Exploring AI scribes in Norwegian general practice: a qualitative individual interview study Authors: Damoun Nassehi, Øystein Hetlevik, Jørgen Breivold Published: 2026 ABSTRACT AbstractObjectiveTo explore how Artificial intelligence (AI) scribes influence Norwegian general practitioners’ (GPs) clinical practice, focusing on their impact on workflow, consultation routines, and professional responsibility, and to explore GPs reflections on AI technology and its alignment with their needs, expectations and ethical considerations in primary care.DesignQualitative individual interview study using systematic text condensation (STC).SettingGeneral practice in Norway.SubjectsTen GPs with more than 3 months’ regular use of an AI scribe. Participants were self-selected early adopters.ResultsThe GPs in our study reported that using AI scribe markedly saved time, while reducing stress levels. The technology improved consultation structure and patient focus, though some GPs missed their personal writing style and found the language overly standardised. The tool sometime struggled capturing nuances during complex discussions. Occasional transcription errors, and hallucinations highlighted the need for careful review. GPs expressed concern regarding data privacy and regulatory uncertainty but maintained a stance of critical trust. Overall, the AI scribe was seen as a major innovation that enhances efficiency and patient care, provided its outputs remain under vigilant control.ConclusionAccording to the respondents, AI scribes influence GPs daily routines and professional identity by enhancing documentation efficiency, reducing workload and enabling GPs to focus more on patient care. However, they may introduce new areas for error and subtle shifts in clinical practice, such as reliance on AI-generated text. Responsible integration requires collaborations on guidelines, training and further research to balance efficiency gains with patient centred care and address long-term impacts.Keywords: General practice, primary care, artificial intelligence, large language model, generative AI, AI scribe, professional identity, clinical documentation Introduction Artificial intelligence (AI) is increasingly integrated into healthcare to streamline both clinical and administrative tasks. One emerging application in primary care is AI scribes, based on large language model (LLM) technologies, for automatically documenting medical consultations. These systems are designed to alleviate the administrative burden on general practitioners (GPs), which is a growing concern across many healthcare systems. A scoping review from 2025 reported reduced documentation burden, enhanced clinician satisfaction and increased workflow efficiency using AI scribes, although the evidence was sparse [1]. This is in line with a systematic review of qualitative studies by Martinez-Martinez et al. [2], indicating that AI could help address workload and decision-making and improve efficiency. However, while LLM-based AI systems show potential, they also face significant challenges. For instance, these tools have been found to produce notable errors [3] and having gaps in accuracy and contextual understanding [4]. Prior qualitative research among Norwegian GPs found a strong openness to adopting new digital tools to support clinical work [5]. Although AI was not specifically examined in that study, it identified key motivators for technology adoption, such as improving workflow efficiency, maintaining high-quality documentation, and reducing time spent on administrative tasks. These incentives align closely with the intended benefits of AI scribes and raise the question of whether such tools are fulfilling these needs in practice. Building on this context, the aim of the study was to explore how AI scribes influence Norwegian GPs clinical practice, with a focus on their impact on workflow, consultation routines, and professional responsibility. The study also sought to explore GPs reflections on the use of AI technology and its alignment with their needs, expectations and ethical considerations in primary care. Material and methods Settings and participants From January to March 2025, the first author conducted semi-structured interviews with 10 GPs from various clinics to explore their experiences using AI scribes. This study utilised video conferencing technology to include GPs from across Norway. The GPs were predominantly specialists in general practice and male (N = 7). All GPs worked in clinics with more than one physician. In a couple of clinics, all the GPs had adopted AI scribes, but this was not the norm. The GPs mainly used transcription software from the company Noteless (N = 8), while the remaining two (N = 2) used Medbric’s AI scribe. Most GPs (N = 8) were recruited through a recruitment notice posted on the first author’s blog, LinkedIn page, and the informal Facebook group Allmennlegeinitiativet, which is frequently visited by the majority of Norwegian GPs. In addition, two GPs were found through referrals from GPs who had already been interviewed. We only included GPs who had used AI scribes for more than three months to ensure familiarity and experience with the product. As the technology was rather new, no GP had used AI scribes for more than one year. We recognized the potential that this could lead to an overly favourable perspective on AI scribes but determined that firsthand experience with the technology was of greater significance. More detailed demographic data on the GPs is presented in Table 1. Table 1.Demographic data of GPs (n = 10).CategoryValueAge (years) Median38 Range31–54Gender (n) Female3 Male7Profession (n) Specialist in general practice7 Trainee in general practice3Experience as physician (years) Median9 Range0.5–35Place of medical degree (n) Norway5 Other country5Workplace (n) City or town5 Village5Geographical region (n) Northern1 Middle1 Western7 Southern0 Eastern1 Demographic data of GPs (n = 10). Data collection The interviews featured open-ended questions, enabling physicians to reflect on their experiences with cues such as: ‘Can you tell us why you chose to use AI-powered transcription services?’ and ‘What has been your experience with using AI-powered transcription services in practice?’. After this invitation to share their experiences with AI scribes, we asked them about the perceived advantages and challenges of using these tools. This was followed by a couple of questions exploring whether the use of the AI tool had any impact on the GPs practice and their care for patients, as well as in which situations they found the tool to be most useful. At the end of the interview, we asked about their concerns regarding patient privacy and data security and, finally, if they had any thoughts on the future use of AI tools in general practice. The conversations were audio-recorded and transcribed using OpenAI’s WhisperAI, run locally on a PC using Python – a setup that ensured all audio data was processed on a local machine without transmission to external servers, thereby protecting the confidentiality of participant data. Transcriptions were compared with the audio recordings to ensure their validity. Each interview lasted between 20 and 41 min, with both the mean and median being 29 min. For the purposes of this study, the names of the GPs have been anonymised, and pseudonyms have been used in their place. Data analysis The data were analysed using systematic text condensation [6]. The analysis followed a four-step process. First, the transcripts were thoroughly read to ensure a comprehensive understanding, and preliminary main themes were discussed. Second, meaningful units aligned with the study’s aim was identified, decontextualised, and after discussion organised by related text elements in code groups and subcategories. Third, code groups were condensed into specific content, through the construction of narratives centred on key units within each subcategory, as identified by the authors. This process was assisted using ChatGPT-5 (OpenAI, 2025). All AI-generated condensates were carefully reviewed by checking for consistency between all meaning-bearing units from Step 2 and the AI-proposed condensates in Step 3. We examined each condensate and corrected any discrepancies identified. In this manner, Step 3 was validated and aligned with our own analysis of the material. The authors retained full responsibility for the final analysis and interpretation. Finally, recontextualization was conducted, where the content of each code group was summarised, and an analytical text with categories and subcategories was composed, and adjusted through discussion, merging and relocation. This study is grounded in mild constructivism, consistent with Malterud’s framework for systematic text condensation, in which meaning is understood as co-constructed between participants and researchers rather than discovered as objective fact. All three authors are practising GPs with experience of the Norwegian primary care system, which provided insider knowledge of the clinical context and enabled nuanced interpretation of participants’ accounts. The first author (DN) has extensive experience with digital health technology, has tested and built AI scribes, but has deliberately chosen not to use one in his own clinical practice, providing informed critical distance from the technology. The other two authors have both tested AI scribes in clinical practice. Results Four main themes were constructed from the interviews, together capturing the GPs experiences of using AI-based transcription tools in everyday clinical work. These themes span the impact on the GP role and workflow, changes in the doctor–patient relationship, perceptions of the quality and reliability of clinical notes, and broader reflections on professional responsibility and ethical frameworks (see Tables 2 and 3). Table 2.Overview of main thematic categories and subthemes.Main categorySubthemes1. Impact on GP workload and work stylea) Time savings, stress reduction and efficiencyb) Adaptation of work style and consultation flow2. Impact on doctor–patient relationshipa) Presence, communication and patient involvementb) Information practices and patient reactions3. Impact on clinical notesa) Documentation quality and structureb) AI-generated language and stylec) Situational appropriateness of an AI scribed) Errors, misunderstandings and hallucinationse) Technical aspects and system stability4. Concerns for privacy, quality and impact on professional identitya) Privacy and data securityb) Quality assurance and control routinesc) Trust and technological restraintd) Professional identity and the future Overview of main thematic categories and subthemes. Table 3.Example of analysis process using systematic text condensation (STC).StepExample1. Overall impressionThe initial reading indicated