Sermo research Report | AI in healthcare | 2026
Only 34% of clinicians who use AI always verify its output
Adoption of artificial intelligence tools among healthcare professionals (HCPs) is widespread, with 92% of surveyed HCPs reporting that they use them in their work. However, AI training, output verification, and documentation still lack a common approach.
Sermo surveyed thousands of U.S. members to understand which AI tools healthcare professionals use, how they check AI output, what patients bring into consultations, and where experiences differ by role and practice setting.
At a glance
- Documentation practices are inconsistent: 33% routinely document AI use in patient charts, while 34% say they do not document it at all.
- Verification is not standard practice: Just 34% say they always check AI output against a second source.
- Two platforms dominate: OpenEvidence (63%) and ChatGPT (53%) are the most commonly used AI tools.
- Liability concerns change behavior: 52% of HCPs say liability worries have changed how they use or document AI.
- Concerns about patient AI use are widespread: 98% of respondents have at least one specific worry about patients’ use of AI to interpret health-related information.
How does AI use in your practice compare?
Join a community of verified healthcare professionals to discuss how AI is changing your day-to-day workflows, for better or worse.
Introduction
Why does AI in practice matter to HCPs?
This report benchmarks several facets of AI use —tools, verification habits, access, training, and patient impact—so healthcare professionals like you can see exactly where you stand compared with peers.
For physicians
AI tools’ outputs and generated responses aren’t always accurate. If not caught, false or mischaracterized information may persist in patient records, potentially affecting patient care and outcomes. Physicians are ultimately still accountable for decisions made based on information from AI tools. Knowing which tools your colleagues use, how they verify and document output, and how to check AI’s work is becoming a necessary skill.
For advanced practice providers and other HCPs
AI outputs are often passed through many team members. Scribes draft, APPs edit, nurses act on summaries. If these outputs aren’t verified, mistakes can cascade downstream, causing miscommunication, increased workload, and even patient risk.
Key Findings
AI in healthcare: the current state
Five key findings define where the healthcare profession stands on AI today.
1. AI adoption is high, but young
AI adoption is no longer a future trend in healthcare; it is already embedded in day-to-day clinical practice. Nearly all HCPs surveyed by Sermo reported using AI tools as part of their work, signaling widespread interest in technologies that can support documentation, information gathering, clinical decision-making, and administrative efficiency.
However, adoption does not necessarily equate to maturity. The data suggests that healthcare’s AI journey is still in its early stages, and many clinicians are still learning how to incorporate AI effectively, evaluating where it adds value and developing best practices for responsible use. As organizations and individuals gain experience, healthcare is likely to move from broad experimentation toward more sophisticated, standardized, and outcome-driven applications of AI.
Over half of respondents who use AI (55%) have less than a year of experience, and only 14% have more than two years.
2. There is no single approach to documenting AI use
Routine documentation is highest in hospitals and health systems, with 39% of respondents who work in this setting reporting that they “routinely” document AI use in patient charts. In comparison, 28% of respondents in group private practice and 24% of respondents in independent private practice regularly do so. Another 28% of respondents in other practice settings (including academic/teaching, government/VA, and locum tenens) routinely note AI use in their records.
These fragmented responses suggest clinicians are not operating under a consistent documentation norm, pointing to real ambiguity or a lack of guidance on documentation expectations.
34% of HCPs using AI don’t document it at all, 33% routinely document it in charts, and 26% document it only sometimes.
Do you document your use of AI in patient charts?
| No, I do not document AI use | 34% |
|---|---|
| Yes, routinely | 33% |
| Sometimes, depending on the situation | 26% |
| I am not sure what the expectation is at my institution | 8% |
- No, I do not document AI use (34%)
- Yes, routinely (33%)
- Sometimes, depending on the situation (26%)
- I am not sure what the expectation is at my institution (8%)
3. AI output verification is common, but not universal
94% of healthcare professionals apply at least some scrutiny to AI output, but only 34% always check it against a second source.
A significant portion of HCPs (41%) reported being self-taught in AI, or learning through peer recommendations/informal discussion (32%). Only 18% received institutional training or onboarding. Because documentation practices also vary widely by setting, the finding suggests clinicians need standardized AI training that includes protocols for verifying outputs.
AI training is associated with more consistent verification of AI output. 49% of respondents who received AI training via CME or another accredited program reported that they “always verify” AI output against a secondary source, as do 39% of respondents who received institutional or onboarding training. Members with no training show the weakest verification patterns: 9% “rarely” verify and only 31% “always” do.
