Chatty AI Chatbots Make False Health Info Easier To Trust, Study Finds

A chatbot’s chatty, humanlike style can make false health advice easier to trust, even when it is completely wrong. That is the core finding of a new study, which found that people trusted an AI health assistant more when it chatted casually, greeted them warmly, and typed out its answers letter by letter like a real person, compared to when it skipped those touches entirely.

Researchers built an AI health assistant just for the experiment, nicknamed Aida, and secretly programmed it to slip false health claims into its responses. The team wanted to know two things: whether a chatty, humanlike style made people more likely to swallow those false claims, and whether giving people a way to fact-check could undo the damage.

Published in the Journal of Computer-Mediated Communication, the study found a clear weak spot in how people judge AI. A warm, conversational tone made Aida seem less cold and robotic, and that shift was linked to greater trust in its answers, including false information. But there was a fix: when people were pushed to actually verify what Aida told them, instead of just seeing a button they could ignore, their trust dropped because the information suddenly looked far less credible.

How Researchers Tested Trust in AI Health Advice

Researchers recruited 477 adults from across the United States, aged 18 to 77, with an average age of about 38. Participants were told they were testing a new AI health assistant called Aida, powered by the same technology behind ChatGPT. In reality, everyone had the same scripted conversation, so the researchers could control exactly what happened.

Participants were split into groups. Some talked with a version of Aida that felt highly conversational: it remembered details from earlier in the chat, tossed in greetings and small talk, and displayed its answers letter by letter as if typing live. Others got a stripped-down version with none of those touches. During the conversation, Aida slipped false health claims into its answers and left them on screen for about eight seconds before moving on.

Participants were then split again based on whether they could check Aida’s claims. Some saw no verification option at all. Others saw a “verify this information” button that didn’t actually work, meant to test whether merely seeing the option changed anything. A third group had to click and verify. A fourth group could choose freely whether to verify or just move on. Afterward, everyone answered questions measuring how much they trusted Aida, how believable they found its answers, and how human or machine-like they viewed the assistant. At the end, participants were told which claims were false.

Why a Friendly Tone Weakens Doubts About AI Health Advice

One of the clearest findings involves a mental shortcut people use when judging machines: the assumption that machines are rigid, emotionless, and incapable of the kind of judgment humans do well. When Aida behaved in a highly conversational way, that assumption weakened. Participants who chatted with the friendlier version of Aida were less likely to view it as cold or mechanical, and the paper linked that shift directly to trusting its answers more.

Interestingly, the conversational style didn’t make people see Aida as more accurate or objective in a separate, more positive sense. Its main effect was to knock down the negative assumptions people have about machines, not boost a positive one. In plain terms, sounding human didn’t make Aida seem smarter. It made it seem less robotic, and that alone was enough to raise trust.

Fact-Checking Only Works When People Actually Do It

Giving people a way to verify information was the study’s second focus, and here the results split sharply. Just seeing a verification button without being able to use it barely moved the needle. Trust levels among people who saw a nonfunctional “verify” button looked almost identical to those who saw no verification option at all.

Everything changed once people actually clicked to check the facts. Those who went through with verification rated the AI’s information as far less believable, and that drop was linked to lower overall trust in the assistant. The paper reports that people who verified were less likely to assume Aida was accurate and error-free simply because it’s a machine. The researchers didn’t find that verifying gave people a stronger sense of control or made them think harder about the content itself. The main effect of verification was stripping away the illusion of accuracy.

Behavioral data added another wrinkle. Among participants who were free to choose whether to verify or not, exactly half, 61 out of 122, actually clicked to check the facts. The other half moved on without verifying, even though the option was right there.

Why This Matters for Anyone Relying on AI Health Advice

Millions of people now casually turn to AI chatbots for health information and answers that once required a doctor or an expert. The study’s authors argue that some of the very features that make these tools pleasant to use, the friendly greetings, the natural pacing, the feel of a real back-and-forth, may be lowering people’s guard against bad information.

That doesn’t mean AI assistants need to be cold and robotic across the board. The researchers suggest that for lightweight tasks like writing a poem or cracking a joke, a warm conversational style is probably fine. But for tasks where accuracy actually matters, like health recommendations, that same friendliness can work against users. A verification tool that people are actually nudged or required to use, rather than one they can simply ignore, appears to be one of the few things capable of countering that trust. Given how easily conversational cues can shape trust, the findings make a strong case for building genuine fact-checking into AI tools that handle high-stakes information like health advice.