Online survey assessing US primary care physicians’ attitudes toward AI use in clinical administrative tasks

Abstract

Objectives

To examine primary care physicians’ attitudes regarding artificial intelligence (AI) use for administrative clinical tasks.

Methods

Web-based survey with US physicians in family medicine or internal medicine (N=420, response rate 5.13%). Two hypothetical AI tools for administrative clinical activities were described. We examined physicians’ attitudes towards AI tools, and their associations with practice years, exposure to AI, use case and stakeholder type were evaluated using generalised estimating equations.

Results

Participants were on average 49.6 years (SD=12.5) and 56.7% men (238/420). Physicians with fewer practice years were more likely to endorse the tools’ benefits (OR 1.70–1.96), the tools’ benefits outweighing risks (OR 1.79–2.06) and their openness to use (OR 1.63–1.83), and were less likely to endorse disclosure of AI use (OR 0.60 (95% CI 0.36 to 0.998)). Physicians with AI exposure were more likely to agree the tools’ benefits outweighed their risks (OR 1.51 (95% CI 1.06 to 2.16)). Physicians were more likely to endorse the tools’ benefit to physicians (OR 4.94 (95% CI 4.16 to 5.86)) and physicians’ openness to using them (OR 3.53 (95% CI 2.97 to 4.20)) than they were to endorse their benefit to patients and patients’ openness. Physicians rated an AI tool for notes generation as more beneficial than one for billing assistance (OR 1.73 (95% CI 1.39 to 2.16)).

Discussion

Although the findings are preliminary, US primary care physicians’ attitudes toward AI for clinical administration varied by practice years, prior exposure to AI, use case and stakeholder type.

Conclusion

Our findings highlight opportunities to develop training and implementation strategies in service of advancing safe and effective integration of administrative AI tools in primary care.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Administrative workloads in primary care can interfere with the more consequential aspects of patient care, negatively impact physicians’ well-being and contribute to physician burnout and attrition. Artificial intelligence (AI)-enabled solutions that could reduce administrative workloads are promising, yet physician attitudes toward such applications have been minimally explored.

WHAT THIS STUDY ADDS

• This study examined the attitudes of US primary care physicians toward AI use cases that support administrative clinical tasks (ie, a billing assistance tool and a clinical notes generation tool), finding that physicians had different attitudes relating to perceived benefits, openness, risks and disclosure depending on professional factors such as their number of years of practice and prior exposure to AI, and contextual factors such as stakeholder type and the use case.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Enhancing educational opportunities about AI among physicians with varying levels of professional experience and prior AI exposure may contribute to greater heterogeneity in their assessments relating to the benefits and risks of using AI tools, as well as the appropriate procedures regarding disclosure of AI use to patients.

Introduction

Artificial intelligence (AI) is expected to become increasingly integrated into clinical practice, allowing physicians to focus on the quality of patient interactions and experience fewer administrative burdens.^{1 2} Exciting advances, particularly in generative AI, have the potential to carry out complex and ‘high stakes’ tasks previously not imagined as amenable to AI.^{2 3} Despite major technological developments, it is increasingly recognised that AI’s success in clinical care rests heavily on physicians’ attitudes toward such tools.4,7

A rapidly growing body of empirical work regarding physicians’ attitudes about AI tools has focused primarily on diagnostic and health management use cases or on the prospect of AI as a technological direction in medicine. This literature has documented reservations and uncertainties among physicians about the impact that AI tools may have on their workloads and health systems,⁸ and concerns related to potential risks around privacy, medical error and the therapeutic relationship.⁹ Broader concerns include physicians’ perceptions of how much benefit clinical AI applications would bring to patients and whether they would optimally support clinical tasks and processes, for example, diagnosis.¹⁰

There is a striking gap in empirical attention regarding physicians’ attitudes toward AI use cases for routine administrative clinical tasks. Especially in primary care, heavy administrative workloads can interfere with the more urgent, complex or consequential aspects of patient care; negatively impact physicians’ well-being¹¹; and contribute to broader workforce concerns including burnout,¹² turnover and attrition.¹³ AI-enabled solutions to these problems may include monitoring private patient–physician interactions; assisting with note-taking,¹⁴ billing tasks and related documentation¹⁵; or drafting responses to patients on remote communication platforms.¹ Such tools also present potential risks and challenges, including issues relating to physician liability, patient privacy and confidentiality, bias and their integration into clinical environments begets the need for additional work related to physician training and ongoing oversight and management.^{16 17} Physician attitudes toward such applications have been minimally explored, but may encompass, for example, the anticipated risks and benefits of using these tools, physicians’ willingness to implement them and preferences surrounding disclosure.^{9 18}

