Generative AI Shapes the Full Lifecycle of Health Misinformation, Review Finds

Generative artificial intelligence is not merely another channel through which false health claims travel, according to a new narrative integrative review published in Frontiers in Medicine. The review, authored by Xiaoli Nan of the University of Maryland, argues that large language models and related generative systems now operate at every stage of the health misinformation lifecycle, from the production of misleading content through its detection, correction, and downstream consequences for individuals and institutions. Drawing on research from communication science, public health, human-computer interaction, and artificial intelligence, the review proposes a Lifecycle-Based Framework that treats these four domains as interdependent and recursive rather than as a fixed linear sequence.

The review’s starting point is production. Generative AI systems, particularly large language models, can produce fluent, context-sensitive, expert-styled health content on demand and at scale, unconstrained by the human expertise, labor, and coordination that limit traditional content creation. Because these systems are optimized for probabilistic language prediction rather than factual verification, the same tools that expand access to health information also create conditions for the rapid generation of misleading, incomplete, or fabricated content. The review emphasizes that AI-generated misinformation is not limited to blatantly false statements; it can also take the form of incomplete explanations, fabricated details, decontextualized evidence, or persuasive narratives that appear medically credible despite lacking factual grounding.

A central mechanism is hallucination, defined as the generation of information not grounded in verifiable sources or established evidence. The review distinguishes intrinsic hallucinations, in which generated content contradicts the source material provided to the model, from extrinsic hallucinations, which introduce information that cannot be verified from available sources. Empirical evaluations cited in the review show that large language models can generate incorrect or misleading health information in response to common queries. In assessments of ChatGPT’s responses to vaccination myths, responses were often well-structured and aligned with public health messaging in tone, yet sometimes contained factual inaccuracies, oversimplified mechanisms, or misleading causal explanations. Analyses of COVID-19-related content found that ChatGPT blended accurate information with outdated guidance or unsupported claims, producing outputs that lay users found difficult to critically evaluate. Studies in medical contexts also document that models may fabricate references, invent clinical details, or construct internally coherent yet false explanations, and hallucinated outputs often retain a high degree of linguistic confidence, which may amplify their persuasive impact despite their inaccuracy.

Beyond isolated errors, the review identifies amplification of false narratives, adversarial vulnerability, and deepfakes as additional production mechanisms. Experimental and observational studies indicate that large language models can generate explanations mirroring well-known misinformation tropes, including vaccine skepticism, alternative cancer treatments, and conspiracy-based reasoning, frequently supplying extended rationales with causal structure, scientific-sounding language, and moral framing that increase the perceived plausibility and memorability of misinformation. At the system level, medical large language models have been shown to be susceptible to data-poisoning attacks, in which strategically manipulated training or fine-tuning data induce models to generate false or misleading diagnostic explanations and treatment recommendations, a dynamic the literature describes as misinformation pollution. At the interaction level, adversarial prompts introducing distorted premises or misleading contextual cues can steer medical models toward incorrect clinical content, which the models then elaborate on using fluent, clinically styled reasoning that lends misinformation an appearance of medical legitimacy. Notably, both attack types remained effective even against models fine-tuned for medical accuracy. Advances in synthetic media further allow highly realistic fabricated audio and video depicting clinicians or public health officials making statements that never occurred, and deepfake generation can be coupled with model-produced scripts to create coordinated multimodal misinformation.

The review then turns to detection, where generative AI plays a double role. Large language models can assist automated fact-checking by generating provisional judgments, offering rationales, and pointing to evidence. In one health-related comparison, ChatGPT-4 and Google Bard were assessed against expert judgments on sleep-health misconceptions, illustrating that such models can often identify myths and provide corrective explanations while still producing notable errors and inconsistencies. Researchers are also developing structured workflows in which models first identify specific claims, retrieve relevant evidence, evaluate whether the evidence supports the claim, and report how certain the judgment is. Generative AI can additionally support pattern recognition and classification, processing volumes of content that human reviewers cannot feasibly track and identifying recurring talking points, reused templates, and narrative variants, including classifying narrative types such as conspiracy frames, pseudoscientific rationales, and testimonial narratives.

