Artificial Intelligence Scribes in Psychiatry

Abstract

Artificial intelligence (AI) scribes for clinical documentation are likely to be among the first AI tools to affect the day-to-day practice of psychiatry, yet many psychiatrists are unfamiliar with them. This article introduces psychiatrists to AI scribes, including their potential benefits and risks. AI scribes may enhance efficiency, reduce physician burnout, and improve patient–physician rapport by automating documentation processes. Key concerns include privacy, accuracy, and the ability to capture nuanced patient communication. Initial recommendations for responsibly integrating AI scribes into psychiatric practice are provided, emphasizing the importance of balancing technological benefits with careful oversight.

Case

Dr. S is a psychiatrist at a local health system, where she has a busy general outpatient clinic and covers the inpatient unit and emergency room every other month while on call. Although she provides excellent care, she is increasingly frustrated by the burden of clinical documentation. She worries that she spends more time looking at her computer screen than at her patient during new outpatient evaluations, thus negatively affecting her rapport with patients. She often spends an hour at home each night after clinic catching up on notes, which, she suspects, is contributing to a growing sense of burnout. Because she only covers the inpatient service several times per year, she is less familiar with their electronic medical record and struggles to document efficiently.

When Dr. S raised these concerns during her annual performance review, her chair mentioned that she just came back from a meeting where a presenter demonstrated the use of an artificial intelligence (AI) scribe in a primary care clinic. The chair asked Dr. S to investigate whether AI scribes may be helpful for addressing her concerns about the burden of clinical documentation and useful to other psychiatrists in the department. Although she has a limited knowledge of AI tools, Dr. S is intrigued by this suggestion. She has many questions: Will these tools work in psychiatry? How much time could they save me? Will patients feel comfortable with me using this technology? Are they cost-effective? What about safety and privacy concerns? With these questions in mind, Dr. S began researching the potential benefits and risks of adopting this technology in her department.

AI shows enormous potential to revolutionize all aspects of psychiatry, including clinical care, patient engagement, education, research, and administration. Although the potential applications of AI to psychiatry were first explored in the 1960s (1, 2), recent advances in AI have allowed for many novel uses, including therapy chatbots, educational resources, and clinical support tools (3–5). AI scribes for clinical documentation are likely to be among the first AI tools to affect the day-to-day practice of psychiatry, yet many psychiatrists are unfamiliar with them.

This article introduces practicing psychiatrists to AI scribes for clinical documentation. We begin by briefly describing this technology’s foundations before considering potential benefits and risks of its use. We then offer preliminary recommendations for responsible use of AI scribes. In this article, we also consider several unique aspects of applying AI scribes to psychiatric care.

What Is AI?

First coined in the 1950s, the term AI (or artificial intelligence) refers to computer systems designed to perform tasks that typically require human intelligence, such as reasoning, language comprehension, and decision making. Over the past 70 years, advancements in AI have led to widespread adoption and increased utility. AI enables many familiar technologies, such as facial recognition on smartphones, targeted advertisements on websites, and automatic spam e-mail filtering.

Recent advances in generative AI have reignited interest in the potential applications of AI. Generative AI, capable of creating new content such as text or images, powers popular chatbots such as OpenAI’s ChatGPT and Google’s Gemini. The rapid adoption of these tools has prompted discussion about the application of generative AI to medical practice, education, and research. In this article, we explore a single application of generative AI to clinical medicine—the AI scribe—because it is likely to be among the first AI tools used by psychiatrists in routine clinical practice.

As AI terminology may be unfamiliar to many psychiatrists, a glossary of terms used in this article is provided in Box 1.

BOX 1. Glossary of terms

AI (Artificial Intelligence)

Computer systems designed to perform tasks that typically require human intelligence, such as reasoning, decision making, and problem solving.

AI Scribe

A generative AI tool that automatically generates clinical documentation from patient encounters.

Generative AI

AI systems that are capable of creating new content, such as text, images, or music, by learning patterns from existing data.

Hallucination

A plausible but false response generated by an AI model.

What Are AI Scribes?

AI scribes are tools that automatically generate clinical documentation from patient encounters. Because there are various commercially available AI scribes, in this article, we focus on general principles that apply across tools.

