Will AI Replace Doctors? What the Clinical Relationship Still Carries
- 5 days ago
- 9 min read
By Dr. Ernest Wayde, PhD, MIS
Clinicians are increasingly asking whether AI could replace them, and not just for tasks like documentation or triage. The deeper question is whether it could replace the relationship they have with their clients and patients.
Consider an intake session. A client answers the same question three times, each time slightly differently: once flat, once with a strained laugh, once after a long pause. No single answer stands out on its own. Together, they suggest something the client isn't saying directly. The clinician notices this because of experience, not because any one answer contained an obvious signal.
That kind of noticing, picking up on a pattern across small variations rather than any single data point, is close to the center of what people mean when they ask whether AI could replace a therapist or a doctor.
The fear, named plainly
Two fears travel together under that question. The first is job displacement. It is a reasonable concern. Clinicians and physicians build careers over a decade or more of training, and the tools now handling documentation, triage, and information delivery are advancing quickly. Patients and clients depend on these professionals staying in the field, in practice, and in a position to be paid for the work they do. Dismissing that fear as overblown would be its own kind of denial.
The second fear runs underneath the first, and it's less about employment and more about function. It asks whether the human part of the relationship, the part that isn't a task, is actually necessary, or whether it has just been preferred out of habit and comfort while a machine slowly gets good enough to do the same job.
What AI is actually doing
Some of what AI is doing in clinical settings is well documented.
Ambient AI scribes, tools that listen to a clinical conversation and generate the note automatically, have moved from pilot programs into real deployment (Preiksaitis et al., 2026). A 2026 cohort study of emergency department encounters, published in Annals of Emergency Medicine, found that ambient AI was used in about one in nine eligible visits, with over a third of attending physicians using the tool at least once, though a small group of frequent users accounted for most of the actual use (Preiksaitis et al., 2026). A separate randomized trial of 238 outpatient physicians across 14 specialties, published in NEJM AI, found that one of two ambient AI scribe tools tested, Nabla, significantly reduced documentation time compared to usual care, while the other tool, DAX, showed no significant change. Both tools showed potential improvements on burnout and work environment measures, a reminder that ambient AI tools don't perform uniformly, even within the same study (Lukac et al., 2026).
Diagnostic support tells a more complicated story. A 2025 study in the Proceedings of the National Academy of Sciences found AI models matching or exceeding most individual physicians' accuracy on diagnostic case vignettes, with the strongest results coming from hybrid teams that combined physicians and AI rather than either working alone (Zöller et al., 2025). A separate meta-analysis in npj Digital Medicine, pooling 83 studies, found the opposite pattern for messier, real-world cases: generative AI performance drops noticeably when it's handed the kind of detailed information found in an actual electronic health record, rather than a clean, isolated vignette (Takita et al., 2025). The honest summary is that AI is capable in narrow, well-defined tasks, and its performance is less reliable the closer a case gets to the complexity of actual practice.
None of this settles the question of what AI means for the relationship. That's a different claim, and it needs its own evidence.
What AI hasn't touched
AI can get significantly better at diagnosing, documenting, and answering questions without that progress touching the relationship at all. The two are not the same axis. A tool could reach expert-level accuracy on a task and still have nothing to offer the part of care that depends on being known by someone over time.
That part of care is measurable, not just felt. Research on the therapeutic alliance finds it predicts outcomes independently of which technique a clinician uses, which means the relationship does some of the work itself, not just delivers the treatment (Flückiger et al., 2018). Two things in particular explain why: clinical judgment, and accountability.
Clinical judgment
Clinical judgment is a capacity that develops through training and accumulated practice, and it shows up most clearly in exactly the kind of moment described at the start of this piece: noticing that something doesn't fit, before being able to say what.
A study on nursing practice, published in Nursing Education Perspectives, found that years of clinical experience were the strongest independent predictor of clinical judgment scores when comparing novice and expert nurses, a stronger predictor than the number of training simulations completed (Shinnick & Cabrera-Mino, 2021). That finding is specific to nursing, and it shouldn't be stretched further than the study itself supports. But it's consistent with what most clinicians would say about their own development: judgment sharpens with time in the room, not just time in the classroom.
Picture the intake scene again. A clinician who has sat with hundreds of clients recognizes the shape of an answer that keeps changing, even without being able to name the pattern immediately. That recognition is judgment being exercised, built from years of exposure to variation that no single case could teach on its own, rather than a rule being applied.