“In other MDs’ notes, I can see where they used AI or DAX Copilot. Clear that the AI actually picked up the wrong name of medication, and it got populated forward…for a chemo the patient had no business being on. Similar issues with people not reading dictations, but now because it comes with so much additional fluffy text, it’s harder to spot that wasn’t the clinician’s intent. I have called the MDs and made them change their notes.”
How do you typically confirm the accuracy of AI outputs?
| I sometimes verify, depending on the situation | 60% |
|---|---|
| I always verify against a secondary source | 34% |
| I rarely verify the output | 5% |
| I do not use AI in my practice | 1% |
- I sometimes verify, depending on the situation (60%)
- I always verify against a secondary source (34%)
- I rarely verify the output (5%)
- I do not use AI in my practice (1%)
4. AI tool preferences vary more by practice setting than by care team role
As AI experience increases, use of AI tools rises, suggesting that more experienced AI users tend to use a broader mix of tools rather than relying on a single solution.
Among respondents, care team roles do not vary much in their choice of AI tool – the spread is minimal. Clinicians use general AI tools broadly and often in combination, while work-related AI tools may see higher adoption when they are specialized for professional tasks or embedded directly into existing workflows. Rather than replacing one another, the data suggests that respondents use both categories of AI tools for different purposes.
Overall, survey respondents report using the HIPAA-compliant tool OpenEvidence most often (63%), with ChatGPT second at 53%.
Which of the following AI tools do you currently use?
| OpenEvidence | 63% |
|---|---|
| ChatGPT | 53% |
| EHR-embedded AI | 42% |
| Gemini | 21% |
| Claude | 18% |
| Other tool | 16% |
| Abridge | 12% |
| NuanceDAX | 11% |
| Perplexity | 4% |
| Nabla | 2% |
| Do not use AI tools | 1% |
- OpenEvidence (63%)
- ChatGPT (53%)
- EHR-embedded AI (42%)
- Gemini (21%)
- Claude (18%)
- Other tool (16%)
- Abridge (12%)
- NuanceDAX (11%)
- Perplexity (4%)
- Nabla (2%)
- Do not use AI tools (1%)
5. The perceived value of AI is higher among more experienced users
60% of surveyed HCPs say AI’s value relative to its cost has increased over the past year, and that number climbs the longer the respondent has used AI: 70% of respondents with 2+ years of AI experience report increased perceived value, compared to just 45% of those who have used it for less than six months.
The survey shows an association between AI experience and perceived value, but it does not establish that more time using AI causes that change. A number of factors could influence value perception:
- Self-selection: HCPs who find AI unhelpful may stop using it, so the long-tenured group may include users who already found value.
- Early adopters: Clinicians who started using AI earlier may be more tech-enthusiastic, or may have access to better tools.
- Tool and setting: Longer-tenured users may work in settings with better-supported AI tools.
Among experienced AI users, ChatGPT use rises from 39% (<6 months) to 66% (>2 years), and OpenEvidence from 50% to 70%.
Experienced AI users show higher adoption across leading tools
| ChatGPT | 39% | 50% | 60% | 66% |
|---|---|---|---|---|
| OpenEvidence | 50% | 60% | 71% | 70% |
- Under 6 months
- 6-12 Months
- 1-2 years
- Over 2 years
HCP Benchmarks
AI use by role: physicians vs. APPs
Survey results show similarities and differences in how APPs (n=1022) and physicians (n=1820) receive AI training and use AI tools in practice. Here’s how your role compares:
Physicians:
- Physicians lean on general-purpose AI more than any other group — 22% use Claude and 25% use Gemini, which is roughly twice the usage rate of APPs (11% and 13%). Physicians aren’t just adopting AI more; they’re exploring it more broadly.
- Physicians are more likely than APPs to report learning to use AI through self-taught methods: 46% reported learning through trial and error, versus 31% of APPs, who are more likely to have gotten institutional training (21% APPs vs. 16% physicians) or learned AI from a colleague (36% APPs vs. 30% physicians).
- Physicians show the largest difference in perceived AI value. 43% of physicians with less than 6 months of AI use report higher perceived value vs. cost; that jumps to 72% at two-plus years, a 29-point spread (vs. 23 points for APPs).
- 46% of physicians report concern about erosion of the patient-physician relationship due to patient AI use, versus 38% of APPs. This is consistent with physicians carrying more of the direct relationship and liability exposure in most care models.
APPs:
- APPs respondents reported more frequent access to institutional support, with AI more likely to be provided and paid for by their employer (34% vs. 30% physicians), and institutional training more common.
