AI and Your Doctor
An independent analysis of patient behavior, physician resistance and the trust gap reshaping primary care. Patients are using AI. The question is whether their physician is in the conversation.
Updated November 2026: redesigned; renumbered from White Paper to Research Brief No. 10. The KFF figure is corrected to 56% (not confident telling true from false). Claims CMT could not verify against an original source were removed: the "1 in 4" provider statistic, the Google search share, the Harvard 88% vs 96% comparison, the Sermo poll and the 6.3% primary care error figure.
Start here: the questions this brief answers
Half of U.S. adults now use AI for important health decisions, even as their comfort with AI falls. Research shows AI is strong at naming a final diagnosis and weak at early reasoning, so a human is still needed. Physicians make errors too. Practices that respond with curiosity rather than dismissal, and use AI to give time back, are positioned to lead.
Does AI really misdiagnose 80% of cases?
No. It missed appropriate early differentials over 80% of the time but got final diagnoses right over 90% of the time in the best models. Go to the full answer ↓
How many patients use AI for health decisions?
51% of adults in a January 2026 national survey. Go to the full answer ↓
Aren't physicians more accurate?
Usually, but 23% of seriously ill hospitalized adults in one study had a diagnostic error. Go to the full answer ↓
How should we respond when patients bring AI output?
With curiosity, not dismissal. See the conversation checklist. Go to the full answer ↓
What does this mean for membership practices?
Use AI to give time back, and be reachable so patients rely on it less as a substitute. Go to the full answer ↓
Part 1 · The headline and the nuance
"AI misdiagnoses 80% of cases" is not what the study said
In April 2026, a Mass General Brigham study in JAMA Network Open tested 21 large language models on 29 standardized clinical vignettes, feeding information step by step as a physician would receive it.1 The models failed to produce an appropriate differential diagnosis more than 80% of the time early in a case. But with complete information, the best models reached the correct final diagnosis more than 90% of the time; across models, final-diagnosis success ranged from about 60% to over 90%.1 The authors' conclusion was calibrated: these tools still need a human in the loop and close oversight.
How the same AI models performed at two stages of a case
Practice InsightSource: Rao et al., JAMA Network Open, 2026.1 Bars are drawn at the reported thresholds. The gap is the story: AI names answers well and reasons through uncertainty poorly.
Both halves of the finding matter to patients and physicians. AI is good at naming an answer once the facts are in, and weak at the open-ended start of a case, which is exactly where a physician earns trust.
Part 2 · What patients are doing
Half of adults made a health decision with AI and no clinician
of U.S. adults used AI to make an important health decision without consulting a medical professional (January 2026).2
Population Dataof adults aren't confident they can tell true from false in AI chatbot answers (2024).3
Population DataWhat adults say they use AI for in health (January 2026)
Population DataSource: Ohio State University Wexner Medical Center survey of 1,007 adults, fielded by SSRS, margin of error ±3.5 points.2
The pattern is not blind faith. Openness to AI in care is falling even as use rises,2 and most adults doubt chatbots' health accuracy.3 Meanwhile, the share of adults reporting "a lot of trust" in physicians and hospitals fell from 71.5% in 2020 to 40.1% in 2024.4 Patients are filling a gap, not rejecting their doctors.
Patients are not embracing AI because they want less humanity in their healthcare. They are using AI because they want more of it.
Michael Tetreault, Editor-in-Chief, Concierge Medicine Today (CMT's view)5
Part 3 · Whose error rate?
Hold AI and ourselves to the same standard
Diagnostic error is a human problem too. In a 2024 study of 2,428 seriously ill hospitalized adults at 29 academic centers, 23% had a diagnostic error, and errors contributed to harm in about 18%.6 The study looked only at patients who died or were moved to intensive care, so it does not give an overall hospital error rate. That is not an argument against physicians. It is an argument against pretending any diagnostician, human or machine, is perfect.
Physicians are already adopting AI themselves: two-thirds reported using it in 2024, mostly for documentation and administrative work.7
Part 4 · What resistance communicates
Caution builds trust; dismissal ends it
Dismissal
You shouldn't trust that.
Tells the patient their effort was foolish, and they may stop sharing what they read.
Caution with curiosity
That's an interesting starting point. Let's look at what you found together.
Keeps the physician in the loop, which is where the research says a human belongs.1
Teaching
AI is good at naming possibilities. My job is to figure out which ones fit you.
Explains the differential-diagnosis gap in plain language.
Patient AI conversation checklist for your practice
0 complete
Nothing you check is saved or sent anywhere. Educational self-review only.
Part 5 · The leadership move and the business reality
Be the director, not the reactor
Burnout is easing but still affects about 42% of physicians,8 and a 2024 survey found more than a third of burned-out primary care physicians planned to stop seeing patients within one to three years.9 Membership practices were built partly as an answer to that: smaller panels and physicians patients can actually reach. These models are not anti-technology. They are pro-relationship.
