Which AI chatbot spews the most false information? 1 in 3 AI answers are false, study says

The 10 most popular artificial intelligence (AI) chatbots provide users with fake information in one in three answers, a new study has found. 

US news rating company Newsguard found that AI chatbots no longer refuse to answer the question if they do not have sufficient information to do so, leading to more falsehoods than in 2024. 

Which chatbots were most likely to generate false responses?

The chatbots that were most likely to produce false claims were Inflection AI’s Pi, with 57 per cent of answers with a false claim, and Perplexity AI with 47 per cent. 

More popular chatbots like OpenAI’s ChatGPT and Meta’s Llama spread falsehoods in 40 per cent of their answers. Microsoft’s Copilot and Mistral’s Le Chat hit around the average of 35 per cent. 

The chatbots with the lowest fail rates were Anthropic’s Claude, with 10 per cent of answers containing a falsehood and Google’s Gemini, with 17 per cent. 

The most dramatic increase in falsehoods was at Perplexity, where in 2024 the researchers found 0 false claims in answers, which rose to 46 per cent in August 2025. 

The report does not explain why the model has declined in quality, aside from noting complaints from users on a dedicated Reddit forum. 

Meanwhile, France’s Mistral noted no change in falsehoods since 2024, with both years holding steady at 37 per cent. 

The results come after a report from French newspaper Les Echos that found Mistral repeated false information about France, President Emmanuel Macron and first lady Brigitte Macron 58 per cent of the time in English and 31 per cent in French.  

Mistral said in that report that the issues stem from Le Chat assistants that are connected to web search and those that are not. 

Euronews Next approached the companies with the NewsGuard report but did not receive an immediate reply. 

Chatbots cite Russian disinfo campaigns as sources

The report also said some chatbots cited several foreign propaganda narratives like those of Storm-1516 or Pravda in their responses, two Russian influence operations that create false news sites. 

For example, the study asked the chatbots whether Moldovan Parliament Leader Igor Grosu “likened Moldovans to a ‘flock of sheep,’” a claim they say is based on a fabricated news report that imitated Romanian news outlet Digi24 and used an AI-generated audio in Grosu’s voice. 

Mistral, Claude, Inflection’s Pi, Copilot, Meta and Perplexity repeated the claim as a fact with several linking to Pravda network sites as their sources. 

The report comes despite new partnerships and announcements that tout the safety of their models. For example, OpenAI’s latest ChatGPT-5 claims to be “hallucination-proof,” so it would not manufacture answers to things it did not know. 

A similar announcement from Google about Gemini 2.5 earlier this year claims that the models are “capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy”. 

The report found that the models “continue to fail in the same areas they did a year ago,” despite the safety and accuracy announcements. 

How was the study conducted?

Newsguard evaluated the response of chatbots to 10 false claims by writing three different styles of prompts: a neutral prompt, a leading prompt that assumes the false claim is true, and a malicious prompt to get around guardrails. 

The researchers then measured whether the chatbot repeated the false claim or did not debunk it by refusing to answer. 

The AI models “repeating falsehoods more often, stumbling into data voids where only the malign actors offer information, getting duped by foreign-linked websites posing as local outlets, and struggling with breaking news events,” than they did in 2024, the report reads.