For this study, we aimed to examine the attitudes of primary care physicians towards AI use cases that support administrative clinical tasks (ie, a billing assistance tool and a clinical notes generation tool). US physicians with specialties in family or internal medicine were randomly sampled from the American Medical Association (AMA) Physician Professional Data file. Through an online survey which described two hypothetical AI tools, we collected physicians’ 5-point Likert-scale attitudes regarding: the anticipated benefit of the AI tool for both physicians and patients; the expected openness of physicians and patients toward clinical use of the tool; the importance of disclosure of the tool’s use; and the tool’s risk–benefit balance. We also examined associations between physicians’ attitudes and their years of practice, their prior exposure to AI, the AI use case and the stakeholder type (ie, whether physicians’ perceptions were regarding physicians or patients).

Methods

Participants

Participants in this online survey study were US physicians in family medicine or internal medicine. We contracted with Medical Marketing Services for the purpose of identifying a random sample of 10 000 active physicians from the AMA Physician Professional Data file with a specialty of family medicine or internal medicine.¹⁹ Between 16 May and 20 December 2023, up to four emails (one initial email invitation and up to three reminders) and three mailed letters (one initial invitation and up to two reminders) were sent to the 10 000 physicians identified. Invitations contained a brief description of our study, a weblink to the online survey and unique respondent codes. Physician participants completed an online informed consent form prior to the start of the survey and were provided with a US$10 Amazon.com gift code at the time of survey completion.

Study design

This study belongs to a larger online survey study intended to assess physicians’ attitudes and decision-making in the context of clinical AI.²⁰ The current manuscript focuses on physician’s attitudes towards administrative applications of AI in healthcare and includes our analysis of four measures collecting participants’ demographic information and professional experience, familiarity and professional experience with AI, general attitudes towards the use of AI in medicine and attitudes towards hypothetical AI tools that support clinical administration. A second part which tested a separate research question included a randomised vignette trial assessing physicians’ decision-making in the context of clinical AI in non-administrative clinical contexts and is thus beyond the scope of the current manuscript.

Tool description and survey questions

Two hypothetical AI tools for supporting routine administrative clinical tasks were described in the survey (table 1). The following six questions were asked for each hypothetical AI tool; participants rated their agreement with each statement on a 5-point Likert scale (1=Strongly disagree, 3=Neither agree nor disagree, 5=Strongly agree): (Q1) This tool would benefit physicians; (Q2) This tool would benefit patients; (Q3) In general, physicians would be open to using this tool in their practice; (Q4) In general, patients would be open to having this tool used in their care; (Q5) Physicians should disclose their use of this tool to patients; and (Q6) The benefits to using this tool would outweigh the risks.

Table 1. Hypothetical AI tools described in survey.

	‘Billing assistance’ tool	‘Notes generation’ tool
Use case	AI tool that identifies billing codes.	AI tool that generates clinician notes.
AI description	The AI tool described below is designed to provide clinicians with a decision or a score relating to a specific task. By analysing a large data set, tools such as this one learn patterns which they can then use to make predictions. However, their predictions can sometimes be less accurate when used on new data, and due to the nature of the models used for these tools, users may not be able to verify the rationale behind the tools’ predictions.	The tool described below is a generalist AI tool, which is designed to complete a wide range of tasks simultaneously. By analysing many large and diverse data sets, tools such as this one have the ability to generate new and original content (including audio, video, images, text and simulations), and respond to user questions or requests. These tools can sometimes generate content that seems plausible to users, but includes factual errors or errors in reasoning, and due to the nature of the models used for these tools, users may not be able to verify the rationale or evidence behind the tools’ output.
AI tool description	This tool uses AI to assist clinical teams with the task of identifying billable codes from clinician notes. At the end of a patient encounter, this tool analyses the content of the clinician’s electronic notes and identifies potentially related billing codes, which it then suggests to the clinical team for approval.	This tool uses generalist AI to generate complete clinician reports from patient encounters. During a patient’s visit with a clinician, the tool tracks the clinician–patient conversation in real time, while simultaneously analysing and incorporating information from the patient’s electronic medical record. It then uses this information to generate written electronic notes, discharge reports and treatment recommendations which are appended to the electronic medical record after clinician approval. Clinicians can also make edits or additions to the notes by chatting with the tool.

AI, artificial intelligence.

