Ambient AI scribes, also referred to as ‘digital scribes’ [1] or ambient voice technology (AVT), are emerging tools designed to passively capture clinician–patient conversations and automatically generate documentation within electronic health records [2]. This is a rapidly expanding market which may reduce the documentation burden on clinicians, thereby addressing one of the major drivers of burnout in primary care [3].

Despite this promise, their use raises important challenges [4, 5]. Concerns include accuracy and reliability of transcribed notes, risks of errors or omissions in documentation, medico-legal liability, patient consent, and the potential loss of nuance in clinical communication [6]. Furthermore, biases in speech recognition and summarisation may result in systematic disadvantages for certain groups of patients, such as those with strong regional or foreign accents, or individuals with speech disorders. The rapid pace of product development also creates uncertainty about consistency, integration into existing record systems, and implications for professional practice and for regulation.

Although interest in ambient AI scribes is growing, there is little empirical evidence on how widely they are being adopted in frontline general practice and how they are perceived by clinicians who use them [7, 8]. Few studies have directly explored the lived experiences, uptake, and opinions of doctors regarding these tools [7]. Given the apparent interest and rapid adoption in primary care settings, this study seeks to fill that gap by examining general practitioners’ (GPs) use of, and attitudes toward, ambient AI scribes in the UK.

This survey follows the format of previous studies we have conducted into UK GPs’ experiences and opinions with AI tools [10, 11]. We devised an anonymous, approximately 5-minute questionnaire that defined “ambient AI scribes” for respondents as background systems that capture consultation dialogue and automatically generate documentation such as notes or referral letters in the electronic health record (see Appendix 1). A screening question ensured that only current users of ambient AI scribes proceeded with the survey; here we also asked whether GPs who had not used these tools intended to do so in the near future.

The instrument consisted of 8 core closed-ended questions, with some conditional follow-ups and a single free-text prompt. Items addressed which products were being used (multi-select, including examples such as Nuance DAX, Heidi Health, Suki AI, DeepScribe, Augnito, and Amazon HealthScribe), the frequency and clinical significance of transcription errors, how often such errors were corrected, and whether errors were more common in particular patient groups. Other questions covered medico-legal concerns, perceived changes in cognitive workload, burnout, time spent on documentation, overall note quality, and approaches to patient consent, including the proportion of patients who declined when consent was sought. Aside from the 8, core closed-ended questions, the survey included 7 demographic questions including about weekly working hours, and the average number of patients seen per day. The survey concluded with an optional free-text box inviting respondents to share additional reflections. All closed-ended questions were mandatory for submission, while the free-text question was optional.

Prior to launch, the instrument was piloted with five UK-based GPs to ensure clarity and usability. The final survey was designed to take 3–5 minutes to complete. Survey design and reporting followed the Checklist for Reporting Of Survey Studies (CROSS) guidelines (see Appendix 2).

The study protocol was approved by the Faculty of Psychology Ethics Committee at the University of Basel, Switzerland (Nr 017-25-1). Participants were provided with study information and gave informed consent electronically before participation. The survey was hosted on a secure server maintained by Doctors.net.uk. All responses were anonymised and encrypted prior to transfer for analysis. Personal identifiers, including email addresses, were removed before data handling. The study complied with the EU General Data Protection Regulation.

Participant characteristics and responses to survey opinion items were summarised through descriptive statistics. To analyse whether users and non-users of ambient AI scribes differed in gender, age, role, or practice size, a chi-square of independence test was carried out for each variable. If significant, it was followed by a post-hoc two-proportion z-test, where one variable level was tested against all others. Bonferonni-adjusted p-value was used. Given our non-random sample, to analyse whether differences with GPs in the GMC Registry, a chi-square goodness-of-fit test was used. Significance levels were pre-set to p < .05 for all tests. Analyses were carried out by AK in JASP (v0.95.1.0) and (v2.6.2). Tables and figures were created by AK.

To capture richer perspectives, the survey included a single optional free-text item: “Please add any comments about the topic or the survey (1–2 brief comments).” Free-text responses were cleaned to remove trivial entries (e.g., “none,” “N/A”). Due to limitations with the data set, full inductive coding was not applicable. CB reviewed the data, developed descriptive codes, and iteratively refined coding decisions through discussion. Where responses contained multiple ideas, multiple codes were assigned. First-order codes were then grouped into higher-order categories to summarise recurrent patterns. This process was also independently carried out by GC, and through discussions, decisions were reached.

Of the 141 respondents, 29 (21%) provided free-text comments, generating a total of 689 words after data cleaning. Most entries were concise, often comprising short phrases or 1–3 sentences. Through inductive coding and iterative analysis, four overarching categories of views on AI scribe use in general practice were identified: (1) Perceived benefits in workflow and patient care; (2) Accuracy and reliability concerns; (3) Ethical and professional considerations; and (4) Future expectations and system-level challenges. Quotes are anonymised and presented verbatim; participant number, gender, and age are in parentheses. Overall, GPs’ views captured a blend of enthusiasm and caution. While AI scribes were recognised as valuable in alleviating workload and enhancing patient focus, concerns about accuracy, professional identity, and system integration underscored the need for ongoing refinement and support.

This convenience sample survey of UK GPs provides the largest snapshot to date of current and intended use of ambient AI scribes in frontline practice. While uptake remains limited (14%), a further 39% of GPs anticipate near-term adoption, underscoring the rapid diffusion of these technologies. Users reported substantial gains in efficiency, with most experiencing reductions in both time spent on documentation and cognitive load. Importantly, perceived errors were typically minor, though a non-trivial proportion were considered clinically significant, and concerns were heightened in complex and multilingual consultations. These patterns highlight the dual potential of ambient AI scribes: to alleviate workload and improve documentation quality, but also to exacerbate inequities and introduce medico-legal risk if governance is weak. Clear national regulation, robust safety monitoring, and consistent consent practices are urgently needed to ensure safe, equitable deployment. Prospective, multi-site research incorporating objective workflow and safety metrics—as well as patient perspectives—will be crucial to determine the long-term role of ambient AI in primary care.