However, the review catalogs persistent limitations of AI-based detection. Auditing and evaluation studies repeatedly find error-proneness in both directions: false positives that label true claims as false, and false negatives that fail to flag misinformation. Chatbot accuracy varies by topic, claim type, and source attribution, implying unstable operating characteristics for real-world deployment. Detection systems can inherit bias from training data, benchmarks, and evaluation criteria, performing unevenly across high-resource and low-resource languages, which may reduce detection quality for communities already facing barriers to reliable information. Many health misinformation claims are partially true, decontextualized, or framed with misleading causal inference, so surface-level factual overlap can mask misleading conclusions, and fluent, medically styled explanations containing subtle errors or outdated guidance do not resemble obvious hoaxes. Detection also faces adversarial pressure, as producers adapt content to evade flags, generating endless paraphrases and narrative variants with minimal human labor. Transparency poses a further challenge: stakeholders may not know why a claim was flagged, what evidence was used, or how uncertainty was handled. The review concludes that AI is best used as a support tool rather than a standalone judge of accuracy, with human expertise needed for contextual interpretation and final decisions.

At the correction stage, the review finds that generative AI can produce scalable, tailored, and interactive corrective messages. Early empirical studies show that AI-generated responses to misinformation can reduce endorsement of false beliefs under some conditions, particularly when responses are coherent, evidence-based, and dialogic rather than purely declarative. Models can adapt explanations to different literacy levels, cultural frames, or emotional concerns, addressing specific misconceptions rather than offering generic myth-versus-fact formats, and emerging intervention studies show that interactive, AI-mediated dialogue can improve vaccine confidence and uptake intentions compared with static information formats. Experimental research on AI-mediated dialogues also indicates that structured conversations with chatbots can reduce conspiracy beliefs and misinformation endorsement by encouraging reflection and evidence-based reasoning. Generative AI is well suited to support inoculation approaches, which build psychological resistance by exposing individuals to weakened forms of misleading arguments and refutations, and emerging research demonstrates that generative systems can help produce scalable prebunking content, although direct health-domain trials of AI-generated inoculation messages remain limited.

These corrective capabilities come with significant sociotechnical risks. The review highlights trust displacement, whereby users come to rely more on AI-generated explanations than on human health professionals or public health agencies, partly because people often apply a machine heuristic, perceiving computational systems as objective and less biased than human sources. AI-driven correction also raises questions of epistemic authority, since normative judgments about acceptable uncertainty and evidentiary thresholds become embedded in technical systems, potentially presenting evolving or debated health knowledge as fixed and settled. A further risk is reinforcing misinformation: repeating a false claim even for refutation can increase familiarity and perceived truth, a pattern associated with the continued influence effect, and automated corrections may inadvertently restate or elaborate on the narratives they seek to counter. Governance and accountability remain unresolved, as responsibility for misleading AI-generated corrections may be distributed across developers, platform operators, deployers, and end users, while proprietary models limit external auditing.

Finally, the review examines impact. Empirical research indicates that AI-generated health misinformation can degrade knowledge accuracy, with users often accepting incomplete or misleading medical explanations as evidence-based and struggling to distinguish valid guidance from erroneous or speculative content. Exposure can distort risk perceptions and health beliefs, and misinformation exposure has been linked to decreased adherence to public health guidelines, greater openness to unproven treatments, and increased vaccine hesitancy. Three persuasive mechanisms stand out: linguistic fluency, which functions as a cue to credibility even when accuracy is low; narrative-style, mechanistic explanations that make inaccurate claims easier to encode and more resistant to correction; and dynamic personalization that adapts content to users’ concerns and beliefs. Notably, studies show that labeling content as AI-generated does not reliably reduce perceived accuracy or sharing intentions. At the societal level, the review links AI-generated misinformation to declining trust in health authorities, heightened epistemic uncertainty amid information overload, and potential amplification of health inequities, since vulnerable communities may rely more heavily on conversational AI systems and culturally tailored misinformation can extend misleading narratives into underrepresented communities. The framework ultimately conceptualizes the lifecycle as recursive and adaptive, in which eroded trust, institutional strain, and detection failures continually reshape how subsequent misinformation is produced, received, and countered, requiring coordinated, multilevel strategies that preserve human oversight, institutional accountability, and public trust.

Subject of Research: The role of generative AI across the health misinformation lifecycle

Article Title: Generative AI and health misinformation: a lifecycle-based communication framework

Article References: Nan, X. (2026). Generative AI and health misinformation: a lifecycle-based communication framework. Frontiers in Medicine, 13, Article 1801842. https://doi.org/10.3389/fmed.2026.1801842

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1801842

Keywords: generative AI, health misinformation, large language models, hallucination, deepfakes, fact-checking, inoculation, risk perception, vaccine hesitancy, health communication, trust displacement, public health