The process for using AI scribes is outlined in Figure 1. After obtaining consent from the patient, the psychiatrist activates the AI scribe on a smartphone or computer and proceeds with the interview as usual. Behind the scenes, AI scribes begin by converting audio recordings of the encounter into a written transcript using speech recognition algorithms. This process is similar to dictation software, which is already used by many psychiatrists. The AI scribe then summarizes the transcript’s content and outputs a structured note organized in the typical format (i.e., subjective, objective, assessment, plan). Because AI scribes are capable of making errors, psychiatrists must review the documentation for accuracy and completeness and make necessary edits before signing their note.

FIGURE 1.

Workflow for using an AI scribe^a

^aAI, artificial intelligence.

Several AI scribes are being integrated into electronic health records (EHRs) to streamline workflow and enhance efficiency. Many AI scribes include customization options, such as selecting between narrative or bulleted documentation styles, to fit user preferences and clinical needs.

The Coming Wave

AI scribes are rapidly improving in their accuracy and usability and are being increasingly incorporated into clinical practice. A 2023 systematic review did not identify any large-scale clinical studies of commercially available AI scribes (6). Since then, there have been several publications describing pilots of AI scribes in clinical settings.

In March 2024, Tierney and colleagues (7) published the results of a pilot study at the Permanente Medical Group where an AI scribe was used by 3,442 physicians in different specialties in over 300,000 patient encounters. The authors reported favorable response to AI scribes from both physicians and patients, high note quality, and some evidence of improved clinical efficiency. The results of several smaller pilot studies of AI scribes are described later in this article when we discuss the potential risks and benefits of these tools.

Although there are currently no published empirical studies of AI scribes in psychiatry, many psychiatrists have already adopted these tools and described their experiences in online groups and forums. A 2019 global survey of 791 psychiatrists revealed that 75% of respondents believed that technology would eventually replace humans for clinical documentation, with 47% expecting this to occur within the next 4 years (8). Similarly, a 2023 survey of 138 psychiatrists who attended an American Psychiatric Association (APA) information session on AI found that 70% of respondents “somewhat agreed/agreed” that “documentation will be/is more efficient” from using generative AI tools (9). Psychiatrists must, therefore, be prepared to consider the potential benefits and drawbacks of this technology before deciding whether to adopt it in their own practice.

Potential Benefits

The clearest potential benefit of AI scribes is increased efficiency in clinical documentation, which takes up a significant portion of physician time. A time-and-motion study of 57 physicians across multiple specialties found that, during a clinic day, physicians spent 27% of their time on direct clinical face time and 49% of their time on EHR and desk work (10). Early results of AI scribe implementation studies show promise for improving documentation efficiency. For example, Cao and colleagues (11) found that AI scribes allowed a group of 12 dermatologists and physician assistants to reduce total daily time in EHRs from 90 min to 70 min. In a different study, eight clinicians from various medical specialties and allied health professions participated in simulated patient consultations, both with and without the assistance of an AI scribe. Documentation completed with the help of an AI scribe took 26% less time to complete on average and scored higher on documentation quality with the use of a validated assessment tool (12).

Improved efficiency in clinical documentation may also lead to improved physician satisfaction and reduced burnout. In a survey of 1,792 practicing physicians across specialties, Gardner and colleagues (13) found that over 46% of respondents felt that they had insufficient time for documentation and that 38% of respondents reported moderately high or excessive time spent on the EHR at home, both of which were associated with higher rates of burnout. Similarly, a 2019 survey of 1,524 outpatient physicians by Gaffney and colleagues (14) found that respondents spent an average of 1.8 hours per day documenting outside of office hours. Here, too, AI scribes are showing early promise. In a large-scale pilot study at the Permanente Medical Group, AI scribe adoption was associated with a significant decrease in EHR use outside of working hours (7).

By enabling psychiatrists to focus more on their patients and less on the EHR during interviews, AI scribes can improve both patient rapport and satisfaction. Although the empirical evidence summarized in systematic reviews is mixed (15, 16), surveys suggest that many physicians are concerned that EHRs negatively affect patient rapport and communication (17). Preliminary data from pilot work at the Permanente Medical Group found that 81% of surveyed patients reported that their physicians spent less time looking at the computer screen after the implementation of AI scribes (7). Similarly, about three-quarters of dermatologists’ patients surveyed by Cao and colleagues reported that “my visit felt more like a personable conversation” after the implementation of an AI scribe (11).