Learn the tools your patients use
Try the same chatbots your patients use with real (de-identified) questions, so you know their strengths and failure modes.
Invite AI into the visit
Ask patients to bring what they found. Treat it as history, not as a threat.
Use AI to give time back
Most physician AI use is documentation.7 Reinvest saved time in the conversation AI can't have.
Protect privacy
Never enter identifiable patient information into consumer tools. See Brief No. 36 on cybersecurity and AI privacy.
Learn → Build → Lead
Putting it to work
AI names answers well and reasons through uncertainty poorly; patients know it's imperfect.
Adopt the conversation checklist and a written staff AI policy.
Be the physician who understands AI best in the room, and use it to give patients more of you.
Methods, limitations and evidence types
How this brief was built
CMT re-verified each statistic in the original white paper against the original publisher. Five claims that relied on aggregators or could not be traced were removed. This brief does not report concierge or DPC practice data.
Every key finding is labeled by evidence type. Labels describe the type of evidence, not its value. Each type answers different questions. Funding is disclosed on every source.
What we don't know
- AI capabilities change quickly; the 2026 study tested models available at submission.
- Survey data show what patients report doing, not whether outcomes improved or worsened.
- No study yet compares AI use or outcomes among concierge patients specifically.
How to cite this brief
Concierge Medicine Today. AI and Your Doctor. CMT Research Brief No. 10, Version 2.0. November 2026. conciergemedicinetoday.net.
External review: This brief has not been externally peer reviewed. Corrections: CMT corrects factual errors promptly and notes the change in the updates line. Send corrections through conciergemedicinetoday.net/contact.
Keep reading
Related CMT Research Briefs
References
Sources and citations
- Rao AS, et al. Large language model performance and clinical reasoning tasks. JAMA Network Open. April 2026. doi:10.1001/jamanetworkopen.2026.4003. 21 models, 29 vignettes. jamanetwork.comEvidence type: Practice Insight · Funding: not stated (check article)
- Ohio State University Wexner Medical Center. Public comfort with AI in health care falls. News release, April 7, 2026. SSRS probability-based panel, 1,007 adults, January 16 to 20, 2026; margin of error ±3.5 points. wexnermedical.osu.eduEvidence type: Population Data · Funding: not stated (commissioned by Ohio State University Wexner Medical Center; fielded by SSRS)
- Presiado M, Montero A, Lopes L, Hamel L. KFF Health Misinformation Tracking Poll: Artificial intelligence and health information. KFF. August 15, 2024. 2,428 adults, June 3 to 24, 2024. kff.orgEvidence type: Population Data · Funding: not stated (designed and analyzed by KFF; fieldwork by SSRS)
- Perlis RH, Ognyanova K, Uslu A, et al. Trust in physicians and hospitals during the COVID-19 pandemic in a 50-state survey of US adults. JAMA Network Open. 2024;7(7):e2424984. jamanetwork.comEvidence type: Population Data · Funding: not confirmed by CMT; see the article's funding section
- Concierge Medicine Today. Editor-in-Chief commentary via the CMT Media Desk. conciergemedicinetoday.netEvidence type: Industry Research · Funding: Concierge Medicine Today (self-conducted; no outside funder)
- Auerbach AD, Lee TM, Hubbard CC, et al. Diagnostic errors in hospitalized adults who died or were transferred to intensive care. JAMA Internal Medicine. 2024;184(2):164-173. 2,428 records, 29 academic centers. psnet.ahrq.govEvidence type: Population Data · Funding: not stated (check article; authors declared no competing interests per Harvard Gazette)
- American Medical Association. 2 in 3 physicians are using health AI, up 78% from 2023. AMA News Wire, February 26, 2025. Survey of about 1,200 physicians. ama-assn.orgEvidence type: Industry Research · Funding: not stated (AMA survey)
- Medical Economics. Physician burnout declines overall, but gaps persist across specialties, AMA data show. April 16, 2026. medicaleconomics.comEvidence type: Industry Research · Funding: not stated (AMA Organizational Biopsy data, reported by Medical Economics)
- Horstman C. A poor prognosis: more than one-third of burned-out U.S. primary care physicians plan to stop seeing patients. Commonwealth Fund, December 6, 2024. commonwealthfund.orgEvidence type: Population Data · Funding: foundation (Commonwealth Fund, self-conducted)
Disclaimer
This brief is educational and informational only. It is not medical, legal, tax, financial or accounting advice, and it is not media training or a substitute for advice from a qualified professional. Concierge Medicine Today is an independent trade publication: no organization paid for favorable coverage, no practice model is favored, and funding is disclosed for every source cited. Product, company and organization names are used for identification and context only and do not imply endorsement. While we work to be accurate, this content is not without possible error or omission. © 2007-2026 Concierge Medicine Today, LLC. All rights reserved.
Concierge Medicine Today · independent trade publication since 2007 · Learn → Build → Lead