that GPs perceived AI transcription tools as both liberating and risky – improving efficiency and clinical presence while introducing new ethical and technical concerns2. Identifying meaning units‘It sometimes writes things I never said … like adding details I didn’t perform to an examination’3. Condensation and codingMeaning unit: AI adds incorrect or fabricated informationCondensed statement: The system may generate clinically misleading contentCode: Errors and hallucinations4. Synthesising to analytical textSub-category: Errors, misunderstandings and hallucinationsCategory: Clinical NotesAnalytical text: GPs must critically review AI-generated drafts, as the tool may insert subtle inaccuracies that could compromise clinical documentation if left uncorrected Example of analysis process using systematic text condensation (STC). Impact on GP workload and work style Time savings, stress reduction and efficiency GPs consistently reported that the AI scribe significantly reduced time spent on writing notes, often saving three to five minutes per patient and roughly one to two hours daily. Much of the documentation was now completed during the consultation rather than afterwards, eliminating the need to catch up on notes in the evening. While proofreading the AI’s output was still necessary, it was far less time-consuming than manual note-writing. GPs also felt less drained and could stay on schedule, addressing last-minute patient concerns without running late, because the tool was documenting in the background. Some GPs mentioned that the scribe helped rekindle a sense of work satisfaction: I hardly spend evenings anymore finishing up the writing (…) it’s a whole new everyday life. (Gordon) Overall, the respondents reported that the tool improved efficiency, saved time, reduced stress and enhanced focus on patient care. Adaptation of work style and consultation flow Using the scribe required doctors to adjust their consultation style, particularly by verbalising more of the clinical examination and decision-making process. Initially this felt unnatural, but most found that narrating their actions and findings soon became a routine and even improved the consultation flow: Several doctors began providing a running commentary during physical exams, which kept patients informed and ensured the AI captured all the details. You must verbalise everything you do and everything you find, which can feel a bit strange for us at first. However, once you have learned it, it works quite well. (Ivy) This change also had a pedagogical benefit by engaging patients and making the consultation more transparent. Overall, the adaptation in work style was seen as positive, leading to smoother workflows and more interactive consultation. Impact on doctor–patient relationship Presence, communication and patient involvement GPs reported that the scribe allowed them to be more present and attentive during consultations. Freed from the keyboard, they maintained eye contact, used open body language, and shared clinical reasoning aloud, which patients appreciated. One GP noted it this way: The less time you spend typing, the more time you have with the patient. (Bob) GPs felt more relaxed, leading to consultations that were more patient centred. Overall, the AI tool strengthened their rapport with patients by enabling clearer communication and reducing distractions. Information practices and patient reactions Most GPs informed patients about the recording and transcription, either directly or by an office notice, and they universally reported that no patient had outright refused the use of the scribe. The typical patient reaction ranged from neutral to curious, with some patients showing interest in the novel technology. One GP said: They are often impressed. “Yes, this was exactly what we talked about,” they say. (Theodore) However, some doctors admitted to occasionally failing to inform every patient, acknowledging this as an ethical grey zone. Impact on clinical notes Documentation quality and structure A prominent finding was that the GPs often had experienced the AI-generated notes as more comprehensive and better structured than the ones doctors would write on their own. I think they are easier for others to read. (Robert) GPs also observed that these notes captured details in real time that they themselves might have missed or omitted. Some GPs went so far as to say the documentation quality with AI was as good as or better than their usual standard. They felt that having more complete notes could enhance continuity of care – for instance, by preserving more patient information for future visits than might otherwise be recorded. A few did note that the AI’s thoroughness resulted in lengthier notes than necessary, but the consensus was that the scribe met its promise of not only speeding up notetaking, but also producing organised, detailed notes that covered all the key points of the consultation. AI-generated language and style While the content of the notes improved, some GPs noted that the AI’s standardised language and style made their notes feel uniform and impersonal, with a loss of their personal ‘voice.’ This artificial or overly formal tone occasionally made it harder for some GPs to recognise their own notes and recall patients during follow-up consultations: It seems artificial, as every journal entry consists of history, findings, assessment, plan. All presented in bullet points. (Robert) There were also instances where the language was overly verbose or included slightly unconventional medical terminology, requiring the physician to edit the wording to match their usual tone. Most found the stylistic issues relatively minor – they could always tweak the phrasing if needed – whereas the gains in consistency and detail were significant. Several GPs appreciated that the AI’s transcript could capture patients’ own words more directly than a manually written note might. At the same time, a few doctors cautioned against over-reliance, emphasizing the importance to scrutinize its output and avoid complacency Situational appropriateness of an AI scribe The GPs found that the scribe was especially well suited to certain types of consultations and tasks. In routine somatic consultations the AI performed excellently, capturing the necessary information and structuring it without trouble. GPs also reported that it was very useful for acute or urgent visits, where being able to document on the fly helped them manage time pressure. It also drafted lengthy referral letters or medical reports efficiently. Some GPs found the AI scribe useful in lengthy consultation, particularly in psychiatric and therapy settings where patients require attentive listening. Perhaps it is in psychiatry, when patients need a general practitioner who listens, that this approach is particularly useful. (Theodore) Challenges were sometimes noted in capturing nuances during complex discussions, requiring manual adjustments for accuracy. The GPs noted that the scribe is not a one-size-fits-all solution – some encounters still require a manual or more nuanced documentation approach. Errors, misunderstandings and hallucinations All the GPs experienced occasional errors in the AI-generated notes, which meant that proofreading was essential. Some errors were minor, such as medication names being misheard or mistyped. More troubling were instances of ‘hallucinations,’ where the AI added details that were never actually mentioned. You can get all the hallucinations, additions and such included in the record, and that could potentially pose a risk. (Liam) In a few cases, the AI appeared to mix up context, which could produce serious errors. If a patient spoke about a family member’s health issue during the consultation, the system might incorrectly merge that information into the patient’s own note. One GP recounted how an AI transcript merged two separate issues from a visit into a single narrative, which actually led to an incorrect referral. All these experiences underscored for all the GPs that the AI’s draft could not be taken at face value. The general view was that the scribe dramatically speeds up documentation and captures a lot of information, but the physician remains the final safeguard for quality and must intercept any errors or invented details. Technical aspects and system stability The GPs reported that the transcription system was generally stable, but technical glitches did occur occasionally. Rarely the software could crash, freeze or fail to generate a note at the end of a consultation – events that were highly disruptive when they happened. The GPs had to revert to memory to complete the note, which was frustrating. Another technical limitation was that the system could not distinguish between multiple speakers. It can’t distinguish between several different people who are present in the consultation at the same time. (Ivy) If a patient’s spouse or parent interjected during the visit, the AI might attribute all spoken words to a single speaker, jumbling the content. Language support was also limited – the tool functioned best for Norwegian. Some GPs found that if parts of the consultation switched to English or another language, the transcription quality dropped significantly. The GPs recognised the need to stay sufficiently attentive during consultations and to have fallback plans for documentation when necessary so they could reconstruct the clinical content if technical problems occurred or if they had forgotten to start the recording. Concerns for privacy, quality and impact on professional identity Privacy and data security When discussing the broader context of using AI in practice, GPs frequently raised concerns about data privacy and security. They generally trusted the official assurances given by the software vendors that patient data was handled securely, but many still felt uneasy about the potential for data breaches. We have no legislation in Norway that regulates this, so it is hard to know what happens if a note leaks. (Ivy) This regulatory grey area left GPs unsure about accountability and consequences. The sentiment of having to relinquish control over sensitive information made some GPs uncomfortable. Several GPs admitted that in the absence of concrete regulations, they had little choice but to rely on the technology providers and assume that the proper safeguards were in place. They were aware that the audio and text data were being processed on external servers (often overseas) and that theoretically, anything stored digitally, even temporarily, could be vulnerable to hacking or misuse. The GPs felt that formal frameworks were needed to ensure patient confidentiality is protected and to define legal responsibility if something goes wrong. Despite these worries, all GPs continued to use the scribe because of its clear