Caveats and Potential Risks

Like all new technologies, AI scribes introduce potential risks that must be carefully considered by psychiatrists before these tools are adopted. Foremost among these concerns are privacy and security. The APA notes that AI tools used in clinical practice must comply with applicable health privacy laws and regulations, including the Health Insurance Portability and Accountability Act (HIPAA) in the United States (18). The U.S. Food and Drug Administration does not currently regulate AI scribes, because they are not considered medical devices. Beyond regulatory compliance, psychiatrists must also consider how protected health information is stored and used by the AI scribe, including further training of the AI model. The APA publishes a structured framework to help psychiatrists evaluate privacy and security risks, among other domains, for mental health apps (19).

AI scribes, like any tool, are not perfect and may produce errors, or “hallucinations,” which may take the form of factual inaccuracies, omissions, and unsupported assertions. For example, one study (7) noted that an AI scribe documented that a prostate examination had been performed when the physician had only mentioned scheduling one. Other examples of hallucinations include listing previous medications as current medications or excluding key parts of the patient’s history that were discussed during the appointment. Psychiatrists retain responsibility for the accuracy of their documentation and must, therefore, remain vigilant about reviewing and editing all documentation before signing, preferably immediately after the encounter. Some AI scribe programs are addressing concerns about accuracy by allowing users to review recorded evidence from the interview transcript that supports each portion of the structured note.

In addition to introducing factual errors, AI scribes may also subtly mischaracterize portions of the interview. This consideration may be particularly relevant in psychiatry, where patients communicate symptoms in nuanced, complex, and culturally influenced ways that require interpretation by the psychiatrist. Most AI scribes currently only use audio input and may miss important nonverbal communication in the interview. In other words, AI scribes may capture the text but miss key subtext in clinical encounters. Psychiatrists may, therefore, find AI scribes helpful for capturing fact-based aspects of their documentation (e.g., histories and plan) while continuing to write their own mental status examinations and assessments.

AI scribes also introduce a risk of bias that must be acknowledged and managed. AI scribes, like many generative AI tools, are trained on large data sets that may reflect existing biases in health care, including those regarding race, gender, and socioeconomic status. These potential biases in training data may carry over into output from the AI scribe (20). For example, AI scribes may be more likely to mischaracterize portions of conversations with patients from marginalized communities, who may be less well represented in the training data. The risk of bias further underscores the importance of psychiatrists reviewing and editing all AI-generated documentation.

The innovative nature and the potential risks of this technology make informed consent crucial. Psychiatrists should obtain and document patient consent before using AI scribes to ensure transparency (18) and compliance with applicable state laws regarding recording conversations. We recommend both verbal consents during patient encounters and written consents that provide additional details about the potential risks and benefits of this technology.

Psychiatrists must also carefully consider whether an AI scribe is clinically appropriate for each encounter. Introducing an AI scribe may threaten rapport and trust for a patient who is experiencing paranoia, for example, or it may escalate agitation in someone with persecutory delusions.

Finally, AI scribes also carry a risk of failing to deliver on their value proposition as they are implemented in specific real-world settings. Although many of the previously cited studies describe a positive effect from adopting AI scribes, these findings are not universal. For example, Haberle and colleagues (21) conducted a cohort study comparing 99 outpatient providers with access to an AI scribe to 76 matched individuals in the control group. Over a 6-month period, there were no meaningful differences in documentation quality, provider productivity, or patient satisfaction. Implementing AI scribes entails substantial financial costs and requires considerable training time, which may not always justify the purported benefits in every clinical setting.

Initial Recommendations and Next Steps

In light of the potential risks and benefits discussed, we offer several preliminary recommendations for psychiatrists who are evaluating and implementing AI scribes in their practices (22) (Box 2). We also offer practical recommendations for using AI scribes in clinical encounters (Box 3).

BOX 2. Recommendations for evaluating and implementing AI scribes

Review Employer Policies, if Applicable

Refer to relevant AI policies from the health system, practice, and/or medical school, if employed.

Consult Guidance From National Organizations

American Psychiatric Association, “The Basics of Augmented Intelligence: Some Factors Psychiatrists Need to Know Now” (18).

American Psychiatric Association, “App Evaluation Model” (19).

American Medical Association, “ChatGPT and Generative AI: What Physicians Should Consider” (20).

