Fighting Fake Health Claims With AI: Why Better Technology Still Needs Human Trust

Artificial intelligence is being discussed as a defence against health misinformation online, yet its real value may lie less in automatically deciding what is true or false and more in helping health professionals understand what people are worried about, where misleading narratives are emerging, and how credible information should reach different communities.

A new mini review titled 'AI-enabled health communication to address misinformation on social media', written by Yuqi Hu of the University of California, San Diego and published in Frontiers in Communication, examines how AI could become part of a wider health communication response rather than simply serving as an automated misinformation detector. The review covers research published between 2000 and 2026 and brings together evidence from public health, communication and computer science.

Health misinformation is already found across subjects ranging from infectious diseases and vaccines to cancer, diet, reproductive health, tobacco and e-cigarettes, while generative AI has created another route for producing convincing but inaccurate material. Exposure to misleading health claims can weaken health literacy, encourage unsafe choices and interfere with prevention or treatment, making the problem far more serious than an occasional incorrect social media post.

AI Can Listen to Online Conversations at a Scale Humans Cannot

The strongest potential use identified in the review is AI-enabled social listening, where technology helps health communicators understand not only which false claims are circulating but also the questions, emotions, uncertainties and information gaps surrounding them.

Traditional monitoring based on keyword searches, surveys, focus groups, manual media reviews and expert analysis can struggle with the enormous volume and speed of social media conversations. Machine learning and natural language processing systems can scan much larger amounts of content, classify potentially misleading posts or claims and identify patterns that deserve closer human attention.

The opportunity goes deeper than detection: Topic modelling, semantic clustering and LLM-based summarisation can reveal emerging narratives and recurring concerns, while sentiment and emotion analysis can indicate whether a conversation is being driven by fear, anger, confusion, distrust or uncertainty. Network analysis and bot-detection techniques can help identify influential accounts, coordinated amplification and communities through which misinformation travels.

That context matters: A vaccine rumour rooted in fear of side effects may need a very different response from one driven by distrust of institutions or confusion about scientific evidence.

AI does not necessarily see every community equally well: Models can misunderstand sarcasm, memes, humour, coded language, multilingual conversations and local references. Training data dominated by particular languages or populations can cause systems to overlook misinformation affecting immigrant, refugee, low-income and low-resource-language communities. Monitoring public conversations also creates legitimate privacy and surveillance concerns, making human interpretation and community involvement essential.

From Corrections to Chatbots, AI Could Help Shape the Response

Once a harmful narrative is identified, AI could support the creation and delivery of corrective messages, fact-check explanations, plain-language summaries, multilingual information, chatbot responses and social media content adapted to different platforms.

Large language models can help summarise evidence, simplify technical medical information and draft several versions of a message. Research has already explored AI-generated pro-vaccination messages, responses to vaccine myths and communication intended to reduce misconceptions around mental health.

Another possibility is prebunking, which tries to prepare people before misinformation reaches them. Instead of correcting every false claim after it spreads, communicators can teach audiences to recognise tactics such as cherry-picking evidence, emotional manipulation, fake expertise, conspiracy framing and false balance. AI could identify recurring manipulation techniques and create topic-specific exercises or interactive media-literacy material, although the review stresses that much of this work has yet to demonstrate effectiveness in real-world social media environments.

Chatbots offer a more conversational approach because users can ask questions when they are uncertain rather than simply receiving a static correction. Evidence remains modest: one 2026 study found that short LLM chatbot conversations increased parents' immediate intentions to vaccinate against HPV compared with receiving no message, yet the effect did not last and the chatbot did not outperform standard public health materials.

Personalisation adds another possibility, allowing communication to be adjusted for language, health literacy, emotional concerns, previous misinformation exposure, platform culture and local context. It also creates risks when systems make assumptions about people from incomplete online signals, potentially reinforcing filter bubbles or existing information inequalities.

Technical Accuracy Does Not Mean Better Health Communication

A major weakness in the current evidence is how AI systems are evaluated. Researchers frequently report accuracy, precision, recall, F1 scores, processing speed or the quality of generated text, yet a model that detects misinformation accurately has not necessarily helped a person understand a health issue or make a safer decision.

ChatGPT assistance has shown inconsistent effects on people's ability to distinguish reliable health information from misinformation. Some AI systems with strong technical results have not yet been tested for their effect on actual audiences. Research on social media vaccine interventions also shows stronger evidence for improving knowledge, confidence and attitudes than for producing actual increases in vaccination.

The review argues that future evaluations need to measure comprehension, credibility, belief accuracy, trust, risk perception, intentions, behaviour and sharing. Researchers should also examine whether an intervention narrows or widens gaps associated with language, literacy, socioeconomic circumstances, platform access and institutional trust, while watching for unintended consequences such as greater exposure to false claims, increased scepticism or excessive dependence on automated advice.

AI tools should also be compared with existing alternatives such as clinician communication, community-led messaging, peer correction, established public health materials and non-AI media-literacy interventions, rather than assuming that an AI-generated message is better simply because it can be produced quickly or personalised at scale.

The Same AI Fighting Misinformation Can Also Produce It

The review highlights an uncomfortable contradiction: generative AI can support health communicators while making health misinformation cheaper, faster and easier to produce. AI can generate fluent, persuasive and audience-specific misinformation in many languages and formats. Synthetic text, images, voices, videos and deepfake-style material could make misleading health narratives more personalised and harder to identify, while the same false story can be repeatedly rewritten for different audiences and platforms.

This makes governance central to responsible use. Social listening can cross an uncomfortable line between understanding public concerns and surveillance, especially when people do not expect their posts to be classified or used to create targeted interventions. Unequal model performance across languages, dialects and communities creates another risk of leaving underserved populations with poorer information support.

The review calls for privacy protection, data minimisation, transparency, documentation of data sources and model limitations, clear review and error-correction procedures, human oversight, community consultation, independent auditing and ways for the public to provide feedback.

The message is that AI should support health communicators, clinicians, public health professionals and trusted community messengers rather than replace them. Future research also needs to move beyond the English-language, COVID-19 and vaccine settings that dominate current evidence, while recognising that different social platforms, communities and health topics require different communication strategies.

AI may give public health teams faster ways to spot emerging misinformation and respond at scale, but trust cannot be automated so easily. The most useful systems are likely to be those where machines handle scale, analysis and drafting while people provide the judgment, cultural understanding, scientific responsibility and human connection that credible health communication still depends on.