Health Chatbots Struggle When Questions Push Toward Misinformation
Five widely used generative AI chatbots were put through a deliberately adversarial health audit, and the results were blunt: nearly half of the answers raised concern. In a field where people increasingly turn to search-style prompts for everyday medical guidance, the study tested whether these systems could stay accurate when asked about topics known to attract misinformation, including cancer, vaccines, stem cells, nutrition and athletic performance.
The answer was often no. Across 250 chatbot responses, 49.6% were judged problematic, split between 30% somewhat problematic and 19.6% highly problematic. The tone of the outputs created a particular risk: the responses were generally delivered with confidence and certainty even when the content was incomplete, misleading or unsupported. For a non-expert user, that combination can be more dangerous than a cautious but vague answer, because fluency can disguise error.
Adversarial Prompts Exposed Weaknesses Across Five Platforms
The audit included Gemini from Google, DeepSeek from High-Flyer, Meta AI from Meta, ChatGPT from OpenAI and Grok from xAI. In February 2025, each chatbot received 10 questions from each of the five topic areas, with prompts designed to strain the models toward misinformation or contraindicated advice. The design matters because it mirrors how real-world users often interact with these tools: not as supervised decision aids, but as conversational engines that may be asked to interpret symptoms, evaluate claims or validate risky advice.
Two experts in each category rated the responses as non-problematic, somewhat problematic or highly problematic using predefined criteria. That approach anchored the review in objective coding rather than vague impression. The final pattern was unsettlingly broad: no platform escaped error, and the overall quality did not differ significantly among chatbots overall, with a p value of 0.566. One model stood out in a narrower way. Grok generated significantly more highly problematic responses than expected under a random distribution, with a z-score of +2.07 and p=0.038.
One in five answers crossed into the high-risk range.
Vaccines and Cancer Performed Better Than Stem Cells, Nutrition and Training Advice
Performance varied by subject area, which is clinically important because different health domains place different demands on factual precision. The strongest outputs came in vaccines and cancer, where mean z-scores were -2.57 and -2.12 respectively. Those lower scores suggest better-than-average performance relative to the study’s coding framework. Even there, however, the systems were not uniformly reliable enough to be considered safe without oversight.
The weakest results appeared in stem cells, athletic performance and nutrition, with mean z-scores of +1.25, +3.74 and +4.35. Those domains are especially vulnerable to bad advice because they mix active scientific debate with commercial claims, social media exaggeration and a heavy dose of anecdote. In practice, that creates a fertile setting for hallucinated certainty: a chatbot can sound authoritative while drawing a false line between evidence and speculation.
The study design also helps explain why those categories were so exposed. Questions about supplements, diet, training load, regenerative medicine and unproven interventions often invite oversimplified or absolutist answers, while the underlying evidence base is more conditional and context-dependent. In a conversational interface, nuance is easy to lose and easy to replace with compressed, overconfident guidance.
Reference Quality Fell Short Even When Answers Looked Convincing
Citations were another major weakness. The median completeness score for references was 40%, with an interquartile range of 20% to 67%. In other words, even when chatbots attached citations, the references were often partial, thin or poorly matched to the claims being made. Hallucinated and fabricated citations prevented any chatbot from producing a fully accurate reference list.
That failure carries direct clinical implications. Users who see citations may assume the answer has been checked against reliable literature, when in fact the references may be incomplete, inaccurate or invented. For patients searching for guidance, and even for clinicians using these tools as a quick background check, false referencing can create a dangerous illusion of validation. The problem is not just factual error; it is error wrapped in the appearance of scholarship.
Refusal Rates Were Rare, Which Left Little Room for Safety Barriers
Across the entire audit, chatbot refusals were exceptionally uncommon. Out of 250 total questions, only two answers were refused, both by Meta AI, representing 0.8% of prompts. That low refusal rate suggests the systems were generally willing to answer even when the safest response might have been to decline, defer, or advise consultation with a qualified professional.
In a health context, refusal is not necessarily a flaw. It can function as a safety valve when a prompt seeks diagnosis, treatment substitution, or advice that would be unsafe without clinical context. The absence of that brake leaves the burden on the model to distinguish between what it can answer responsibly and what it should not attempt. This audit suggests that distinction was rarely enforced.
Readability Sat Below the Level Most Users Can Comfortably Parse
Every chatbot response received a Flesch Reading Ease score in the “Difficult” range, defined in the audit as 30 to 50 and roughly equivalent to college sophomore-to-senior reading level. That finding adds another layer of concern. A response can be wrong, but it can also be wrong and difficult to parse, which increases the chance that users will misread qualifiers, miss caveats or accept fragments of advice out of context.
Readability matters most in misinformation-prone topics because people often seek quick clarification, not a dense educational paragraph. When the language is hard to process, the user may cherry-pick the most assertive line and ignore the rest. In that setting, the combination of confident phrasing, incomplete citations and advanced reading level becomes a practical risk rather than a stylistic issue.
Public Education and Oversight Now Sit at the Center of Deployment
The audit’s conclusion is direct: these chatbots performed poorly in health and medical fields that are already saturated with misinformation. Continued deployment without public education and oversight risks amplifying falsehoods rather than reducing them. The findings also expose a gap between consumer-facing convenience and clinical reliability. Generative systems may feel conversationally helpful, but in health domains they can produce polished answers that fail basic tests of accuracy, citation integrity and readability.
For clinicians, the study reinforces a familiar warning. Patients will increasingly arrive with chatbot-generated content in hand, and some of it will be persuasive enough to complicate counseling. For developers, the results point to a deeper product problem: answer fluency is not a substitute for verified evidence, and citation formatting is not the same as citation accuracy. Until safety mechanisms improve, health-related chatbot output should be treated as untrusted text, especially when the question touches cancer, vaccination, stem cell claims, nutrition or performance advice.