Dr. Stein, I have to disagree on this one. Using an AI scribe has allowed me to be a better listener for my patients and freed up some time during visits to allow me to make better more researched and thoughtful decisions. Many times I have missed key moments during patient conversations which potentially affected clinical decision making. There are several AI scribes to choose from now and you can curate the notes to your own liking-it’s not complicated. As with all technology there are pros and cons. Less hours spent tying up notes and more time listening to my patients is a great thing! :)
Seems like my original response got lost. I just want to say I’m glad you’re enjoying it and it’s working for you! I hope it stays that way and look forward to our tools improving. Thanks for commenting!
A few years ago I was contacted by the NYPD Cold Case Unit, about a murder victim who I had pronounced dead on arrival at a now defunct ED I was working in. As part of the process they sent over my old, handwritten chart, and, my goodness, I could not believe how cogent and well-written it was! We have lost a lot with the adoption of click-based EMRs, and AI charting is just an extension of this process. I have never used an LLM-based scribe, but I have been using human scribes for over 10 years, with mixed results. They are necessary in my work because of the high volume of patients I have to see, but they are, let's just say, not 100% reliable. The bottom line is that we are burnt out because we no longer have the time we need to focus on the primary tasks we are trained to do: see patients at the bedside, perform procedures, order and interpret studies, and then write about our findings and plans of care. AI scribes are just a symptom of a deeper malaise, which is that the modern health care system does not give us the time we need to do what we do. The solution? Just staff up with an adequate number of docs and ancillary providers, and the budget be damned!
In medical school and residency, writing an H&P is what helped me actually process and synthesize the HPI, exam, labs and imaging into a coherent plan. Now as a new attending, I still find significant personal fulfillment from a well written A&P. I hope we can find a way to utilize these AI scribes to improve note bloat and tackle the billing/coding nightmares we currently have in our EMR, but I also find myself concerned that part of what makes internal medicine a cognitively stimulating field may be disappearing. Appreciate your article!
The problem I have with AI scribes is simply that they’re not very good. I have now used two different products and they both suffer from the same problems - namely, they cannot take a patient’s meandering, non-linear story and organize it into a concise, coherent narrative that is relevant to the deductions that the doctor is striving towards. The Assessment and Plan are even worse. This then leads to extensive editing which defeats the purpose of the scribe in the first place. When I brought this up to our organization’s MD “AI champion”, the response was that the note just needs to be “good enough”, because doctors are burnt out. This may be sufficient for some doctors, but I think the standard for the medical profession needs to be higher. When I read a note generated by an AI scribe, it’s obvious. It takes much more cognitive load to decipher. Until the technology improves, for the quality level that I’m comfortable with, the older method of taking a few notes during the visit and dictating is much faster. For all the noise made about AI in medicine, I think scribing is the weakest and least useful application because the notes generated are not just bad, they are counterproductive. Even within just the same area of patient encounters and day to day clinical decision making there are much better uses for AI.
We have a pretty heavily promoted AI program for our visit notes with our patients. Our residencies are not allowing the first year family medicine residents or I believe the second years to use the AI product. I think this is reasonable as they definitely do need to learn how to take histories and make diagnoses. For me as an experienced internist the product has definitely shortened my charting time. I generally scan each note and often do a little bit of typing to edit it. And I do find errors occasionally. I have had a couple of patients message me about errors in the notes and I have made the changes they requested. We do request consent from every patient at every visit.
I have had several recent experiences with excellent physicians using AI scribes. I have found that the AI scribe jumbles the sequence of information and even though I came into my sessions with a written document (given to the physician) of the timeline of development of symptoms, and associated timeline of changes in medications--the AI scribe jumbled up the history. I think this was because we talked about some of these events outside of the historical sequence. In my case this did not lead to any devastating consequences, but I do not trust these systems to do the complex cognitive processing that a human clinician does in creating a timeline from a discussion.
An interesting exposition. I’ve read another on Substack, where the physician tried using a scribe but found his editing took almost as much time as writing the notes, so he dropped it.
My oncologist has just begun using a scribe and had a new user’s enthusiasm for it. I will be interested to hear how it goes in a few months.
Thanks again for a thoughtful and thought-provoking piece.
You raise several interesting and important points.
I note that my HMO is being sued by a group of patients over the fact that they were not informed that a "scribe" note program was used to record their session with their physician on his/her iPhone. Apparently, their action has to do with the fact that they are claiming that their privacy was breached in the summarization process when the recording was uploaded to the AI provider's server. The provider is claiming that the summarization happens without any "human" intervention whatsoever i.e. the summary is written entirely by computer algorithm, and then the voice recording is promptly erased. This morning, bright & early at 0730, my specialist began the session by asking my specific permission to record our ENT session; and by 1315, the "summary was available for my "consumption." My wife - who is an assistant program director of mental health services of a large county health provider - made me aware of this situation as they were developing a consent form in their own agency where their IT Department issues clinicians who ask old phones to record patient sessions that automatically erases recordings after three hours. All of this is to say that, while your comments regarding the need for trainees to learn the skill, subtleties, and nuance of writing clinical notes are well taken, it seems to be the goal of many AI note providers to directly "subvert" the process by providing session summaries that are written purely by AI for consumption by the patient & purely for the EHR, all from the same process recording. Personally, I object to being referred to as "her." And so it goes...