- APPs lean on OpenEvidence and ChatGPT at rates close to physicians (60% and 51%, respectively) — but use tools like Claude and Gemini less often.
- APPs’ confidence in AI forms earlier and holds flatter. They start higher (46% report higher value vs. cost with less than six months of use, vs. 43%) and end slightly lower (69% at two-plus years of AI experience).
- APPs report greater concern than physicians across several patient-facing AI risks — 84% worry about patient self-diagnosis or self-medication (vs. 78% of physicians), and 49% worry about delayed care from false AI reassurance (vs. 43% of physicians).
Which of the following AI tools do you currently use?
| Physicians | 66% | 55% | 42% | 25% | 22% | 15% | 14% | 11% | 6% | 3% | 1% |
|---|
- OpenEvidence
- ChatGPT
- EHR-embedded AI
- Gemini
- Claude
- Abridge
- Other AI tool
- Nuance DAX
- Perplexity
- Nabla
- Do not use AI tools
What are your biggest concerns about patients using consumer AI tools to interpret their own health data? (multi-select answer)
| Physicians | 78% | 77% | 75% | 51% | 43% | 46% | 28% | 2% | 1% |
|---|
- Self-diagnosis / self-medication
- Patient anxiety
- Inaccurate / misleading information
- Time correcting misinformation
- Delayed care after false reassurance
- Erosion of physician relationship
- Privacy / data security
- No concerns
- Other
Which of the following AI tools do you currently use?
| APPs | 60% | 51% | 41% | 13% | 11% | 10% | 17% | 11% | 2% | 1% | 0% |
|---|
- OpenEvidence
- ChatGPT
- EHR-embedded AI
- Gemini
- Claude
- Abridge
- Other AI tool
- Nuance DAX
- Perplexity
- Nabla
- Do not use AI tools
What are your biggest concerns about patients using consumer AI tools to interpret their own health data? (multi-select answer)
| APPs | 84% | 79% | 75% | 50% | 49% | 38% | 23% | 2% | 0% |
|---|
- Self-diagnosis / self-medication
- Patient anxiety
- Inaccurate / misleading information
- Time correcting misinformation
- Delayed care after false reassurance
- Erosion of physician relationship
- Privacy / data security
- No concerns
- Other
“I’m unsure at this point whether [the outcomes from AI] will be positive or negative. We just initiated AI in our office, and I would like to give it six months before I form an opinion on long-term concerns. I am concerned about the resources needed to keep AI data centers going, and a lot of communities are against these data centers coming into their communities with concerns such as noise, pollution, and harm to the environment.”
AI Tools
Which AI tools are clinicians using, and how?
HCPs aren’t picking one AI tool—most use several, and which ones depend heavily on their work environment.
Leading AI tool use differs by practice setting
ChatGPT use ranges from 49% among hospital and health-system respondents to 60% in independent private practice. Usage is 55% in group private practice and 57% in other practice settings.
OpenEvidence shows a wider range: 65% of respondents in hospitals and health systems use it, compared to 56% in independent private practice, 55% in group private practice, and 75% in the combined “other” group.
Leading tool use by practice setting
| ChatGPT | 49% | 55% | 60% | 57% |
|---|---|---|---|---|
| OpenEvidence | 65% | 55% | 56% | 75% |
- Employed by a hospital or health system
- Group private practice
- Independent private practice
- Other practice setting
AI access is uneven
16% of clinicians pay for AI tools completely out of pocket
Overall, 31% of respondents report employer-provided AI access, while 15% report bundled access through their EHR.
Employer-paid access is more common among hospital and health-system respondents than among those in independent private practice (38% versus 14%), and independent private-practice respondents are more likely to report paying out of pocket for AI tools (30% versus 9% in hospitals and health systems).
These findings suggest differences in AI access between employed and independent practice settings. For these reasons, clinicians should remember that comparing their own AI use with peers in different practice settings may be misleading.
How AI tools are accessed
| Hospital / health system | 38% | 16% | 9% | 36% | 1% |
|---|---|---|---|---|---|
| Independent / private | 14% | 13% | 30% | 41% | 2% |
- Provided or paid for by my employer or institution
- Bundled into my EHR at no additional cost
- I pay for tools out of pocket
- A combination of the above
- I do not currently use AI tools
Most HCPs taught themselves to use AI
41% of clinicians reported learning to use AI through their own trial and error, while 32% learned through a colleague. Formal onboarding is less common, even where it is most available: 23% of hospital and health-system respondents report institutional training, compared with just 9% in independent private practice.