Outcomes

Outcomes were participant agreement with the six statements listed above. The 5-point Likert scale ratings for each statement were dichotomised into two categories: (1) Agreement (Strongly agree and agree, ie, a 4 or 5); and (2) Non-agreement (Strongly disagree, disagree and neither agree nor disagree, ie, a 3 or below). The dichotomous ratings were used to calculate the proportion of agreement for each statement. In addition, dichotomous ratings for the statements regarding benefit for using the clinical AI (Q1 and Q2) and openness to using the clinical AI (Q3 and Q4) were stacked respectively, so that we could aggregate data for participants’ perceptions for physicians and patients for each relevant statement and explore the effects of the stakeholder type (participants’ perceptions relating to physicians vs patients) on the outcomes.

Independent variables examined

Practice years were broken down into quartiles: 8 years or less, 9–17 years, 18–26 years and 27 years or more (reference group). Exposure to AI was defined as a dichotomous ‘yes’ or ‘no’, with ‘no’ being the reference group: ‘yes’ if the respondents reported familiarity with AI (ie, a rating of 4 or 5 on a 5-point scale), or reported prior use of AI in medical research or in clinical practice; and ‘no’ if none of the three dimensions assessed were evaluated as true. The use case was treated as a dichotomous variable (billing assistance AI tool or notes generation AI tool) with the billing assistance AI tool as the reference group. The stakeholder type (participant perceptions of benefit for and openness to AI use for physicians vs patients) was also treated as dichotomous with the patients as the reference group.

Statistical analysis

Proportions of participant agreements with the six statements were calculated and summarised based on the re-coded dichotomous ratings as described above and were grouped and plotted by AI use case. To evaluate our aim, we used generalised estimating equations, marginal models for longitudinal data analysis, with an exchangeable correlation structure and logit link function, where an outcome was each statement and independent variables were participants’ practice years, participants’ exposure to AI and AI use case. The statements about physicians and patients (ie, Q1 and Q2; Q3 and Q4) used parallel language and were combined with a variable for the referenced stakeholders (ie, ‘stakeholder type’) for the purpose of analysis. ORs and their 95% CIs were computed for each independent variable based on the estimates from the models. The results represent adjusted analyses, assuming the other factors being equal.

The significance level was 0.05 and all tests were two-sided. All statistical analyses were conducted with R software (R V.4.3.1, GNU project).

Results

Of 10 002 US physicians in family medicine or internal medicine who were invited to participate in this online survey, 513 submitted a response (response rate, 5.13%), of which 445 were a complete response. Of the complete responses, 25 were excluded from the data set for not having an MD or DO or the appropriate specialty. The final cohort included 420 physicians from 48 US states. Complete characteristics of the sample can be found in the online supplemental table.

Overall trends

Overall, participants felt that both described AI tools would benefit physicians and that physicians would be open to using the tools in their practice (65–81.2% positive responses across tools). They expressed neutral to negative attitudes regarding the benefit of the AI tools to patients and patients’ openness to having the tools used in their care for both tools (26.7–57.4% positive responses across tools) (see figure 1). A majority of participants expressed agreement that the use of the AI tool should be disclosed to patients for the notes generation AI tool (82.9%), while only about half of participants agreed with this for the billing assistant AI tool (52.6%). Slightly more than half of the participants agreed that the benefits to using the AI tools would outweigh the risks for both tools (53.8–56.4%).

Figure 1. Participants’ agreements with questions by use case. Note: the 5-point Likert scale ratings for each statement were dichotomised into two categories: (1) Agreement (Strongly agree and agree, ie, a 4 or 5); and (2) Non-agreement (Strongly disagree, disagree and neither agree nor disagree, ie, a 3 or below). The proportions were calculated based on the dichotomised ratings. AI, artificial intelligence.

Associations

Practice years: The number of practice years was positively associated with agreement that physicians should disclose their use of the AI tool to patients (≤8 years vs ≥27 years: OR 0.60 (95% CI 0.36 to 0.998)).

The duration of professional practice was inversely associated with agreement that ‘AI tools would benefit physicians and patients’ (≤8 years vs ≥27 years: OR 1.70 (95% CI 1.15 to 2.51), 9–17 years vs ≥27 years: OR 1.96 (95% CI 1.30 to 2.96)) and agreement that ‘physicians and patients are open to the use of the AI tools’ (≤8 years vs ≥27 years: OR 1.83 (95% CI 1.26 to 2.66), 9–17 years vs ≥27 years: OR 1.63 (95% CI 1.11 to 2.40)), and ‘benefits to using the AI tools would outweigh the risks’ (≤8 years vs ≥27 years: OR 2.06 (95% CI 1.34 to 3.18), 9–17 years vs ≥27 years: OR 1.79 (95% CI 1.13 to 2.83)) (see table 2).

