Managing misinformation in healthcare: New risks and strategies

• In an online world, health misinformation is prevalent and can often be misleading

• New threats around health misinformation are emerging from the increased use of AI within healthcare systems

• Healthcare leaders must take action to protect patients and healthcare systems by treating AI tools like clinical devices, with the same level of safeguarding in place before adoption

Much has been discussed about the ways in which users are exposed to health misinformation on social media and the wider internet. It will be a long time before anyone forgets the stark consequences of vaccine misinformation during the COVID-19 pandemic, which led to countless unnecessary deaths.¹ But in 2026, with the increasing use of AI in healthcare, organizations are at risk from new, more complicated types of health misinformation. 

We sat down with José Morey M.D., Eisenhower Fellow and a health and technology consultant for NASA, Forbes, MIT, and the White House Office of Science and Technology, to discuss what examples of health misinformation look like today, with a particular focus on the implications of AI's increasing prominence in the clinical encounter.

A lack of credible information has been found to impact how patients see healthcare organizations.² “Studies show again and again that lower trust can have a direct effect on care, and perception of care, which leads to utilization issues,” says Dr. Morey.  He sees this as an area where healthcare organizations need to be proactive: “Misinformation spreads faster than the corrections, so it’s important that hospitals and clinicians view medical misinformation as a community health risk.” 

To overcome this risk, he suggests developing storytelling teams that are actively engaging on social platforms, sharing reliable, evidence-based health information, and building digital footprints. “Hospitals should encourage and give clinicians dedicated time to work on socials because it is part of their acute and chronic care strategies. This is important because data shows people trust individual doctors more than they do hospitals or institutions,” says Dr. Morey.

Navigating misinformation not only affects individual care, but also impacts healthcare systems operating under value-based care models, where adherence to treatment is fundamental. “Poor understanding of discharge instructions contributes directly to preventable readmissions and health system costs,” explains Dr. Morey, “With up to 40% of readmissions to organizations linked to poor adherence to instructions after discharge, there is a lot of scope to improve.”³ 

This can be achieved with better communication and patient education, with one study demonstrating that communication interventions at discharge were associated with lower readmission rates, higher treatment adherence, and increased patient satisfaction.³ This is particularly important for systems like the US, where consumer assessment scores influence hospital reimbursement under value-based purchasing programs, meaning improvements in patient understanding and communication have direct financial implications.⁴

Dr. Morey explains: “In value-based care, the return on investment (ROI) emerges when verified information increases adherence, which then improves outcomes and leads to higher consumer assessment scores, and therefore reimbursement.” He continues: “A useful way to think about ‘verified information delivery’ is to treat it as a clinical intervention that can be measured like a medication and treated like an operational KPI that could be tracked in electronic health records. Ultimately, the ROI is both better patient care and savings for the health system since avoidable readmissions and better adherence reduce chronic care costs.”

Along with ensuring reliable information is communicated to patients, healthcare organizations must take care that misinformation doesn’t end up in their own systems. Dr. Morey highlights that, “Parts of the misinformation problem now look a lot like cybersecurity.” He explains: “If medical misinformation enters patient portals, chatbots, clinical AI tools, search results, call-center scripts, or clinician-facing decision support, then health systems need technical safeguards. This could be source verification, content provenance, human review, monitoring, escalation pathways, and vendor controls to manage harms from AI systems.”

This is key when working with AI vendors to guard against “institutional misinformation”. As the World Health Organization (WHO) has warned, biased AI can deepen inequity if it is trained on unrepresentative data or deployed into unequal care systems.⁵ 

Dr. Morey believes executives should therefore treat vendor AI review less like software procurement and more like clinical quality assurance: “The core question to ask is not, 'does the model perform well on average?', but, 'who does it fail on, under what conditions, and what happens operationally when it does?’” He outlines four steps healthcare leaders can take when considering a new AI vendor: 

1. Demand a dataset provenance dossier from the vendor: That should include where the training, tuning, and test data came from; what time periods they cover; which hospitals, geographies, and care settings are represented; and how patients are distributed across race, ethnicity, sex, age, language, disability, payer type, and comorbidity burden. If the vendor cannot clearly describe representation gaps, label quality, missingness, and exclusion criteria, that is already a major warning sign. The FDA’s recent AI/ML medical device materials emphasize transparency and lifecycle risk management, including attention to bias and real-world performance, not just initial claims.⁶
2. Request subgroup performance reporting, not just a single Area Under the Curve or accuracy figure: Ask for sensitivity, specificity, calibration, false-negative rate, false-positive rate, and intervention yield by clinically meaningful subpopulations. In health systems, the most dangerous bias is often not dramatic error, but quiet underperformance in groups that already face delayed diagnosis, lower continuity, or weaker access. A model that looks “good overall” can still systematically miss Black patients, welfare patients, non-English speakers, rural patients, or people with fragmented records. The National Institute of Standards and Technology notes that AI bias is not only a data problem but can also come from human and institutional choices embedded in design and deployment.⁷  
3. Audit for label bias: This is important because many healthcare models are trained on proxies that reflect the care system rather than the patient’s true clinical need. For example, cost, prior utilization, referral patterns, charting intensity, or diagnostic codes may be influenced by access barriers and clinician behavior. That means the model can end up learning who gets attention, not who needs attention. The WHO’s guidance explicitly warns that bias in AI can depart from equal treatment and threaten inclusiveness and equity.⁵ 
4. Request a workflow harm analysis: A biased model becomes “institutional misinformation” when its output is granted organizational authority and starts shaping triage, outreach, prior authorization, staffing, or escalation pathways. The executive question should be, “If the model is wrong, who gets less care, slower care, or no outreach at all?” The Office of the National Coordinator for Health’s HTI-1 rule created transparency requirements for predictive decision support in certified health IT precisely because users need information about the model’s purpose, data basis, intended population, and intervention logic in order to judge trustworthiness.⁸

Once an organization has chosen a vendor and tools to work with, the focus becomes how these can be used safely across the system. When rolling out to staff, Dr. Morey again recommends that hospitals should treat AI tools like clinical devices, not just another IT software. He outlines practical guidance for implementation: 

• Create an AI governance board including patients, clinicians, informatics experts, legal teams, and ethicists.

• Require clinical validation before deployment of AI tools in care settings.

• Define acceptable use policies for staff using external AI tools (ChatGPT, Copilot, etc.).

• Review AI-generated clinical content:

  • Require source citations and human-in-the-loop review of these. This is important because LLMs, mathematically, are just really great probability guessers. This forces more reliable information, and it also helps keep the clinicians honest.

  • Flag answers lacking sources.

  • Display confidence scores for AI outputs.

  • Require secondary verification for high-risk clinical advice.

• Require AI training for all of your clinicians so they learn the limitations of the models, and so that they learn about AI drift, hallucinations, biases, and how to audit for these things. Both by setting up protocols and self-checks.

Misinformation in healthcare is not just inconvenient, it has real and tangible consequences. Patients, hospitals, and healthcare systems can suffer greatly as a result of inaccurate information. Health leaders must take action to mitigate against these risks by providing credible information, taking caution when choosing vendors, and providing clear governance and guidance on the use of AI tools.