You've probably done this at least once. Maybe after an appointment that ran 12 minutes — long enough to get a diagnosis, too short to understand it. Maybe when you were prescribed something with a long name and wanted to know what it actually did before taking it. Maybe late at night, when something felt off, and calling anyone seemed like an overreaction.

If so: you're in the majority. Pew Research surveyed 3,488 US adults in late June 2026 and found 34% of Americans use AI chatbots for at least one health task. One in three. The most common reasons: fetching quick health information (28%), understanding symptoms (25%), making sense of what a doctor said. A quarter use AI specifically to try to figure out a diagnosis on their own. Forty-seven percent of those users call the answers extremely or very helpful.

I don't doubt that. These tools are genuinely useful for a specific kind of health question. But I want to tell you about something that happened in England the same month.

The word that got dropped

A woman had an MRI. It came back clean — no nerve damage. The hospital she visited uses an AI "scribe," software that listens to the appointment and types up the consultation summary so the doctor doesn't have to. The scribe wrote "demyelination" in her notes. Demyelination is the kind of nerve damage that can be an early sign of multiple sclerosis. The correct result was "null demyelination," meaning none was found.

One word. Dropped. Reversed the meaning of the test.

She's a health professional herself. She knew to question the summary. The hospital corrected it. But if she'd filed the notes away and trusted them — which is what most people would do — a false nerve-damage diagnosis would be sitting in her medical record.

Healthwatch England, the NHS's statutory patient watchdog, documented this case and others in August 2026. A separate scribe swapped a prescribed drug for a different one with a similar name. A third scribe left out the instruction that a patient needed to get a prescription refilled through their GP. England now has 27 different AI scribe products running in clinical settings. None are classified as medical devices, so there's no standardized safety oversight.

In every single documented case: the patient caught the error. Not the AI. Not the doctor.

Source spread

• Pew Research — "From Diagnoses to Treatments: Why Americans Use AI Chatbots for Health" [academic] — primary data source; 3,488 adults surveyed June 22–28, 2026
• The Guardian — "Doctors' AI scribes get names of drugs and diagnoses wrong, NHS watchdog warns" [skeptic] — documents the Healthwatch England cases with specifics on each error type
• OpenAI — "Improving health intelligence in ChatGPT" [hype] — the June 2026 announcement connecting free ChatGPT to clinical databases and medical guidelines

What's real:

AI is good at health translation. Medical jargon is dense. Asking ChatGPT or Claude to explain a diagnosis in plain English, or to tell you what a drug generally does, is exactly the kind of task these models handle well. If your doctor used the word "hyponatremia" in passing and you have no idea what it means, an AI will give you a working definition faster and more clearly than Googling and landing on a WebMD worst-case article. That's a real, legitimate use.

Access gaps are real. Twenty-two percent of AI health users cited low or no cost as a reason they turn to AI. Not everyone has a primary care doctor they can call with a quick question. If the choice is "AI chatbot at 11pm" or "nothing until Tuesday," AI wins that comparison. That matters.

The tools are getting better. OpenAI added clinical health intelligence to free ChatGPT in June 2026, connecting it to medical databases and guidelines. The gap between "AI health advice" and "actual clinical information" is narrowing in real ways.

What deserves a side-eye:

AI is built to generate fluent sentences. It isn't built to flag when it dropped a word that matters. Here's why the demyelination error is structurally predictable, not a fluke: AI models work by predicting what word comes next based on training data. "Demyelination" appears in medical documents far more often than "null demyelination." So the model generates the shorter, more familiar version — and drops the qualifier that reversed the meaning. This failure mode hits hardest on negations and qualifiers: words like "no," "not," "null," "negative." The cases where one word changes everything.

If AI-generated health documents end up in your file, you're the quality control. In every documented NHS case, the only thing that caught the error was the patient's own attention. Not a review system. Not the doctor. A health professional who recognized a result that didn't match what she'd been told. For everyone else: there's no backstop. The error either gets caught by you, or it doesn't get caught.

Helpfulness and trust don't line up. Forty-seven percent of users find AI health answers very helpful. Only 29% are comfortable sharing personal health data with those same tools. People are getting value while being uncomfortable with what the tool knows about them. Both things are true at once. That tension is worth sitting with.

What to do about it

Five adjustments. None take more than a minute.

• Use AI freely for explanation. Understanding what a diagnosis means, what a lab value indicates, what a drug generally does — AI is reliable and useful for this. Ask it constantly.
• For specific drug names and dosing, verify with the pharmacist. The general drug information is usually right. The specific name can be wrong. That's the error with the most direct consequences.
• If you receive AI-generated documentation from a health provider, read it carefully before filing it away. Pay attention to diagnosis words, medication names, and qualifier words — "no," "not," "null," "negative." A dropped qualifier is the documented failure mode.
• Cross-check anything important. AI is a starting point, not a final answer. For a condition you're actually managing, a prescription you're about to fill, or a test result you're worried about: ask the doctor or pharmacist directly. That was always true. It's just worth saying out loud.
• For mental health: be thoughtful about ongoing reliance. A quick "is this level of anxiety normal" check is different from sustained AI-as-therapist use. Pew found Americans split two-to-one — 39% say AI chatbots do more harm than good for loneliness, versus 19% who say they help. That public instinct is worth taking seriously.

Further reading

• Pew Research — "From Diagnoses to Treatments: Why Americans Use AI Chatbots for Health"
• Pew Research — "Do Americans Think Chatbots Help or Hurt People Using Them for Loneliness, Depression or Stress?"
• The Guardian — "Doctors' AI scribes get names of drugs and diagnoses wrong, NHS watchdog warns"
• OpenAI — "Improving health intelligence in ChatGPT"