benefits. Quality assurance and control routines All the GPs stressed that they viewed the AI-generated note as a draft that required their input and approval. This meant they had established routines to carefully read through and edit the AI’s notes before signing them off. I must discipline myself to read through the notes and quality-assure them before I sign them. (Theodore) GPs described being extra vigilant, especially early on, to ensure no mistakes slipped by. The GPs also acknowledged the temptation to become a bit lax when the AI’s performance was consistently good. They agreed that maintaining strict quality control habits was crucial and a part of their professional duty, but some admitted that if they were very tired or busy, there were occasions when they almost trusted the system too much and could skip the thorough review. Trust and technological restraint The GPs in our study adopted what could be called a stance of ‘critical trust’ towards the AI scribe. They valued the tool’s contributions – its ability to ease workload and improve documentation – but were careful not to place blind faith in it. One GP advised: That is, show a healthy scepticism but also use the tools that can make GP life better. (Eric) The GPs voiced a determination to avoid overreliance on the technology. They gave examples of how they deliberately stay engaged – for instance, by mentally summarising the case in their own words even if the AI has written the note, or by double-checking key details like medication dosages and follow-up plans. A few noted that if one becomes too dependent on the AI, there is a risk of becoming ‘lazy’ or less sharp. Some also mentioned that we will need to train future clinicians to handle these tools smartly and responsibly. Medical education may have to evolve so that new doctors learn to work with digital assistants while still honing their own clinical skills. Professional identity and the future Many GPs saw AI-based documentation as a major innovation in their field. One GP remarked: In all the years I have worked, this is the biggest revolution for me. (Eric) GPs expressed that tools like the scribe are likely here to stay and will become an integrated part of how doctors work. They envisioned that in the near future AI will be built into electronic health record systems and be as routine as the stethoscope or the computer itself in consultations. Most doctors were optimistic about the potential for further integration of AI to enhance various aspects of practice, and they welcomed the idea of offloading more administrative burdens to intelligent systems. GPs discussed the importance of setting ethical and practical limits on AI’s role. For instance, they noted the risk of information overload if every detail of a consultation is recorded verbatim, or the danger of dependency where future clinicians might not learn documentation skills or place excessive trust in AI outputs. They also recognised that while AI might make some tasks easier, it could introduce new ones (such as reviewing long transcripts), and thus the net effect on workload would need to be monitored. The GPs felt these effects were manageable if developments are approached with critical awareness and guided by ethical standards. Discussion Main findings GPs reported that using AI scribe markedly saved time, while reducing and improving stress levels. The technology improved consultation structure and patient focus, as patients’ own words directly shaped the journal text. However, some GPs missed their personal writing style and found the language overly standardised. While effective for routine and long consultations, the tool sometime struggled capturing nuances during complex discussions. Occasional transcription errors, and hallucinations highlighted the need for careful review. GPs expressed concern regarding data privacy and regulatory uncertainty but maintained a stance of critical trust. Overall, the AI scribe was seen as a major innovation that enhances efficiency and patient care, provided its outputs remain under vigilant professional control. Strengths and weaknesses of the study This qualitative study provides insights into GPs experiences with AI scribes during their early adoption in clinical practice. We employed systematic text condensation (STC) for analysis. The credibility of our findings depends not on the choice of method alone, but on the rigour with which it was applied and reported. Additionally, our sample was drawn from GPs across a broad geographic range in Norway and included practitioners with over three months experience, relevant for integrating AI scribes to ease documentation burdens. Despite its contributions, the study has important limitations. First, the GPs were self-selected early adopters of AI scribe technology, mostly recruited via the first author’s professional blog (www.hacking-healthcare.com), LinkedIn page, and the Facebook group Allmennlegeinitiativet, with two additional participants recruited through snowball sampling. This recruitment approach likely attracted technologically proficient and innovation-oriented GPs, which may limit transferability of the findings to the broader GP population. However, many of our core findings – particularly those concerning changes in the doctor–patient relationship, professional identity, and the cognitive implications of delegating documentation – are likely to resonate across healthcare