This data points to a need for practical AI onboarding/training that reaches independent clinicians and professionals working outside conventional employer-supported settings (including locum tenens clinicians).
How did you primarily learn to use AI tools in your clinical practice?
| 41% | 32% | 18% | 7% | 1% |
- Self-directed (trial and error, YouTube, online resources)
- Peer recommendation or informal discussion with colleagues
- Institutional training or onboarding
- I have not received any training
- CME or accredited program
Liability concerns are changing behavior for only a slim majority
Liability concerns are shaping AI use for many healthcare professionals, but they are far from universal. These concerns hold steady across practice settings, challenging the assumption that clinicians working within larger health systems feel significantly more protected from legal or regulatory consequences than their independently practicing peers. The finding suggests that while accountability questions remain front of mind, nearly as many HCPs have not yet changed their workflows in response to these risks. As AI adoption continues to expand, clear regulations, institutional guidance, and documentation best practices will likely play a critical role in helping clinicians use these tools with confidence.
52% of HCPs say liability worries have changed how they use or document AI.
Patients and AI
How AI affects patient safety and consultations
Most clinicians use AI in some form in their work, and the same holds true on the patient side, with many patients bringing AI-generated information into their consultations.
Most clinicians are concerned about patient AI use
Survey respondents were asked about patients using consumer AI to interpret their health information. Only 2% of respondents reported no concern about patients’ use of health-related AI information, leaving 98% with at least one specific worry.
Some of the key concerns cited by clinicians include:
- Self-diagnosis or self-medication (80%)
- AI-driven anxiety (78%)
- Inaccurate information (75%)
- Increased consultation time (51%)
- Delayed care (46%)
What are your biggest concerns about patients using consumer AI tools to interpret their own health data?
| Patients self-diagnosing or self-medicating based on AI outputs | 80% |
|---|---|
| Patient anxiety driven by misinterpreted results | 78% |
| Patients receiving inaccurate or misleading information | 75% |
| Increased consultation time spent correcting AI-generated misinformation | 51% |
| Patients delayed care or ignored symptoms because AI reassured them it was not urgent | 46% |
| Erosion of the patient-physician relationship | 43% |
| Privacy and data security risks | 26% |
| I do not have concerns about this | 2% |
- Patients self-diagnosing or self-medicating based on AI outputs (80%)
- Patient anxiety driven by misinterpreted results (78%)
- Patients receiving inaccurate or misleading information (75%)
- Increased consultation time spent correcting AI-generated misinformation (51%)
- Patients delayed care or ignored symptoms because AI reassured them it was not urgent (46%)
- Erosion of the patient-physician relationship (43%)
- Privacy and data security risks (26%)
- I do not have concerns about this (2%)
AI has made consultations more challenging
56% of respondents say patient-connected AI tools (like ChatGPT Health) have not yet affected their consultations, or they haven’t encountered an issue. However, a significant percentage, 25%, report that AI has had a negative or challenging effect on consults, while only 9% report a positive effect. The recurring theme: Patients arrive having already reached a conclusion, and walking that back takes real time and trust.
“The major issue I have faced is that patients come in with an expectation of what I will be telling them and then are not pleased when what I say doesn’t line up 100% with AI. I wish there was a patient-facing, medically accurate AI tool like OpenEvidence, but for patients.”
Clinicians report identifying potentially harmful AI-generated errors
Respondents who use AI in practice were asked, in their own words, about errors or harmful outputs they had seen from AI. Sermo grouped the answers into categories. Of those who responded:
- 70% reported no harmful output
- 12% described identifying a potentially harmful AI-generated error
- 11% reported a concern or near-miss
Respondents described potential AI-generated errors involving diagnosis, documentation, and medication management, highlighting areas where inaccurate outputs could pose patient-safety risks.
“An H&P from an AI-generated output hallucinated a patient’s hospital course from an outside hospital, and this caused delays in determining if a patient actually had a certain surgery or not, which could have significantly changed anticoagulation management.”
“Patients often come with question lists generated by AI or having anchored on an irrelevant diagnosis generated by AI. It can be very difficult to dissuade them from this diagnosis, or the lack of need for extensive imaging and workup.”
future outlook
What clinicians expect from AI over the next 5 years
Just under half of surveyed HCPs (49%) expect AI to have a net positive effect on patient care over the next few years. However, 24% of respondents expect a mixed outcome from AI, and 15% expect a negative one.
