I have lost count. Over the past year, in conference sessions, webinars, working group meetings, and posts from people I respect, I have heard the same sentence delivered with real conviction.

We need to educate our staff and our patients to use artificial intelligence properly.

Heads nod. The agenda moves on. Two questions go unasked, and I have spent a year waiting for someone to raise either one.

Who does the educating? How do they do it?

The Concern Is Real

Shaili Gupta, an associate professor of medicine at the Yale School of Medicine, described in February what happens when a patient arrives at an appointment having already consulted a chatbot. Some patients come better informed, asking sharper follow-up questions than before. Others arrive convinced the chatbot has already found the answer. Gupta characterized the resulting visit as an exercise in trying to “educate, redirect, and cancel out the misinformation.”

Adam Rodman, a general internist at Beth Israel Deaconess Medical Center and an assistant professor at Harvard Medical School, was blunter in June when the subject turned to patients asking chatbots about treatment decisions. “I tell my patients, just don’t do it.”

Neither physician opposes the tools. Both run artificial intelligence programs at their own institutions. Two clinicians who understand the technology well are describing, from inside practice, a problem they encounter weekly.

The concern is real. Gupta recounted a patient with chest pain who became convinced of a cardiac cause after consulting a chatbot, and who held the conviction through testing that pointed elsewhere. A single patient is an anecdote. In February the Canadian Medical Association published its Health and Media Tracking Survey, conducted by Abacus Data among five thousand Canadians, which found that people who followed health advice from an AI platform were five times more likely to report harm than those who did not. Developers have trained these systems to surface the urgent explanation first, and to shape answers around what a user appears to want. Anyone who has typed a symptom into a chatbot at midnight understands the pull of a confident reply.

From the evidence follows the remedy, sincerely offered and widely repeated. Educate patients before they use these tools. The sentence has the shape of a plan and the substance of a slogan.

Who?

I have read a good deal of the advice since, and I have yet to find the teacher named in any of it. Consider who the candidates might be.

Gupta is the obvious first answer, and she is already doing the work. The difficulty is where the work has to happen.

Robert Shpiner, a clinical professor of pulmonary and critical care at the David Geffen School of Medicine at UCLA, named the problem in a commentary published in Annals of Internal Medicine this week. Patients choose chatbots because the tools answer immediately, Shpiner argued, while the ordinary routes into care do not. A tracking poll by KFF, the American health policy research organization, found in early March that roughly a third of American adults had consulted an AI chatbot about health within the past year. Asking clinicians to teach patients about artificial intelligence asks them to do it inside the appointment those patients could not get.

In Canada the difficulty sharpens. The same CMA survey found that fifty-seven percent of Canadians turned to the internet only after failing to reach a family doctor or another health professional. Margot Burnell, the CMA president and a medical oncologist in New Brunswick, observed that Canadians struggling to access care have little choice about where to turn.

A Liaison Strategies poll in May measured the scale. Forty-six percent of Canadians had asked a chatbot for medical advice within the past year. The people most likely to consult a chatbot are the people least likely to have a physician. Assigning the teaching to clinicians assigns it to someone the student cannot reach.

Health systems are the second candidate. Jeffrey Flaks, president and chief executive of Hartford HealthCare, explained in May why his organization built a patient-facing chatbot rather than warning patients away from the ones they already use. Epic has deployed a similar assistant inside its patient portal. Both decisions are defensible. Neither amounts to education. A health system that builds its own chatbot has made a product decision, and the literacy produced runs one tool deep.

The third candidate goes unmentioned everywhere. OpenAI reported in January that more than forty million people consult ChatGPT for health information every day, and that seven in ten of those conversations happen outside ordinary clinical hours. The companies hold the only classroom that reaches patients at the moment a question occurs to them. Nobody proposes handing them the curriculum, for the obvious reason that a vendor writing its own safety manual invites suspicion.

Every proposal assumes a teacher standing just outside the frame. I have been a patient for some years now, through appointments, portal messages, and lab results opened late at night, and no one has offered to teach me anything about artificial intelligence. I would not know where to enrol.

How?

The second question is harder still. Even with a teacher, what does the teaching actually look like?

We have tried this before, and I was one of the people who tried it.

In the early 1990s I helped build one of Canada’s first commercial internet service providers. Customers arrived with a modem, an account, and no map. The web had no search engine worth the name, no ranking, and no way to distinguish a physician’s page from a pamphlet written by anyone with a text editor. We took the problem seriously enough to hire a librarian, a professional holding a Master of Library and Information Science, whose entire discipline is the organization and retrieval of information. She built directories. We held classes.

The approach did not scale. The shortfall belonged to the method rather than to the librarian, because the web grew faster than any human curator could catalogue it. People learned the internet from each other. A nephew showed an aunt. A colleague forwarded a link. Larry Page and Sergey Brin later solved much of the credibility problem structurally, through ranking, rather than pedagogically, through instruction.

Universities reached for the same solution, and hired the same profession to build it. Sarah Blakeslee, a librarian at California State University Chico, published the CRAAP test in LOEX Quarterly in 2004, a checklist covering currency, reliability, authority, accuracy, and purpose. SCONUL, a British library organization, issued a position paper in 1999 that became the Seven Pillars of Information Literacy.

Sam Wineburg and Sara McGrew of Stanford measured whether the checklists worked. In a 2017 working paper, they compared undergraduates, university faculty, and professional fact-checkers evaluating the credibility of web content. Faculty performed barely better than the undergraduates. Wineburg and McGrew traced the shortfall to the method, because faculty examined each website in depth exactly as the checklist instructs.