systems, as these dimensions of clinical practice are not highly context-dependent. Second, our study focusses solely on GPs perspective, omitting patients’ views on being recorded or having AI-generated clinical notes. The patient perspective is an important direction for further research, as successful implementation of AI scribes ultimately depends on patient acceptance and trust. A methodological limitation worth acknowledging explicitly concerns the use of ChatGPT-5 in Step 3 of the STC process – the condensation of meaning units into narrative text. Malterud herself, along with a growing number of qualitative researchers [7], has argued that large language models are incompatible with the interpretive and reflexive paradigm underlying STC, as probabilistic text generation cannot replicate the situated, first-person condensation that is the method’s defining feature. We recognise this tension. Our decision to use AI assistance at this specific step was pragmatic: the AI-generated condensates produced more precise and articulate formulations than initial drafts, and the process stimulated productive dialogue between two authors (DN and JB), who systematically cross-checked all condensates against the underlying meaning units and corrected discrepancies. We therefore cannot claim to have applied STC in its strict sense. What we can claim is that human interpretive judgement remained central throughout, and that we believe the analytical gains from this approach outweighed the methodological compromise. We acknowledge that this remains a contested area in qualitative methodology, and that future studies should address the role of generative AI in interpretive analysis more explicitly. As all authors are also GPs, they have an inherent understanding of the participating GP’s ontology. This made it easier to recruit GPs as well as made the interviews easier to conduct. Alternatively, blind spots or wrong assumptions on the part of the interviewing author could have a negative impact on the study. As the interviews were conducted using video-conferencing tools such as Teams and Zoom, it is possible that some non-verbal signals could have gone unnoticed. It is also the first authors experience from prior work, that in-real-life interviews tend to be a bit longer compared to those performed over Zoom. Still, it is not clear whether that extra interview time would include significant extra material. Using video conference as standard enabled us to reach GPs from a much larger geographical area. Finally, as with any small qualitative study, the findings are context dependent. All interviews took place within one country’s healthcare system, which may limit transferability. Nevertheless, Norway is among the leading countries in Europe in terms of AI scribe adoption, making these early experiences potentially instructive for healthcare systems at earlier stages of implementation. Future research should include diverse stakeholders – especially patients – and extend beyond single-country contexts. Discussion of findings Our study suggests that AI scribes are not merely an add-on to existing practice but a transformative presence in the consultation room. The GPs reported improved clinical encounters, as AI scribes enabled better eye contact and attentive listening, reducing distractions from the ‘computer in the corner.’ This sentiment echoes reports that AI scribes allow physicians to be more present with their patients – observing nonverbal cues and building rapport without the barrier of a screen [8]. In this sense, the technology reshapes patient interaction and transforms documentation, with ambient scribes capturing detailed transcript that could make medical records longer and more verbatim. The GPs appreciated having thorough notes. Some participants observed that the AI scribe used slightly different phrasing or less formal medical terms, requiring adjustments to ensure accuracy in reflecting examinations. Reliance on AI might affect clinicians’ cognitive processes. Some have argued that automating complex tasks could inadvertently erode practitioners’ critical thinking and decision-making skills over time, as clinicians may come to depend on AI for functions that previously required active mental engagement [9]. These concerns were recently validated when endoscopist were presented with deskilling after exposure to AI in colonoscopy [10]. As Postman [11] argued, ‘new technology is ecological’ – it transforms the entire system. Our findings illustrate that AI scribes create a new type of consultation, with ambient notetaking and redistributed roles of observer and recorder. An important implication of this transformation is the potential for unintended consequences, such as the paradox that increasing efficiency might not reduce workload as much as expected [12]. While AI scribes reduced typing and sped up note closure, GPs reported using the spare time for taking on more patients. Some respondents pointed out that more comprehensive notes may strengthen continuity of care, presumably referring to informational continuity. Building on these observations, we argue that the notes may also feel unfamiliar and be less helpful for the physician when trying to retrieve information that is not documented – a kind of cognitive anchoring that in many cases is an important element of relational continuity, where the patient experiences that the doctor follows their story