Accountability
The second thing the relationship carries is accountability, and the honest version of this point is less comfortable than it might first sound.
The instinct is to frame accountability as a clean advantage: a licensed clinician can be held responsible for a decision, and an AI tool cannot. That's true, but it isn't reassuring, because that same fact cuts the other way. A 2024 analysis in Biomedical Instrumentation & Technology found that under current tort law, a clinician who relies in good faith on an AI tool's recommendation can still be held liable for malpractice if that recommendation turns out wrong and harms the patient, regardless of what the tool suggested, since courts have consistently required clinicians to apply their own independent judgment, whatever the algorithm’s output (Lee et al., 2024). A separate 2024 paper in the Future Healthcare Journal gives this a name: a "liability sink," where the clinician absorbs responsibility for the AI's recommendation while having limited say in how that recommendation was generated (Lawton et al., 2024). And a recent World Health Organization Europe survey found that legal uncertainty is the leading barrier to AI adoption across the region, with fewer than one in ten countries surveyed having liability standards that clarify who else, if anyone, might share that responsibility going forward (World Health Organization Regional Office for Europe [WHO/Europe], 2025).
What this means in practice: the clinician is the one left holding a bad outcome when an AI tool contributed to it, not because the field decided that was the right place for responsibility to sit, but because current law hasn't built another option. That's one more reason the relationship can't simply be handed off to a tool that isn't positioned to carry what the clinician carries.
Where the evidence is strongest
Of everything in this piece, the clearest evidence concerns the therapeutic alliance itself. A major review of psychotherapy research by Flückiger and colleagues, published in the journal Psychotherapy and pulling together nearly 300 studies and over 30,000 patients across four decades, found that the strength of the relationship between therapist and client reliably predicts how well treatment works, in person or online (Flückiger et al., 2018). That means the alliance is doing real work in the outcome, not just supporting whatever technique the therapist happens to use. Decades of consistent findings back that up.
The equivalent research in medicine tells a thinner story. A separate study by Kelley and colleagues, published in PLOS One, looked specifically at the patient-physician relationship outside of mental health care and found it still helps, but the effect was smaller and the evidence base was much thinner (Kelley et al., 2014). Only a handful of trials were designed rigorously enough to test the question directly, and the researchers who ran the review said as much themselves. The physician-patient relationship still matters. It just hasn't been studied, or shown to matter, with the same depth or strength as it has in psychotherapy.
The case for physicians in this piece was never resting on alliance research alone. It rests on judgment and accountability, the two things already shown earlier to carry real weight regardless of specialty. Those don't get weaker just because the alliance evidence does.
Why reassurance doesn't land
Telling a clinician that AI still can't fully do their job doesn't resolve either fear named at the start of this piece.
The displacement fear isn't answered by naming a capability ceiling, because that ceiling keeps moving. What AI can't do reliably today is a moving target, not a fixed boundary, and pointing to today's limitations offers only temporary comfort.
The relational fear isn't answered by a task comparison at all, because it was never about tasks. Judgment and accountability aren't functions that show up on a diagnostic accuracy benchmark. They're properties of a relationship, built through history, context, and the willingness to be responsible for what happens next. No task-level capability score changes that.
What this means for the clinician now
Two things follow from all of this, and both deserve to be said plainly rather than folded into each other.
Task substitution is real, and it's accelerating. Pretending otherwise doesn't serve anyone. The useful question for a clinician today is which tasks are worth ceding, on what terms, and with what oversight, because that shift is already underway and will keep moving regardless of how any individual clinician feels about it.
What doesn't transfer, doesn't transfer because of what it structurally requires, not because AI hasn't caught up yet. Judgment built through years of accumulated practice, accountability that currently concentrates in the clinician by default, and an alliance with a research base showing it functions as a genuine mechanism of change: none of that is a capability gap waiting to close. It's a different kind of thing than a task, and it doesn't get replaced by a better version of the same tool. That's the argument this piece is making, and it's the reason the fear, real as it is, doesn't resolve the way task-substitution reassurance suggests it should.
Understanding where that line actually sits, and what it means for how AI gets integrated into a practice responsibly, is the work we do at Wayde AI.
FAQ
Is AI actually replacing therapists or doctors right now?
AI is handling specific tasks today, documentation, triage, and information delivery among them, with real and growing adoption. Full substitution is a different claim, and current research doesn't show AI replacing the working alliance itself.
