Insights
Should we tell patients we're using AI?
Yes — earlier and more plainly than feels comfortable. And here is the part nobody warns you about: telling them is far less awkward than the weeks you will spend dreading it. Your patients are not asking you to defend the technology. They are asking whether a human being is still paying attention to them. Say that part out loud, and most of the worry leaves the room with it.
Your patients answered this question before you thought to ask it
In December 2024, a team led by Jodyn Platt at the University of Michigan published a national survey in JAMA Network Open — 2,021 U.S. adults, asked directly how much it mattered to them to be told when artificial intelligence was used in their care. Nearly 63% said it was very true that they wanted to be notified. Fewer than 5% said it did not matter. What surprised a lot of practice leaders was where the strongest desire came from: adults 60 and older, the exact group most practices assume would rather not hear about the computer, wanted to be told more than anyone else. That assumption is worth retiring today. Staying quiet feels protective from the inside. You are picturing a confusing conversation on a day someone already feels unwell, and you decide to spare them. From the treatment room chair it reads differently. When patients learn about it later — from a friend, from a news segment, from a disclaimer at the bottom of a letter — the thing they remember is not the software. It is that you knew and did not say. Trust is not usually lost over the tool itself. It is lost over the gap between when you knew and when they found out.
62.7%
of U.S. adults said it was "very true" that they want to be notified when AI is used in their health care; 4.8% said notification was not important to them ([JAMA Network Open, December 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11635529/), 2,021 adults). Read it honestly: this measures what people say they want when asked, not how they react in the moment — and it was taken before ambient documentation reached most practices.
From the field
One practice we work with spent the better part of a month agonizing over the wording of a single sentence about their ambient note-taking tool. Would patients think the visit was being recorded and stored somewhere? Would someone walk out? They finally settled on two lines in the intake paperwork and one sentence spoken out loud at the start of the visit. Over the first month, across several hundred visits, the flood of questions never came. A handful of patients asked something — and it was the same something, nearly word for word: does a real person still read it? One patient asked them not to use it at all. They honored it, which meant the front desk had to carry a flag on that chart and the practitioner had to type that day like it was 2019. That is the honest tradeoff, and it is smaller than the fear of it: someone has to own the exception, every time, or the promise you made quietly stops being true.
In a growing number of states, this stopped being a question of manners
California's AB 3030 took effect in January 2025. If a practice sends a patient communication about clinical information that was generated by AI, it has to carry a disclaimer saying so, along with clear instructions for how that patient can reach a human being. There is one exemption worth sitting with: communications that a licensed provider reads and reviews before they go out are not covered. The law is quietly rewarding the workflow you would want anyway — a person's eyes on anything clinical before it reaches a patient. Texas went further. SB 1188, effective September 2025, permits AI in diagnosis and treatment on the condition that the practitioner personally reviews the AI's output before making the clinical decision, and requires disclosure of AI use to patients. The Texas Responsible AI Governance Act, effective January 2026, adds that the written disclosure must be conspicuous and must come before or at the time of the interaction, not buried afterward. Most states still say nothing at all, and none of this is legal advice — check your own state and ask your malpractice carrier what they expect of you. But the direction of travel is not ambiguous. The practices that will have the easiest few years are the ones already saying it out loud, because the rule, when it arrives, will only ask them to keep doing what they do.
The sentence your patients are listening for is not about the software
You do not need a policy binder or a page of small print. Three things belong in a disclosure, and they take about eight seconds to say: what the tool actually does, who is responsible for the result, and how to opt out. Something close to this works in almost any treatment room. "I use a tool that listens and drafts my notes, so I'm not typing while you're talking to me. I read and correct every note before it goes in your chart. If you'd rather I not use it today, just say so — that's completely fine." Notice what is doing the work there. Not the explanation of the technology, which almost nobody wants. It is the second sentence: a named human being stands behind the output. Notice also the shape of the first one. You are not telling them about a piece of software; you are telling them why you can look at them now instead of at a screen. That is the truest reason to say it out loud, and it has nothing to do with compliance. The whole point of a tool like this is that it buys back attention. Handing that attention to the person in front of you, and telling them plainly that is what you did with it, is not a disclosure. It is the care itself.
So write the sentence this week. Read it aloud once to hear whether it sounds like you. Put it in the intake paperwork, say it at the start of the visit, and decide now who owns the chart flag when someone says no — that is the whole implementation. There is no perfect moment coming, and no version of this conversation that gets easier by waiting. The practices that handle it best are not the ones with the best wording. They are the ones who made it ordinary early, so that by the time anyone thought to ask, the answer had already been sitting there in plain sight for a year.