over time. One GP noted that the AI often produces a more detailed summary. Other researchers have also observed that AI systems can capture a fuller set of clinically relevant information from a conversation than a human might record manually [8]. Another dimension that became apparent is the extent to which AI scribes could capture patients’ own words more directly. This could help better understanding the patients and to help deliver more holistic care [13]. We suggest that if the notes are to be directly based on the patient storytelling and understanding, the medical interpretation may be lacking and may also introduce inequity if too closely linked to the patients’ health literacy. The latest generation of AI scribes represents a substantial leap forward in natural language processing capabilities, compared with traditional dictations systems. Their ability to convert unstructured clinical conversations into coherent is striking. Our findings align with emerging literature on the benefits of AI-driven documentation tools, while also revealing a gap in how clinicians perceive broader impacts. Like the GPs in our study, doctors in other settings have reported substantial efficiency gains from AI scribes. For example, a recent evaluation of ambient AI scribe technology in specialist care documented a 20% reduction in time spent on notes and a 30% decrease in after-hours charting, alongside improvements in same day note completion [14]. These objective gains align with our participants’ impressions that AI scribes improved efficiency helping GPs finish consultations on time and reduce administrative burdens. They are consistent with Blease’s et al. [15] findings, which show that many GPs view automation optimistically as a means to ease paperwork and improve efficiency in primary care. On the other hand, taking a more cautious view of the respondents’ estimates of time saved during a workday – especially given their own recommendation to read the notes carefully – suggests that a claimed saving of 1–2 h may be unrealistic. This is supported by a recent article showing relatively modest actual time savings (22 min) using AI scribes [16]. A particularly important concern is the risk of automation bias – the tendency to trust automated systems too readily [17]. Our GPs mentioning that they still check the AI-generated notes for errors, however, based on the time spent on documentation, it appears as they primarily skim through the notes, as one GP admitted. The same GP elaborated that he had found errors, that he had overlooked, in the patient journal upon the patient’s next visit. The potential for AI scribes to introduce subtle documentation mistakes is well documented. Duggan et al. [14] found that while ambient scribes reduced physicians’ typing, error-prone drafts often required significant editing, potentially negating time savings. The related concept of algorithmic appreciation – the tendency to over-rely specifically on AI-generated outputs – has been explored by Klingbeil et al. [18], who found that such overreliance can lead to unchecked errors if clinicians fail to verify AI outputs. While the GPs generally trusted their AI scribes, deliberate verification is essential to prevent rare mistakes from going unnoticed. Institutions should provide guidance and training to help clinicians audit AI-generated notes effectively. Our study also highlights the broader social, ethical, and clinical implications of delegating notetaking to an AI. In contrast to the narrow, task-oriented enthusiasm we observed, other research has found that clinicians do consider wider impacts when prompted. For instance, Kocaballi et al. [19] engaged GPs in co-designing a future AI documentation assistant and reported that participants actively questioned how such a tool might alter the nature of care. Those GPs raised concerns about medico-legal ramifications – noting that having a full recording and transcript of every consultation could create new legal and privacy dilemmas – and they cautioned against focusing solely on efficiency at the expense of patient-centred care. Similarly, in an exploratory survey of UK GPs about AI in primary care, some respondents voiced social and ethical reservations, including whether patients would find AI involvement acceptable and how increased automation might affect the doctor–patient relationship [15]. Compared to these studies, the GPs in our study showed relatively limited reflection on such issues. They did reflect on the problem with AI scribe hallucinations, factual errors and proclivity to amalgamate several speakers into one patient. They also questioned data security and questions of privacy, but the perceived efficiency gains of the technology made these reflections of minor importance. Finally, our study’s context – where GPs adopted AI scribes in the absence of formal guidelines, reflects a broader situation noted in the literature. Blease et al. [20] observed that despite growing interest in tools like ChatGPT for clinical tasks, there is ‘a lack of guidance about these tools and unclear work policies,’ even as many physicians have begun using them. We found that our GPs navigated ethical and regulatory questions on their own. This underscore concerns that governance is lagging technology. Without clear standards on issues like data privacy, documentation accountability, and integration with electronic health records, clinicians