80% of Users Missed Chatbot Errors Because Fluent AI Sounds Smart

Eighty per cent of participants failed to spot factual errors in a chatbot's answers — even though they had been pre-selected for basic knowledge of the topic, a Radboud University study has found.

Communication scientist Marianna de Sá Siqueira and her colleagues tested how language fluency shapes the credibility of chatbots. Their finding: an eloquent AI system sharing incorrect information gets rated by users just as highly as a curt AI system giving correct answers.

The study, published in Computers in Human Behavior: Artificial Humans (2026, vol. 10, article 100412), lands as millions of people daily ask questions of ChatGPT, Gemini, Claude and other chatbots — systems that still produce so-called 'hallucinations', unreliable answers presented with great certainty as truth.

How did the experiment work?

Participants saw one of four variants of a chatbot conversation about The Lord of the Rings:

• Fluent, human-like language with correct answers
• Fluent, human-like language with fabricated facts
• Short, businesslike answers that were entirely correct
• Short, businesslike answers with subtly fabricated facts

Users rated the fluent-and-correct chatbot most positively. No surprise there. The surprise came in the middle of the comparison: a chatbot writing fluently while spouting factual nonsense scored as intelligent and pleasant as a stiff, stilted chatbot that got its facts right.

Then the harder number: 80% of participants did not notice the errors at all, despite their advance selection for basic Lord of the Rings knowledge.

"Above all, it shows how easily people can be led astray and how quickly we come to trust something that is actually spouting nonsense," said Barbara Müller, associate professor and co-author of the study. "If a text sounds elegant and natural, people automatically attribute intelligence to it. You then subconsciously tend to think: this is fine, I don't need to check it any further."

Why does fluent language beat facts?

The co-authors point to how the technology is built. Professor and co-author Tibor Bosse said: "So-called language models such as GPT and Claude are primarily trained to produce statistically plausible and easy-to-read sentences. These systems do not always take the truth very seriously. They are designed to sound human and convincing, but whether the information is factually correct is of secondary importance."

"People used to complain that chatbots were annoying and knew nothing. Now they sound so clever that we're too quick to believe them," Bosse said.

In short: fluency acts as a credibility signal, and users outsource their judgment to it.

Can a disclaimer fix it?

Bosse doesn't think a one-line warning does the job. Most chatbots show a single line at the start stating the output may not be entirely accurate, but, as he put it, "those companies are certainly not going to indicate of their own accord with every answer that the system is uncertain."

His proposal goes further: "If we really want to protect users from chatbots spouting nonsense, we need to, for example, oblige tech companies to make that uncertainty visible. If a chatbot itself expresses doubt, you prevent people from taking nonsense as the truth."

That puts the burden on regulation rather than user vigilance. For brands deploying conversational AI in customer-facing roles, the study's implication is blunt: a polished voice can mask wrong answers, and most users will never notice — until the error does visible damage.