
AI scribes: The impact on the consultation, the notes and relationships in practice
Today, we’re speaking to Dr Emma Ladds, a GP partner and DPhil candidate at the Nuffield Department of Primary Care Health Sciences at the University of Oxford. We’re here today to discuss the recent article she and her colleagues have written for the journal, titled, ‘Ambient scribes in general practice — help or hindrance?’ Title of paper: Ambient scribes in general practice — help or hindrance? Available at: https://doi.org/10.3399/BJGP.2026.0097 Transcript This transcript was generated using AI and has not been reviewed for accuracy. Please be aware it may contain errors or omissions. Speaker A 00:00:00.320 - 00:01:15.600 Hi and welcome back to BJGP interviews after our summer off. I'm Nada Khan and I'm one of the Associate editors of the Journal. Thanks for listening to this podcast today. In today's episode, we're talking to Dr. Emma lads. Emma is a GP partner and a DPHIL candidate at the Nuffield Department of Primary Care Health Sciences at the University of Oxford. We're here today to talk about the recent article that she and her colleagues have written for the analysis section of the Journal. And the article is titled Ambient Scribes in General Practice. Help or Hindrance? So, hi again, Emma. It's lovely to speak to you about this work. And I think one of the reasons we were interested in picking up this article is that it's really highly topical and it's in an area that I think a lot of people are wondering what to do in their day to day practice. And I think that some of the issues that you've highlighted in the article do speak to some of the concerns that people are raising, but also some of the challenges with the implementation in practice. But I guess before we get onto some of the issues that you raise in the paper, can you just explain what we mean by an AI or ambient scribe? Speaker B 00:01:16.160 - 00:01:45.660 Well, it's the technologies based on artificial intelligence that detects conversations, so between the patient and gp, or between GP and other colleagues. And in consultations, it then generates a summary of what's been discussed, but also does some other things as well. So, for example, it might add some coding labels for specific things that are mentioned, and then all of that goes into the patient's electronic health record. Speaker A 00:01:45.980 - 00:02:15.130 And I think that these tools are often presented, at least to some of the GP colleagues that I know and work with, as almost like an obvious win for general practice. So less documentation, more efficiency, and maybe more time with patients. But what made you want to just take a pause and think about some of the assumptions behind some of those quick wins that people might be thinking about, thinking about. Speaker B 00:02:15.450 - 00:03:31.980 And I mean, I think it's interesting, isn't it, because ever since we started to introduce different technologies into the consultations, there's often been an assumption that every new thing would be a quick win. And even since the introduction of the computer, you make for greater efficiency and it will be easier to store data, to retrieve data, to share data, et cetera. And of course, a lot of the time what those technologies do is they drive particular activities. I think it's been interesting just to think about how introduction of different technologies and platforms and digital approaches has driven a much more transactional, quantifiable way of working. And those things that can be recorded and can be captured in technologies or enabled by technologies have been promoted, perhaps at the expense of some of the less quantifiable, more nuanced activity. So suppose it was looking at that journey of technologies and just trying to think about what this next step might also add to that picture. Speaker A 00:03:32.220 - 00:03:41.820 And one of the things you talk about in the article is that AI scribes could create what you call false efficiencies. But can you unpick what you mean by that? Speaker B 00:03:42.300 - 00:04:53.150 I think there's this assumption that actually the summaries and the codes that AI scribes will generate will be very time saving for gps. And no GP ever has enough time. So everybody's looking for, as you put it, the quick wins. And some of the challenges with the outputs from AI inscribes are that they often produce very long transcripts because they're trying to capture a summary of everything that was discussed and that gets entered into the record. And obviously it takes time for subsequent clinicians to look through those kind of summaries. So that could be a fold sufficiency. But also, I mean, it can add false labels or generate errors and that requires checking as well. So it's not that it just produces an accurate summary and then that's out there. Actually, the clinician still has to go back and double check that everything's been done correctly. So even though they don't have to do the actual typing, there's still a level of processing that they have to do of that summary. So it's not necessarily a time saving tool, really. Speaker A 00:04:53.630 - 00:05:15.390 And I think that's one of the interesting things about it because, yeah, there's that interesting point that, you know, GPS might be getting that time back in some other way, but I guess that's highlighting that that might not actually be quite that simple, that the time that you might get back from not having to type out the notes from a consultation is actually being used elsewhere. Speaker B 00:05:15.390 - 00:06:20.920 Actually, I think that's very true. And I think the other thing that's worth thinking about is what's in the notes, the actual information that's