Why does the therapeutic alliance matter more in psychotherapy than in medicine, according to the research?
The research base is simply deeper and the measured effect larger in psychotherapy, where a large meta-analysis found a moderate, robust relationship between alliance and outcome. The equivalent medical research, tested specifically through randomized controlled trials, found a smaller effect and a thinner evidence base. That doesn't mean the physician-patient relationship is unimportant, only that the evidence supporting it is measured differently.
If AI can't be held accountable, doesn't that make clinicians safer, professionally speaking?
Currently, the opposite is closer to true. Legal frameworks in most places haven't yet clarified who bears responsibility when an AI tool contributes to a bad outcome, and the clinician often carries the liability by default, ahead of any decision about whether that's the fair outcome.
Should clinicians be worried about job displacement?
That worry is reasonable and shouldn't be minimized. Task substitution is real and accelerating. The more useful question is which tasks are worth ceding, under what oversight, rather than whether the shift is happening at all.
About the Author
Dr. Ernest Wayde is the Founder and Principal of Wayde AI, a healthcare AI ethics consulting firm. He works with healthcare and behavioral health organizations on responsible AI adoption, governance, risk management, and implementation strategy.
References
1. Preiksaitis, C., Alvarez, A., Winkel, M., Karamatsu, M., Brown, I., Sama, N., Morris, L., Lee, J.-Y., Gubbels, A., Wahl, E., Frye, A., & Rose, C. (2026). Ambient artificial intelligence scribe adoption and documentation time in the emergency department. Annals of Emergency Medicine, 87(5), 569–574. https://doi.org/10.1016/j.annemergmed.2025.12.017
2. Lukac PJ, Turner W, Vangala S, et al. Ambient AI Scribes in Clinical Practice: A Randomized Trial. NEJM AI, 2025;2(12). http://dx.doi.org/10.1056/aioa2501000
3. Zöller, N., Berger, J., Lin, I., Fu, N., Komarneni, J., Barabucci, G., Laskowski, K., Shia, V., Harack, B., Chu, E. A., Trianni, V., Kurvers, R. H. J. M., & Herzog, S. M. (2025). Human–AI collectives most accurately diagnose clinical vignettes. Proceedings of the National Academy of Sciences of the United States of America, 122(24), e2426153122. https://doi.org/10.1073/pnas.2426153122
4. Takita H, Kabata D, Walston SL, et al. A systematic review and meta-analysis of diagnostic performance comparison between generative AI and physicians. npj Digital Medicine, 2025. https://www.nature.com/articles/s41746-025-01543-z
5. Flückiger C, Del Re AC, Wampold BE, Horvath AO. The Alliance in Adult Psychotherapy: A Meta-Analytic Synthesis. Psychotherapy, 2018. https://societyforpsychotherapy.org/wp-content/uploads/2018/10/Fluckiger-et-al-2018-Alliance-MA-Online.pdf
6. Shinnick, M. A., & Cabrera-Mino, C. (2021). Predictors of nursing clinical judgment in simulation. Nursing Education Perspectives, 42(2), 107–109. https://doi.org/10.1097/01.NEP.0000000000000604
7. Lee B, Naftalovich R, Ali S, Chaudhry FA, Tewfik GL. Liability Exposure of Clinicians in Artificial Intelligence-Driven Healthcare. Biomedical Instrumentation & Technology, 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC10987009/
8. Lawton T, Morgan P, Porter Z, et al. Clinicians risk becoming "liability sinks" for artificial intelligence. Future Healthcare Journal, 2024. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11025047/
9. World Health Organization Regional Office for Europe. (2025). Artificial intelligence is reshaping health systems: State of readiness across the WHO European Region. World Health Organization. https://www.who.int/europe/publications/i/item/WHO-EURO-2025-12707-52481-81028
10. Flückiger, C., Del Re, A. C., Wampold, B. E., & Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340. https://doi.org/10.1037/pst0000172
11. Kelley JM, Kraft-Todd G, Schapira L, Kossowsky J, Riess H. The Influence of the Patient-Clinician Relationship on Healthcare Outcomes: A Systematic Review and Meta-Analysis of Randomized Controlled Trials. PLOS One, 2014. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0094207
This article is part of an ongoing series on AI's role in clinical practice from Wayde AI. If you found this useful, consider subscribing to the Wayde AI Brief, a weekly newsletter on ethical, responsible AI adoption in healthcare and behavioral health: https://the-waydeai-brief.beehiiv.com/
Comments