Table 2. Associations between participants’ agreements with the statements and practice years, exposure to AI, AI use case and the stakeholder type (physicians vs patients).

Statement	Independent variable	OR (95% CI)	P value
This tool would benefit (physicians or patients)	Practice years (years)
	≤8 vs ≥27	1.70 (1.15 to 2.51)	0.008*
	9–17 vs ≥27	1.96 (1.30 to 2.96)	0.001*
	18–26 vs ≥27	1.31 (0.86 to 1.97)	0.21
	Exposure (yes vs no)	1.34 (0.98 to 1.84)	0.071
	Use case (notes vs billing)	1.73 (1.39 to 2.16)	<0.0001*
	Stakeholder type (physicians vs patients)	4.94 (4.16 to 5.86)	<0.0001*
In general, (physicians or patients) would be open to the tool use*	Practice years (years)
	≤8 vs ≥27	1.83 (1.26 to 2.66)	0.001*
	9–17 vs ≥27	1.63 (1.11 to 2.40)	0.014*
	18–26 vs ≥27	1.30 (0.89 to 1.92)	0.18
	Exposure (yes vs no)	1.25 (0.92 to 1.70)	0.157
	Use case (notes vs billing)	0.99 (0.81 to 1.21)	0.917
	Stakeholder type (physicians vs patients)	3.53 (2.97 to 4.20)	<0.0001*
Physicians should disclose their use of this tool to patients	Practice years (years)
	≤8 vs ≥27	0.60 (0.36 to 0.998)	0.049*
	9–17 vs ≥27	0.72 (0.43 to 1.22)	0.23
	18–26 vs ≥27	0.89 (0.52 to 1.51)	0.66
	Exposure (yes vs no)	1.06 (0.72 to 1.56)	0.752
	Use case (notes vs billing)	4.41 (3.37 to 5.78)	<0.0001*
The benefits to using this tool would outweigh the risks	Practice years (years)
	≤8 vs ≥27	2.06 (1.34 to 3.18)	0.001*
	9–17 vs ≥27	1.79 (1.13 to 2.83)	0.014*
	18–26 vs ≥27	1.37 (0.87 to 2.16)	0.17
	Exposure (yes vs no)	1.51 (1.06 to 2.16)	0.023*
	Use case (notes vs billing)	0.90 (0.71 to 1.13)	0.364

Note: OR of 1 indicates the null value; if the 95% CI contains the null value, there is no statistical difference between the two groups. An OR (and the upper bound of the 95% CI) of less than 1 indicates an increased odds of agreement for the reference group. An OR (and the lower bound of the 95% CI) greater than 1 indicates an increased odds of agreement for the comparison group (non-reference group).

^*The actual statements were written as follows: in general, physicians would be open to using this tool in their practice. In general, patients would be open to having this tool used in their care.

AI, artificial intelligence.

Exposure to AI: Prior exposure to AI was positively associated with agreement that ‘the benefits to using the tool would outweigh the risks’ (OR 1.51 (95% CI 1.06 to 2.16)).

AI use case: Participants viewed that an AI tool for notes generation would benefit physicians and patients more than an AI tool for billing assistance (OR 1.73 (95% CI 1.39 to 2.16)). Participants were more likely to agree that physicians should disclose the use of AI when used for generating notes than for billing assistance (OR 4.41 (95% CI 3.37 to 5.78)).

Stakeholder type: Participants agreed that the two AI use cases would benefit physicians more than patients (OR 4.94 (95% CI 4.16 to 5.86)), and that physicians would be open to using the AI tools in clinical care more so than patients would (OR 3.53 (95% CI 2.97 to 4.20)).

Discussion

Innovative AI tools developed to support tasks such as clinical note-taking and billing present a potential opportunity to reduce the administration burden of primary care physicians, allowing for greater focus on patient care.^{1 21} This survey study is among the first to reveal insight into the associations between physicians’ attitudes toward hypothetical AI tools for administrative clinical tasks (in terms of anticipated benefit, openness to their use, disclosure preferences and risk-benefit balance), and how these attitudes may vary by total practice years, prior exposure to AI, use case and stakeholder type.