there. Yes, it's a summary of the consultation, but actually generating that summary when the clinician does it, there's a level of processing, there's a level of sort of cognitive thinking, of reflection, of thinking about what it was that was happening during the encounter between patient and gp, what was actually being said, or a lot of the time, what was not being said, and the AI scribe won't detect that. And a lot of clinical reasoning sometimes happens for gps, sometimes happens behind the scenes when they're doing that retrospective processing. And as one of my colleagues said, you know, she often thinks about other investigations she'd like to add or things that she might not have made overt to the patient, which perhaps need to be made over to the. So it's those sorts of subtleties that aren't captured in just a very superficial representation of a consultation. That's the content of a consultation. Speaker A 00:06:21.720 - 00:06:53.030 And I think that struck me that taking notes isn't simply an administrative task, it's actually part of clinical thinking. And I certainly, I've tried AI scribes, but how I do my consultations is I write on a piece of paper because I think that's how I process it, and then I move from the paper to the notes and that's part of how I work through the problem. And I guess that's what you're getting at here just in terms of sort of the clinical thinking about actually putting the notes onto the record. Is that what you're saying? Speaker B 00:06:53.350 - 00:08:00.370 Yeah, I think that's right. And I think it's a reflection about how we all develop as individual practitioners as well. I mean, the way you do your consultations may be slightly different to the way that I do my consultations, and the way that you write your notes may be slightly different to the way that I write my notes. But I think often when you've worked together with people for quite a long time, you learn hidden meanings and you pick up that they may phrase things in very particular ways and that conveys a deeper level of meaning to you because of that kind of relational knowledge that you have of your colleagues. So for example, if I write one consultation in my note, my partners know that it was an extremely long consultation that was probably quite circular. We probably didn't come to a very conclusive outcome and probably the patient is going to want to come back and see me again rather than one of them. You know, there are these subtleties in communication that actually at the moment the AI scribes aren't good enough or nuanced enough to be able to detect. And I don't know if they'll ever get to that kind of level of something. Speaker A 00:08:01.130 - 00:08:28.010 But I suppose an AI scribe is never going to really capture clinicians voice, is it? And I think that's what I see when I see some colleagues who are using AI scribes that the. There seems to be quite a lot of detail there which is great, but it doesn't have that clinician's voice in it. So it's hard to really unpick some of the hidden meaning behind sort of what's going on there. And I think that's why I don't use it, because I don't feel it reflects my thought process when I'm going back to look at notes. Speaker B 00:08:28.150 - 00:09:41.210 Yeah, yeah, no, I think that's exactly right. And I think it's extremely good at trans. At sort of capturing transactional interactions and transactional material. And I think that during my detail, I was observing a GP using one of the AI scribes, and we. I watched a series of consultations and in one, which was a very transactional discussion phone call with a patient where he was just adjusting her medication doses, he said, oh, yes, the scribe will be very good in this one. I'll just use the scribe. And it was. It was very accurate. It captured exactly what they. What they discussed. And then he had a consultation with a very complex patient with a number of comorbidities and lots of psychosocial difficulties in the background. And he actually turned the scribe off before the consultation because he just said, it won't capture this. This not the sort of thing it will be any good at. And I just thought that was very interesting, that whilst it is very good at capturing that more not simple is the wrong word, but you know, that. That much more transactional encounter, it just can't quite cope with the relational stuff that often, I think, is GP says now our bread and butter. Speaker A 00:09:41.370 - 00:09:56.020 And I think you make that point in the paper, that a consultation is sometimes much more than just the word, words that are spoken. So looking at those examples, what do you think that an AI scribe might struggle to capture in that sort of more complex or relational type of care? Speaker B 00:09:56.820 - 00:10:51.900 I mean, I think that kind of complex care, there are so many uncertainties, aren't there, ranging from sort of diagnostic uncertainty to prognostic uncertainty. There's often uncertainty, I find, in thinking about what my patient is doing, feeling about me, you know, that level of kind of psychodynamic uncertainty. And I think all of that sort of stuff is likely to be overlooked by the AI scribe. And if you take a very simple example, I mean, often in general practice we're going through a sort of process of diagnostic reasoning where we're weeding out diagnoses, or you'll do a particular investigation thinking, well, that will exclude X, but sometimes the AI scribe might actually put X into the coding as a diagnosis. There isn't that well, this is possible to be captured. So I think that can be another sort of potentially problem as well. Speaker A 00:10:52.140 - 00:11:25.470 Yeah, you talk about that quite nicely in the paper that, about