Last updated Nov 8th, 2026
About 1 in 6 US adults said they used an AI chatbot at least once a month to find health information or advice in a KFF poll of 2,428 adults on AI and health information run in June 2024, and more than half of adults who use AI were not confident that chatbots give accurate health information.
Governments can now use AI to tailor health guidance, answer citizens' questions, catch emerging misinformation and produce plain-language translations in minutes. Each of those gains can also damage public trust or health outcomes when an AI-assisted message goes out untested and without anyone accountable for it. At The Decision Lab, we design and test public health communication with governments and health bodies, including the risk messaging for Mila's COVI contact-tracing app, which we took into working sessions with Health Canada.
Across all four shifts, our recommendation is the same. Use AI to generate and adapt candidate messages, test those candidates with the people the agency serves, measure the behavior each message is meant to change, and give every message a named human owner. We call that last requirement the named-owner standard: one person is accountable for the message, its sources, its escalation rules and its correction process. The examples in this guide come from Canada, the United States, France and Italy, and the lessons apply to US health agencies at every level of government, which face the same pressure to communicate quickly and equitably.
The four shifts at a glance
- Personalization. AI lets an agency tailor a health message to a citizen's language and circumstances, including the barriers most likely to stop them acting, and choose which version each person receives and when.
- Conversation. Public-health chatbots answer citizens' own questions in place of one page of guidance written for everyone.
- Misinformation response. AI can surface emerging health rumors sooner and, in research settings, reduce belief in false claims through a structured, evidence-based conversation, while also producing confident errors of its own.
- Plain language and translation. Agencies can produce plain-language and multilingual versions of AI-assisted health messaging in minutes, so reaching people who are usually left out of campaigns costs less.
1. AI personalization for public-health messages
What changes with AI
Public health campaigns have always segmented their audiences, but most segments were broad, such as an age band or a region, because writing and testing many versions of one message was expensive. AI makes each extra version cheap to draft, and machine learning can find groups that standard demographics miss.
What the evidence says
We saw this when we helped the Canadian Cancer Society design a text-message quit-smoking service with machine learning. We surveyed 122 smokers and ex-smokers and analyzed the results with decision trees and cluster analysis. The strongest dividing line was identity: whether smoking was part of who someone was separated smokers before risk tolerance or social support did. Four groups emerged, and each received its own conversation flow, with messages timed to the moments of the day each person had said were hardest.
What can go wrong
Personalization by a government carries obligations that a company or a charity can sometimes sidestep. Citizens can reasonably ask which of their data a public body used to decide what they were told, and a message that feels attentive to one person can feel like surveillance to another. Tailoring can also widen health gaps, because when a model learns mostly from people who already use public health services, the groups with the thinnest data are the ones most likely to get the generic message.
What government teams should do before launch
Decide which data may be used for targeting and publish that decision in plain language. Check who ends up receiving the generic version, and test the tailored messages with those groups specifically. Our work with Mila, the AI research institute, on how the COVI contact-tracing app communicated risk treated privacy as a design requirement from the start. We ran weekly surveys of Canadians on their trust in tracing apps and their privacy concerns, then designed English and French messages that changed with a person's circumstances, such as their occupation and health history, and replaced blunt exposure alerts with graded risk guidance. The project informed the move away from binary exposure alerts toward graded risk levels, a change Apple and Google later made to their exposure notification system.
2. Public-health chatbots and human escalation
What changes with AI
Most government health guidance has been one-way: a web page or an ad that says the same thing to everyone. Conversational AI lets a citizen ask about their own situation. The World Health Organization's prototype digital health promoter, S.A.R.A.H., launched in 2024, answers questions in 8 languages on topics such as quitting tobacco and managing stress.
What the evidence says
The evidence that chatbot conversations change health attitudes is encouraging but short-lived. In a randomized experiment on a COVID-19 vaccine chatbot published in 2021 in the Journal of Experimental Psychology: Applied, Sacha Altay and colleagues gave French participants a chatbot that answered common questions about the vaccines. Among the 338 people who used it, the number holding positive views of the vaccines rose by 37%, from 145 before the conversation to 199 after, while views barely moved among 305 people who read a short paragraph about the vaccines instead. One to two weeks later, the two groups no longer differed. We read that pattern as a design brief: a conversation can move someone in the moment, so it should end with a concrete next step, such as a link to book the appointment, while the person is still persuaded.