Consider who those faculty were. People with doctorates, trained to weigh evidence for a living, some of them teaching the checklist to their own students, scoring level with first-year undergraduates. The fact-checkers did better because they had picked up a different habit at work, opening new tabs and investigating the publisher rather than studying the page in front of them, a practice Wineburg and McGrew named lateral reading.

Health care now proposes the same remedy, aimed at patients rather than students, against a technology moving considerably faster than the web ever moved. The remedy failed when librarians delivered it. No one has said who delivers it this time.

I am not arguing against helping patients. I am arguing that a paragraph in a policy paper is not a plan, and that education positioned as a precondition for use becomes something other than education. Banks made the identical argument when customers wanted to move money online. Stockbrokers made it when investors wanted to place their own trades. Every argument was sincere. Every argument lost.

What to Hand Patients Instead

Three suggestions follow, offered in the spirit of the advice I am criticizing. Each one names who acts and what they do.

First, give patients their data in portable form. Much of what gets diagnosed as a literacy gap is an access gap, because a chatbot knows nothing about the person typing.

Bill S-5, the Connected Care for Canadians Act, would require vendors to make health information technology interoperable, defined as letting a user access their electronic health information and exchange it with other health information technologies. Senator Pierre Moreau sponsored the bill on February 4. The Senate passed it on May 26, the House gave it first reading on May 28, and second reading has not been reached. An earlier version died on prorogation in 2025.

OpenAI launched a service on January 7 that connects patient portals directly to a chatbot, four weeks before Moreau introduced the bill. Parliament is debating whether Canadians may hand over their records.

Second, put Rodman’s stoplight in front of patients at the point of care. Rodman tells patients not to ask a chatbot about treatment decisions, and Rodman also wrote the clearest guidance any patient has been handed. The physician who takes the risk most seriously is exactly the one you want writing the instructions. Green means go. General questions, diet, appointment preparation, and making sense of your own lab results and visit notes. Yellow means proceed with care. New symptoms belong here, and Rodman suggests asking the chatbot to interview you the way a physician would rather than reciting a list, so the tool probes for what you do not have. Red means stop. Treatment decisions belong with a clinician who can examine you.

A national organization should settle the wording and the format once, rather than leaving every hospital and health authority to invent its own. Healthcare Excellence Canada exists to turn a proven approach into something others can adopt locally. Hospitals, clinics, and health authorities could then add the guidance to discharge instructions, lab result notifications, waiting room material, and their own websites. No class convenes. No teacher is hired. The guidance reaches the patient holding the document at the moment the question arrives.

Third, ask patients what they are already using. Add a question to the intake form and to the history. Have you looked into this concern with an AI tool, and what did it tell you? The question costs one line and surfaces the belief a clinician would otherwise spend the visit discovering by accident. Gupta is describing the discovery happening the hard way, in the middle of an appointment, when a patient arrives already convinced. Asking at the start turns the interruption into information.

We Are Not Waiting

Notice what the three suggestions share. Each one assumes patients are already using these tools, because they are. Nobody learned the internet before using the internet. People opened a browser, made mistakes, asked a nephew, and worked it out, while librarians and universities built curricula most of them never encountered. Use came first and understanding followed. Reversing the order for artificial intelligence will fail exactly as it failed before, because the tool is already in the patient’s hand and no class has been scheduled.

I began with a question I have heard asked many times and have yet to hear answered. Who educates the patient? The honest answer is that the education started without anyone deciding to begin it. Patients are learning these tools at midnight, from each other, in the same disorderly way my father learned the internet in his seventies. Asking whether patients should be taught before they use artificial intelligence is asking whether to permit something already well underway. The useful question is what gets handed to the people already learning.

Sources

Shaili Gupta — Yale School of Medicine. Yale News, February 12, 2026. “Using AI chatbot for health advice? Keep these tips in mind.”

Adam Rodman — Beth Israel Deaconess Medical Center / Harvard Medical School. Harvard Health Publishing, June 11, 2026.

Robert Shpiner — David Geffen School of Medicine at UCLA. Annals of Internal Medicine, July 28, 2026.

KFF — Tracking Poll on Health Information and Trust, conducted February 24 to March 2, 2026.

Canadian Medical Association — 2026 Health and Media Tracking Survey, published February 10, 2026. Conducted by Abacus Data.

Liaison Strategies — Canadian AI in health care survey, released May 27, 2026.

Jeffrey Flaks — Hartford HealthCare. U.S. News & World Report, May 6, 2026.

OpenAI — “AI as a Healthcare Ally,” published January 5, 2026. ChatGPT Health launched January 7, 2026.

Sarah Blakeslee — “The CRAAP Test,” LOEX Quarterly 31, no. 3 (2004).

SCONUL — 1999 position paper, Seven Pillars of Information Literacy.

Sam Wineburg and Sara McGrew — “Lateral Reading: Reading Less and Learning More When Evaluating Digital Information.” Stanford History Education Group Working Paper, September 2017.

Bill S-5, Connected Care for Canadians Act — 45th Parliament, 1st session. Sponsor: Hon. Sen. Pierre Moreau. Senate third reading May 26, 2026. House of Commons first reading May 28, 2026.

Healthcare Excellence Canada — formed 2020 from the amalgamation of the Canadian Foundation for Healthcare Improvement and the Canadian Patient Safety Institute.