Leading concerns cited by respondents include:
Accuracy and safety risks (13%)
Degradation of workforce and clinician relationships (12%)
Overreliance on AI (8%)
Leading benefits cited by respondents include:
Documentation relief (20%)
Greater efficiency (16%)
Decision support (12%)
“AI can provide general education for patients in the next five years, and as long as patients continue to confide in their doctors regarding diagnosis and management, these changes will prove to be positive.
However, with any introduction of new technology or policy, the benefits and risks will exist. The benefit-risk ratio will continue to be user-dependent based on the user’s mindset regarding AI and ability to perceive concepts face-value compared to comparing it with other sources.”
Conclusion
What Sermo survey findings mean for clinical teams
HCPs aren’t lacking caution with AI; they lack a shared training and documentation standard that matches how quickly adoption has already happened.
AI use is already common among clinicians, though the practices around it remain uneven.
- 9 in 10 respondents are already using it, and most are getting real value from it — with that value perception appearing to grow the longer they use it.
- Respondents differ in their AI practices, including how they access it, learn to use it, verify outputs, and document it.
- Patient questions are often shaped by consumer AI interfaces, sometimes requiring HCP vigilance to correct misinformation.
- Findings suggest inconsistent training, verification, and documentation practices among respondents.
- HCPs are catching AI mistakes before they affect patients through widespread spot-checking, sharp awareness of where AI can mislead, and real caution around liability.
The findings point to three practical areas for teams to examine:
1. Training: Give clinicians a way to learn appropriate uses, limits, and common failure modes that does not depend entirely on trial and error. One important takeaway is that the standard should reach independent clinicians and professionals in less institutionally supported roles as reliably as it reaches those in hospital and health system settings.
2. Verification: Define who reviews AI-assisted material and what level of verification is expected before it informs a note, order, or patient communication.
3. Documentation: Clarify when and how AI assistance should be recorded within the organization’s own policies and applicable requirements.
Survey findings
Methodology
This report draws on a survey of 3,067 U.S. healthcare professionals on Sermo, a verified peer community for HCPs. The data does not reflect a comprehensive census of all U.S. healthcare professionals.
Respondents were grouped into three provider types:
MD/DO physicians, Advanced practice providers (nurse practitioners and physician assistants) and other healthcare professionals (including midwives, nurses and podiatrists). Unmatched specialty respondents were excluded from provider-type comparisons.
Physician and APP subgroup findings above are drawn from segment-level crosstabs (practice setting and AI-use tenure). The “other practice setting” group combines academic/teaching institutions, government/VA settings, locum tenens, and other practice-setting responses.
Respondent practice setting segments
| 1,446 |
| 535 |
| 583 |
| 503 |
- Employed by a hospital or health system (1,446)
- Independent private practice (535)
- Group private practice (583)
- Other practice setting (503)
Results are descriptive, not causal. Differences between respondent groups have not been tested for statistical significance and should be interpreted as directional rather than conclusive. Responses are confidential and reported in aggregate. Findings reflect self-reported AI use, verification habits, training, and perceived patient impact at the time of fieldwork. Question bases vary: some questions were shown only to qualifying respondents (such as current AI users), and multi-select questions may total more than 100%. The sample size (n) for each result appears in the relevant chart.
Survey Fielding Dates: June 22 – September 11, 2026
Sample size: 3,067
Demographics: Healthcare professionals across all U.S. regions
Respondent provider type segments
| 1,820 |
| 1,022 |
| 91 |
| 134 |
- MD/ DO physician (1,820)
- APP (1,022)
- Other HCP (91)
- Unmatched specialty (134)
Frequently asked questions
AI in practice FAQs
In this survey of 3,067 U.S. Sermo members, 92% reported using AI tools in clinical practice.
OpenEvidence (63%) and ChatGPT (53%) are the most commonly used tools overall. Use and access to AI tools varies by role and practice setting, and respondents may use more than one tool.
94% of respondents in this survey who use AI report applying at least some scrutiny to AI output, with 34% of that segment reporting that they always check it against a second source.
Respondents most often cited concerns about self-diagnosis or self-medication (80%), AI-driven anxiety (78%) and inaccurate information (75%) when patients use consumer AI to interpret health information (note: respondents could select more than one answer). 51% said correcting misinformation from patient AI use has added time to consultations.
70% of the 2,833 respondents who use AI described no harmful output from its use. 12% described an error that could have harmed a patient, and 11% described a near-miss or concern. Many of the examples involved diagnosis, documentation or medication management, and several clinicians described that they had caught the problem by reviewing the output before acting on it. These are self-reported accounts that Sermo has not independently verified and may not reflect a universal experience.