What can go wrong
The World Health Organization's 2024 guidance on large multi-modal models in health warns of false or incomplete statements that could harm people making health decisions, and of automation bias, where people overlook a model's errors or hand it choices it should not make. A citizen in distress who meets a dead end in a chatbot may also give up on getting help altogether.
What government teams should do before launch
Define escalation triggers, such as signs of distress, a question about a specific symptom or medication, repeated confusion, or a direct request for a person, and make the handoff to a human fast. Under the named-owner standard, those escalation rules belong to the same person who owns the chatbot's answers. Speed of handoff was central to our redesign work for the Wellness Together Canada mental health platform. User research and behavioral design informed a national mental health service, funded by the Government of Canada, that connected users to a counsellor after an average wait of 12 seconds and served more than 1.2 million people in its first year. Almost a quarter of its users said they would not have looked for help anywhere else if the platform had not existed.
3. Using AI to detect and respond to health misinformation
What changes with AI
AI shortens the time between a health rumor appearing and an agency noticing it. A guide to AI for local health department communicators from the National Association of County and City Health Officials (NACCHO), published in January 2026, suggests using AI to scan social media and community chat boards for local health concerns, and to stress-test a draft by asking what questions critics would raise.
What the evidence says
Research suggests AI can also reduce belief in false claims directly. In a 2024 Science study on dialogues with AI, Thomas Costello and colleagues had 2,190 Americans discuss a conspiracy theory they believed with GPT-4. Their belief in the theory, rated on a scale of 0 to 100 before and after, fell by about a fifth on average, and the drop was still there two months later. That study tested a structured conversation built to present evidence, and a general-purpose chatbot has not been shown to have the same effect.
What can go wrong
The same tools create a new source of error that citizens struggle to judge. In the KFF poll, more than half of all adults were not confident they could tell what is true and what is false in the health information AI chatbots give them. A government chatbot that answers wrongly does more damage than an anonymous one, because it borrows the authority of the agency.
What government teams should do before launch
Restrict any citizen-facing tool to vetted sources, and set out how its errors will be corrected publicly, the same way the agency would correct a press release. The NACCHO guide also tells communicators never to enter protected health information into AI tools.
4. AI for plain-language and multilingual public-health communication
What changes with AI
According to the Centers for Disease Control and Prevention's considerations for generative AI in public health, local health departments increasingly use generative AI to draft plain-language advisories and multilingual outreach messages for topics such as respiratory season and heat events.
What the evidence says
So far the evidence describes adoption more than effect. The same CDC page reports, from NACCHO's 2024 profile of local health departments, that communications drafting is the most common generative AI use among departments adopting AI. We have found far less published evidence on whether AI-simplified or AI-translated guidance improves what citizens understand or do.
What can go wrong
The WHO guidance notes that the best-performing models are not equally accessible or affordable, and a fluent translation can still miss the way a community talks about illness. Faster drafting improves equity only if the versions reach the people who need them and make sense to them.
What government teams should do before launch
Review translated and simplified versions with members of the communities they are written for. We built this principle into our work with Health Canada on more equitable, human-centered health research, where we turned the evidence on engaging people with lived experience into guiding principles and more than 30 practical recommendations.
AI disclosure and public trust in government health communication
Disclosure is a safeguard that runs across all four shifts. The CDC's generative AI considerations recommend disclosing generative AI use in public-facing products and report that public disclosure helped the agency build trust with partners and the public. Disclosure also has costs that agencies should plan for. In a US survey experiment by Benjamin Toff and Felix Simon, published in the International Journal of Press/Politics, labeling news as AI-generated lowered how trustworthy readers found it, even though they did not rate the articles as less accurate or less fair. Listing the sources the AI used largely offset that drop.
