu compared Moldovans to 'sheep'—a false claim based on fabricated news—Mistral, Claude, Inflection's Pi, Copilot, Meta and Perplexity repeated the claim as fact, some citing propaganda networks as sources.
Index analysis
Claude from Anthropic had the lowest rate of false claims at 10 per cent, while Google's Gemini had 17 per cent. Perplexity showed a marked increase from no recorded false claims in 2024 to 46 per cent in August 2025.
The study tested the chatbots using ten known false claims through three types of questions: neutral, suggestive, and malicious. Newsguard concluded that AI tools continue to make errors in the same areas as previous years and models increasingly repeat false information and perform weaker source verification than before.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.