are left to use personal judgment on matters that could carry institutional or legal consequences [19]. Several GPs described a regulatory landscape they had not actively investigated, assuming vendor compliance – a pattern we interpret as reflecting an implicit, unarticulated need for clearer professional guidelines and frameworks, including clarification of liability in cases of serious AI errors. Prior commentaries have urged medical communities and regulators to establish clear boundaries for AI in practice [20]. In essence, our findings reinforce those calls; just as Postman [11] cautioned against ‘the surrender of culture to technology,’ letting AI scribes proliferate without robust guidelines or ethical oversight could amount to embracing the technology’s influence on clinical practice unchecked. Defining protocols for reviewing AI-generated notes, obtaining patient consent, and managing data will be essential to balance the benefits of AI scribes while safeguarding professional standards and patient care. Meaning of the study The transformative nature of AI scribes requires GPs and other stakeholders to critically engage with these tools, rather than adopt them unreflectively. Left to its own momentum, technology can shape practice in unintended ways. Postman’s [11] cautionary perspective is relevant here, focusing only on immediate benefits, such as time savings, risks overlooking deeper shifts in work patterns and professional norms. While the GPs were enthusiastic about regained time and smoother workflows, fewer reflected on data control or the implications of automating documentation. To ensure AI scribes serve clinical goals, GPs should take an active role in their implementation ensuring accuracy, providing feedback to developers and contributing to guidelines and policies. If GPs remain passive, AI documentation tools risk evolving primarily under commercial or technical pressures rather than clinical and ethical priorities. However, active participation through pilot programs, user groups, or professional committees can help shape AI integrations that align with patient care values. In essence, the introduction of AI scribes presents an opportunity for GPs to redefine what aspects of documentation should be automated and what must remain under human purview. Conclusion The GPs in this study reported that AI scribes provide practical benefits by improving documentation efficiency, reducing burdens and allowing greater focus on patients during consultations. However, our analysis also indicates that AI scribes introduce subtle changes in clinical practice, such as reliance on machine-generated text and shift in the GPs role. These changes highlight the need for greater awareness and discussion in the medical community. As AI adoption in healthcare accelerates – driven by rapid technological development, commercial availability, and policy-level endorsement – clinicians, informatics leaders, and policymakers must collaborate to establish guidelines for AI scribes, addressing data privacy, error correction and documentation standards. GPs should actively contribute their insights to ensure regulations are practical and aligned with clinical goals. Training and education will be important to help practitioners understand AI’s capabilities and limitations. Further research is needed to explore the long-term impacts of AI scribes on clinical decision-making, clinician–patient communication and health record quality. Questions remain about sustained use, patient trust and whether clerical tasks alleviate burnout over time. While AI scribes show promise in reducing documentation burden in primary care, their integration must be guided by the medical community to balance efficiency gains with the core values of patient centred care. Acknowledgements We thank all participants contributing to the study, as well as associate professor Stefán Hjörleifsson for his feedback on the initial draft of the paper. Funding Statement This study was supported by the Research Committee at the Department of Global Public Health and Primary Care, University of Bergen, Norway. Ethics statement The study’s handling of personal data was approved by the Norwegian Agency for Shared Services in Education and Research (SIKT) (ref. 984639). The project adheres to guidelines from the National Committee for Research Ethics in Social Sciences and the Humanities (NESH). Participants were fully informed and provided written and oral consent. Generative AI statement Three AI tools were used in this study. First, audio recordings were transcribed using WhisperAI (OpenAI, 2024), run locally on a personal computer via Python. No audio data was transmitted to external servers at any point. Second, ChatGPT-5 (OpenAI, 2025) was used exclusively in Step 3 of the Systematic Text Condensation process to assist in drafting written condensates for each subcategory. The identification of meaning units, coding, and recontextualisation (Steps 1, 2, and 4) were conducted by the authors without AI assistance. All AI-generated condensates were reviewed and corrected by DN and JB. Third, ChatGPT (OpenAI, 2025) was used to assist with linguistic refinement of the manuscript. All content decisions, interpretations, and conclusions are solely the responsibility of the authors. Disclosure statement Dr. Damoun Nassehi has investments in global index funds that include shares in AI companies. References