this uncertainty because general practice is often about symptoms that don't fit neatly into a diagnostic category or you're wondering about a diagnosis so you're querying it as you continue with investigations and more consultations and things. And I guess that does create a challenge for AI tools that sometimes seem more designed to classify and summarize and structure information a bit differently. So I thought the example that you gave in the paper was actually very nicely done. Speaker B 00:11:25.950 - 00:12:55.130 There's a lot of talk, or there has been over the years and increasingly still in general practice about the cohort of patients who present to us for whom there is no formal diagnostic label that can be applied. And, and that doesn't mean they're not suffering and they don't have trouble and that they shouldn't be worthy of an appointment. It just means that you can't give them a hard diagnosis at the end and say, well this is X and take Y and it will get better. Actually your, your therapeutic input is through the witnessing of their distress and that kind of human connection. And I guess one of my biggest worries with AI scribes, or not of them per se, but the sort of devoted promotion of them without considering the, than the potential negatives, is that actually because they learn on the kind of data that they're designed to collect, you'll end up with a self perpetuating model where AI scribes will get much better at facilitating transactional consultations and then actually they will just be used within transactional consultations and people will get more and more and more likely to do much, much more transactional consultations in order to use the technologies. And so you'll have technologies driving our values rather than a sort of consideration of what we're actually what we want to promote in general practice, which for me is still about human connection and relationships and I guess sort of bearing witness to the human condition. Speaker A 00:12:55.610 - 00:13:07.300 And having thought about all this, I wonder if you are against the use of AI scribes or is it really about being more thoughtful about how and when we use them. What are your thoughts about that? Speaker B 00:13:07.700 - 00:13:56.690 I think it's that I'm not brilliant with technology, but I'm definitely not somebody who wants to stand in the way of this. I think technology can facilitate our sort of behaviors and our efficiencies enormously. What I don't want it to do is to drive everything. And I think My concern, for example, if you take the executive summary of the table 10 year plan, for example, the word care is used 78 times, but only one of them is as the verb to care. And I just think there's this idea that actually technologies are going to drive models of care or enable faster care, but nobody is actually thinking about what it means to care and how technologies can stand in the way of that sometimes. Speaker A 00:13:56.770 - 00:14:26.270 And I think the way that some people are using AI scribes and how it's being rolled out is a bit patchy. So I know in some practices, for instance mine, There are some GPs who are using it and I know of some practices where it's almost a requirement to use it and that's kind of the, the, the way that they've decided to go forwards. But I think we probably. AI scrubs are likely to become increasingly common. But do you have any thoughts about what good implementation would look like or what would you like to see happen in that? Speaker B 00:14:26.270 - 00:15:57.850 I think I'd like GPC been more involved in the conversations around the development of them and particularly thinking about whether it's possible to enable the technology to capture some of these nuances and subtleties that I've sort of talked about. But I think it's also important for us just to think about what the commercial models and the background is that's driving some of the development of the scribes themselves and what the different agendas might be underlying some of that development. So, for example, you have to think about what the value of data is. I mean, why are people developing such ambient scribes? Actually, it's not for the good of mankind. It's because there's a commercial interest in doing so, and that commercial interest is tied up in data. And so it's in the interests of AI scribe developers often to produce more and more and more data. And so you're looking at more and more and more content. But that might not be the most helpful thing for a practicing gp. You know, actually what we might need is for our ambient stripes to generate a very succinct portrayal of what went on that could be easily skimmed. If you look back at some of the old Lloyd George records, for example, you know, there will be chest infection, amoxicillin, and that's the extent of a consultation. And now think about how lengthy our health records are. And I sometimes wonder how much added value some of that information has brought to things. Speaker A 00:15:58.650 - 00:16:16.570 And I guess in this space, I wonder if you have any thoughts about what you would like GPS to think about before they switch on an AI scribe in their consultation room. So what could be going through someone's mind about the use or good use of it, really, before they think about using it? Speaker B 00:16:17.300 - 00:17:16.280 I think they need to think about what their consultations are for, what are the core activities that they're still trying to enable in their consultation and not to let the AI strive come in the way of them doing that. And so if it is all just about transactions or if it is just about the titration of blood pressure, I kind of argue no consultation is ever just a transaction.
- Transcript