MEC Practice Academy 11.19
Blog|Articles|November 16, 2026

When AI disagrees with the doctor, patients trust the doctor less

Fact checked by: Todd Shryock
Listen
0:00 / 0:00

Key Takeaways

  • A discordant AI recommendation can transfer reputational harm to clinicians, indicating patients evaluate the human–algorithm dyad rather than separate sources.
  • Aggregate positivity toward AI telehealth clusters around operational tasks and low-emotion guidance, where convenience and nonjudgmental interaction are valued.
SHOW MORE

A Penn State experiment found that an AI second opinion that contradicts a doctor makes patients see that doctor as less credible, and even lazier. Here's where patients welcome AI in telehealth, where they still want a human and why disagreement needs careful framing.

In a 2026 controlled experiment out of Penn State, 135 people talked through a stress-management plan with a doctor. Half the time, an AI second opinion appeared and agreed with the doctor's advice. The other half of the time it disagreed. When it agreed, patients trusted the recommendation more. When it disagreed, something stranger happened: patients didn't just doubt the AI. They ended up rating their human doctor as "lazier."

Nobody told the doctor to try harder. Nobody changed what the doctor said. The only thing that moved was a chatbot's opinion, and it was enough to make a real physician look worse in the eyes of the person they were treating. That finding is the clearest evidence yet of something telehealth has been quietly building toward for years: patients no longer evaluate AI and their doctor separately. They evaluate them as one act, and the algorithm now gets a vote in how the human is judged.

A 2026 analysis of more than 25,000 online mentions of AI-enabled telemedicine found an overwhelmingly positive sentiment score across every platform it measured: news, social media, blogs and video. A separate study of youth mental health telehealth use found 68% positive sentiment against 22% negative, roughly three to one. Add it up and it's tempting to conclude the public has simply made its peace with AI in medicine.

But it hasn't, and the real picture is not that rosy. People have made peace with a specific, narrow slice of it.

Look at where that positivity actually concentrates: convenience, scheduling and general information, the parts of care that were never emotionally loaded to begin with. Some patients even say an AI feels less judgmental than a person, which matters for anyone who's avoided a doctor's office out of shame or fear. That's a genuine benefit, not a consolation prize.

But that same positivity thins out fast the moment the stakes rise. In a qualitative study of cancer survivors using AI conversational tools, every single participant interviewed said they preferred talking to a human when the conversation touched anything emotionally difficult.

So it's safe to say that patients like AI for the parts of health care that were always closer to logistics than medicine. They still want a person for the parts that were always about being seen, not processed. The industry's habit of citing one average sentiment score obscures that split entirely. That's not a small oversight. It's the difference between deploying AI where it helps and deploying it where it quietly erodes the thing patients actually came for.

The widely repeated industry claim that public sentiment toward telehealth runs roughly 13 times more positive than negative traces back to no findable original study. It shows up in research roundups, gets cited as established fact and dissolves the moment you try to find where it actually came from. That's the kind of number that makes its way into a pitch deck, not a peer-reviewed finding, and it's worth being skeptical of any AI-in-health statistic that sounds too clean to check.

To sum it up, the Penn State finding is the one that should worry health systems rushing to bolt an AI second opinion onto every visit. If a chatbot's disagreement is enough to make a competent doctor look lazy to the person sitting across from them, then deploying AI without thinking hard about how disagreement gets framed isn't a neutral technology choice. It's a decision about whose reputation absorbs the risk when the algorithm and the clinician don't line up, and right now, that risk is landing on the clinician by default, without anyone actually deciding it should.

AI in telehealth isn't failing. By most available measures, it's landing better than skeptics expected. But "patients like it" and "patients like it everywhere, for everything" are two very different claims, and health care keeps quietly swapping one for the other.

Yodhin Aggarwal, BDS, MHA, is a clinician and healthcare administrator.


Related to this article

AMA targets downcoding, AI licensure; Jacksonville pharmacy owner indicted in $51M fraud; drug-response gene variants found in every participant — Morning Medical Update © everythingpossible - stock.adobe.com
The top news stories in medicine today: The AMA moves against insurer downcoding and says AI can't be licensed as a physician; a Jacksonville pharmacy owner is indicted in a $51 million health care fraud scheme; an NIH study finds every participant carried a gene variant that can affect drug response.