Health messages differ from news articles, and we suspect the effect will be smaller for an agency people already rely on. Until there is direct evidence, we recommend a specific disclosure that covers five points:
- What the AI did, such as drafting or translating the message.
- Which sources it drew on.
- Who reviewed the output, meaning the message's named owner.
- How a citizen can reach a person if they need human help.
- How errors will be corrected.
How to test AI-generated health messages before launch
AI multiplies the number of messages an agency can produce, but it does not tell the agency which one will change behavior. In a megastudy of vaccination text reminders led by Katherine Milkman and published in PNAS in 2022, 689,693 Walmart pharmacy patients received one of 22 text reminders or no message. On average the reminders raised vaccination rates at the pharmacy by 2.0 percentage points over the no-message group across three months, and the best performer, two texts saying a vaccine was waiting for the patient, was a message neither experts nor lay people predicted would win.
AI can help pick the candidates worth testing. Man-pui Sally Chan, Dolores Albarracín and colleagues built a system that pulled HIV prevention and testing messages from US social media and selected the most actionable ones for campaigns in 42 counties, reported in PNAS Nexus in 2025. Men who have sex with men rated the AI-selected messages as more actionable and personally relevant than control messages, and public health agencies and community organizations were substantially more likely to post prevention messages when they used the system. Messages a person had vetted performed better than unvetted ones.
We take the same approach with governments. When the City of Rome asked us to increase participation in its participatory budgeting campaign, we developed a range of message framings and tested them with nearly 10,000 Romans, including how different communities in the city responded, so the city could reach groups whose voices are often underrepresented.
A testing sequence for AI-assisted health messaging looks like this:
- Define the behavior. Name the action you want, such as booking a screening, and how you will measure it, before generating any messages.
- Generate and filter. Use AI to draft or select many candidate messages, then have communicators and subject-matter experts remove anything inaccurate or off-tone.
- Test with the audience. Compare the remaining messages in an online experiment with people who resemble the target population, including the groups with the least data.
- Run a field test where possible. A randomized trial against a control group gives the strongest evidence, though a controlled experiment is not always practical in an emergency.
- Monitor after launch. Track the behavior itself and retire messages that stop working, and have the named owner review chatbot conversation logs on a fixed schedule.
How The Decision Lab helps governments with AI and health communication
We work with national and local governments and public-health bodies on the behavioral side of responsible AI for government health communication. We diagnose why citizens do not act on health guidance, then design the messages and conversational flows that address those barriers and test them against a control group before they scale. Our published work in this area ranges from risk communication for COVI to national mental health and smoking cessation services. If your agency needs help choosing an AI vendor or building the model itself, a technology partner is the better first call, and we are happy to work alongside one.
Frequently asked questions
How are governments using generative AI in public health communication?
Most agencies start by drafting plain-language advisories and translations, which staff review before release. Some are also testing public-health chatbots that answer citizens' questions and models that decide which message each group of citizens receives, alongside tools that monitor social media for health rumors. The strongest evidence of behavior change comes from messages tested before launch.
Are public-health chatbots safe for governments to use?
They can be, if a chatbot draws only on vetted sources and a named person owns its answers and escalation rules, with a fast route to a human. The WHO warns that these models can produce false or incomplete statements and that users may trust them too readily, so handoff triggers should be defined before launch.
Should government health messages disclose that AI was used?
The CDC recommends disclosure for public-facing products. Research on news suggests a bare "AI-generated" label can lower trust, while listing the sources the AI used largely offsets that effect. A useful disclosure covers five points: what the AI did, its sources, who reviewed the output, how to reach a person and how errors will be corrected.
Can AI personalization make public health communication less equitable?
It can. Models trained mostly on people who already engage with health services tend to serve those people best, and translated or simplified text can still miss how a community talks about illness. Testing messages with underserved groups and reviewing versions with community members before launch reduces that risk.
How do you know whether AI-assisted health messaging works?
Measure the behavior it was meant to change, such as vaccination or screening rates, and compare it with a control group that did not receive the message. In a 22-message vaccination megastudy, neither experts nor lay people predicted the best performer, so how good a message looks in review says little about how many people it will move.