American Medical Association, “Principles for Augmented Intelligence Development, Deployment, and Use” (22).

Evaluate Privacy and Security Concerns

Confirm that the AI scribe complies with relevant federal and state laws and regulations (e.g., HIPAA).

Key questions include the following: How does the AI scribe use patient data? Where and how long is clinical information stored? What security measures are in place to protect patient data?

Consider Pricing Models and Value Proposition

Many AI scribes offer tiered monthly subscriptions that are based on the number of users, volume of clinical encounters, and/or specialized features.

Model potential financial impact (e.g., increased efficiency, improved satisfaction/retention of clinical staff).

Pilot a Candidate AI Scribe

Start with a small-scale pilot of early adopters to assess performance and acceptability to both psychiatrists and patients.

Some AI scribes offer free trials.

Incorporate feedback and lessons from pilot into training and support plan for wider implementation.

Develop Training and Support Plan

Initial training sessions should introduce functionality, describe risks and benefits, and demonstrate how to incorporate the AI scribe into the workflow.

Ongoing support should be customized for your organization, but it may include help desk support, recurring workshops, or shared online resources.

Evaluate Its Effect on Practice

Periodically assess efficiency, note quality, physician satisfaction, and patient reaction and/or satisfaction before and after implementing the AI scribe.

Determine whether the AI scribe can be customized for psychiatric encounters or improve over time with user feedback.

BOX 3. Recommendations for using AI scribes in clinical encounters

Consider Appropriateness for Each Clinical Encounter

The AI scribe may threaten the therapeutic alliance with patients who experience paranoia or a general distrust of health systems or technology.

The AI scribe may not provide useful output for disorganized patients, patients with acute delirium, or patients with major neurocognitive disorders.

Obtain and Document Patient Consent

The consent requirements of individual institutions will vary, but verbal and written consents are recommended.

An example of obtaining verbal consent is as follows: “I am using a new tool that securely records our conversation and helps me write my note. This product is HIPAA compliant, but there are safety and security risks with any electronic tool. May I use it during our appointment?”

Written consents may include more detailed information about AI scribes, including potential risk and benefits.

Once the AI Scribe Is Recording, Conduct the Interview Normally, With a Few Tweaks

Place the microphone on a surface between you and the patient for best audio quality. Conduct the psychiatric interview as usual. Summarize findings from the interview—particularly your assessment and plan—together with the patient.

Review AI Scribe Output for Accuracy and Edit as Needed

Check for key omissions or AI hallucinations, preferably immediately after the encounter. The AI scribe is likely more helpful for documenting a patient’s history of the present illness, histories, and plans. We recommend documenting mental status examinations and assessments independently.

Do Not Use Other AI Tools During Clinical Encounters

Do not enter protected health information into general consumer AI products (e.g., ChatGPT), because these tools do not adhere to federal patient privacy laws.

As previously mentioned, there are currently no published empirical studies on AI scribes in psychiatry. Although a small but growing literature describes the use of AI scribes in other specialties, these findings may not generalize to psychiatry, given the unique nature of the psychiatric interview. This suggests an opportunity to systematically evaluate the effect of AI scribes on clinical efficiency, documentation accuracy and bias, and patient satisfaction in psychiatry. We anticipate that the literature will begin with quality improvement projects and feasibility studies on a smaller scale, laying the groundwork for larger and more rigorous multisite evaluations in the future. Continued integration of AI scribes into EHRs may facilitate the passive collection of efficiency data. Mixed-methods designs could provide a more comprehensive understanding of usability and acceptability of AI scribes for psychiatrists and patients, particularly given the novelty of this technology.

Conclusions

AI scribes hold significant promise for improving clinical documentation in psychiatry by enhancing efficiency, reducing burnout, and strengthening patient rapport. Early implementation studies in other specialties have shown mostly positive outcomes, such as decreased time spent on EHRs and high acceptability to patients and physicians. However, psychiatrists must remain vigilant about the potential risks, including privacy concerns, AI hallucinations, bias, and the nuanced nature of psychiatric care that may not be fully captured by the AI scribe. Careful consideration of these factors and adherence to recommended guidelines will be crucial as psychiatrists navigate the integration of AI scribes into their practice, ensuring that the benefits are maximized while minimizing potential drawbacks.