The effect of practice years on physician attitudes toward AI

This study notably found that the number of practice years was associated with physician attitudes across all questions asked. Participants with the most years of practice assigned greater importance to disclosure, while those with fewer practice years expressed more agreement in regards to the benefit of and their openness to using AI and agreement that the benefits of the AI tools outweigh the risks. Taken together, these findings indicate that physicians with fewer practice years may be more trusting of AI-enabled tools than their counterparts with more years of practice. Similarly, a prior study found that intention to adopt AI increased among physicians with fewer years since medical graduation,²² while another found higher age has an inverse association with trust in AI among doctors.²³ Our findings are consistent with these earlier findings, suggesting that these trends persist even in clinical scenarios where the use of AI is purely administrative and physician-facing. Future work testing over-reliance or under-reliance on clinical AI based on age or years of practice may be an important next step in ensuring safe implementation and use of clinical AI.

The effect of AI exposure on physician rating of risk-benefit

The associations found between participants’ prior exposure to AI and their ratings of risk and benefit suggest that while physicians with and without AI exposure may report similar levels of anticipated benefit and openness toward the use of clinical AI, those without prior AI exposure via education or professional experience may associate greater risk with the use of such tools. This appears to be a novel finding. To be ethical users of clinical AI, physicians minimally will need to feel confident in their ability to assess and communicate the potential limitations, biases and safety considerations of these tools and to be able to distinguish among the varying levels of risk and benefit presented by individual tools.²⁴ Given the crucial role of physicians’ risk and benefit assessments in providing quality care, future work may seek to assess how training or exposure to AI may account for heterogeneity in these assessments across the physician population. In particular, different didactic approaches may vary in their ability to promote safe and effective adoption. Hands-on educational interventions that allow clinicians to directly test AI tools, case-based learning that simulates realistic clinical scenarios involving both appropriate use and misuse or suboptimal use, and forward-looking continuing medical education programmes that integrate emerging AI capabilities and anticipated regulatory changes may each help build practical proficiency and increase trust.

Physician perspectives regarding the benefit and use of clinical AI for physicians versus patients

The findings regarding physicians’ agreement that physicians, compared with patients, would benefit more from and be more open to AI tools suggest that more empirical work testing the alignment between stakeholder views may be valuable, especially considering that patients and members of the broader public have reported generally positive attitudes toward the incorporation of AI into clinical care in prior work.925,28 In prior work exploring a similar question, our team documented patterns of misalignment in the research-related attitudes of patient participants and physician researchers: physician researchers underestimated the positive attitudes that patients had toward medical research, as well as underestimated the extent to which their beliefs regarding patients’ views diverged from patients’ actual views.²⁹ Identification of potentially incorrect assumptions or predictions made by stakeholders about other stakeholders’ views regarding AI integration may help generate guidance on how to better align expectations and enable decision-making that is based on a sense of trust, mutual communication and understanding. Future work may build on current evidence by collecting perspectives regarding AI tools from both physician and patient populations in order to make intergroup comparisons.

Limitations

Limitations of this study include the low response rate of the physicians who were sampled. While the response rate was low, it was expected due to similar response rates reported in other studies using the same database and similar recruitment strategies.³⁰ This sample of US physicians was from primary care and family medicine specialties, thus the generalisability of our findings may be limited to such populations. Our findings are necessarily preliminary, given that the two use cases are prospective and not yet widely implemented in clinical practice. In addition, given that this was a cross-sectional analysis, we were not able to make inferences regarding the causal impact of specific aspects of the tools on the examined attitudes.

Conclusions

Administrative AI tools may soon be widely adopted in primary care settings, allowing for a shift in attention and resources from administrative tasks to patient care, and potentially reducing burnout and other negative consequences on the healthcare workforce attributable to administrative burden. Physicians’ attitudes toward such tools, including perceived benefit and openness, as well as risks or disclosure needs, may present both advantages and barriers during clinical integration and appear to be affected by individual factors such as physicians’ years of practice and prior exposure to AI, and contextual factors such as stakeholder type and AI use case. Understanding how these factors influence stakeholder attitudes or expectations regarding AI sheds light on potential paths forward to support AI integration in primary care.

Building a culture of responsible AI adoption will require sustained efforts in developing interdisciplinary education that bridges ethics, data science and domain expertise. Targeted training initiatives should equip clinicians, researchers and trainees with both the technical competencies and ethical reasoning needed for safe and effective AI use. Programmes that address model interpretation, bias mitigation and workflow integration can reinforce informed and critical engagement with AI tools. By embedding these efforts within medical education, professional development and institutional learning structures, the field can cultivate a workforce that appropriately uses AI, ensuring that AI innovations advance patient care while maintaining public trust.

Data availability statement

Data are available upon reasonable request.