Studies over the past year find frequent errors in AI answers. Research on students suggests that verification habits can be taught, but that they weaken when people are in a hurry.

AI chatbots regularly produce answers that sound authoritative but contain errors. Recent research suggests the people relying on them often do not check, especially when they are short on time.

The technical cause is well understood. The US National Institute of Standards and Technology (NIST) explains that generative systems create outputs by predicting patterns in data, which can produce accurate text but also factual errors and inconsistencies. These fabricated or unsupported details are commonly called hallucinations.

How often errors occur

The largest study to date on news answers was coordinated by the European Broadcasting Union and led by the BBC. It had journalists from 22 public service media organizations evaluate more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity, working in 14 languages. Overall, 45 per cent of responses contained at least one significant issue, and 81 per cent had some form of problem.

Sourcing was the most common flaw, affecting 31 percent of responses. Accuracy problems affected 20 percent and missing context 14 percent. The study was published in October 2025, and its questions were put to the assistants in late May and early June of that year.

A more recent test focused on fact-checking. UK fact-checker Full Fact spent five months sending neutral questions about claims its reporters were checking to ChatGPT, Gemini and Grok, and reported its results on August 28, 2026. It identified 39 major errors, with 67 separate incorrect responses across the models. Most models wrongly said an AI-generated image of a Ryanair cabin window was real, and others placed pictures and footage in the wrong country.

In some cases the chatbots corrected themselves days later, often citing Full Fact's articles, but in others they kept giving wrong answers. The trial used a selected set of disputed claims and has not been independently replicated, so it does not establish a general error rate.

Stanford researchers reported in April that advanced reasoning modes gave minimal improvement in fact-checking, and web search helped more but inconsistently.

Time pressure and verification

A study published in January in a ScienceDirect journal looked at whether students could be trained to check ChatGPT's output. Students who received a warning about ChatGPT's flaws verified answers more often in a task involving summaries of academic sources, though the warning had no significant effect on their stated intention to verify. Time pressure reduced fact-checking, and the authors said training must address the trade-off between speed and accuracy.

A high-stakes example

The gap between an AI answer and a checked one has already shown up in government. CNN reported on September 18 that a chatbot used by a US special operations analyst inaccurately identified a ship's cargo, and officials found the error only shortly before a planned operation.

What comes next

Use is growing. The Reuters Institute's 2025 Digital News Report found that 7 percent of online news consumers use AI assistants for news, rising to 15 percent of under-25s.

Full Fact's conclusion is that chatbots cannot be relied on as a foolproof way to check claims, especially in breaking news. Stanford's researchers offered similar advice: be cautious when using chatbots to verify claims, particularly about politics.