Yes, but to clarify, my point is not about the note as the goal, it’s about what we accomplish when we write the note: synthesis, deeper thinking, and ownership. I sadly suspect that what you described is playing out all over our country.
This really resonates! The “documentation time saved” story is seductive, but you’re naming the deeper variable: ownership of the clinical narrative. For many of us, the note isn’t (only) clerical output; it’s where messy signals get integrated into a coherent model of the patient, and where our memory of the case gets consolidated. If AI scribes convert that moment from synthesis → supervision, we may relieve exhaustion while quietly eroding the parts of medicine that create meaning: judgment, mastery, and longitudinal knowing. Two things I hope health systems don’t miss:
1. Efficiency gains are not benign unless they’re protected. If saved minutes are immediately “re-monetized” into more throughput, we’ll get faster clinics, not better care (and not less burnout).
2. Training risk is real. Crafting a crisp HPI and a thoughtful A/P is a learned cognitive skill. If we outsource the scaffolding too early, we may create a generation that’s great at “reviewing plausible text” but less practiced at clinical reasoning in writing.
A practical middle path I’ve found compelling: let AI help with capture/structure (HPI, ROS scaffolding, meds, vitals), but keep the A/P as the clinician’s authored synthesis and make “protected time” an explicit implementation requirement (e.g., a schedule buffer that cannot be backfilled). Otherwise the technology will succeed at what it’s best at (speed), while we lose what patients actually come for: a mind taking responsibility for their story.
Curious how you’d measure “success” beyond minutes saved; continuity, diagnostic calibration, and patient trust feel like the outcomes that matter most here. 🙂
Very well stated. Being a doctor is partially being a craftsman, and the clinical note is your craft. I am old and sometimes crotchety and take a lot of pride in my notes. Thus, I became a slow convert to the AI-generated note. Part of the reason that I am a convert is that computer-generated text is now the norm, so most physicians are not viewing the clinical note as part of the craft and the note bloat and acronyms make the notes useless. I wrote an essay on the AI-generated note, which is also an homage to Lawrence Weed (https://roberteidus.substack.com/p/the-demise-of-the-clinical-encounter). For me personally, the biggest drawback of the AI-generated note is that the words are not my words (craftsman). The biggest advantage is that it takes me away from the computer and gives me more eyeball to eyeball contact. You did hit the nail on the head however, that the organizations that physicians work for (virtually all are employed who use AI-generated notes) are using it to increase throughput and not to reduce stress
You make me think, is the note necessary? I've always taken it for granted that the note is sacrosanct but that may not be so. Certainly notes have morphed into distorted records with the administrative takeover of medicine.
I certainly think so. As I noted later in the piece, "Documentation has become distorted by billing and compliance, often failing to reflect meaningful clinical thinking. If notes are bloated and templated, automating them may simply acknowledge reality ... but automating hollow documentation does not restore its meaning; it codifies the hollowness. The solution is not to remove clinicians further from the narrative, but to reclaim its purpose."
I share your concerns, especially in regards to trainees. Mastering formulating in writing a concise but clinically comprehensive/relevant HPI, and even more so an A/P takes years from intern to resident to fellow, then junior attending then senior attending. It seems the proverbial cat is already out, now time will tell…personally I use the AI scribe to gather the HPI (then proofread), while the A/P is dictated with thought and nuance…
When I read a heavily templated note it’s the intellectual equivalent to an eating junk food.
I am not optimistic. Those who make the decisions re these technologies have a different agenda. This is where our loss of autonomy most affects us.
Maybe as patients we can have our own personal health record and AI navigator and then “they” consult each other and then the doc and patient can have a nice conversation and mutually work through what the AI generates. I am only half joking. Then the doc writes a “real” note limited to the pertinent issues ie a true synthesis of the clinical encounter that when read later provides the narrative thread to the patients clinical “story”.
It would be very nice if I as a patient had an opportunity to correct the incorrect information in the AI generated notes. In my experience I have not thought that it would be useful to even attempt that, it would just be seen as a nuisance by my excellent physician.
Thank you. I think “skilling“ is a major risk of all AI-assisted use cases, in medicine and outside of it. It’s important to be careful not to lose skills.
Dr. Stein, I have to disagree on this one. Using an AI scribe has allowed me to be a better listener for my patients and freed up some time during visits to allow me to make better more researched and thoughtful decisions. Many times I have missed key moments during patient conversations which potentially affected clinical decision making. There are several AI scribes to choose from now and you can curate the notes to your own liking-it’s not complicated. As with all technology there are pros and cons. Less hours spent tying up notes and more time listening to my patients is a great thing! :)
Seems like my original response got lost. I just want to say I’m glad you’re enjoying it and it’s working for you! I hope it stays that way and look forward to our tools improving. Thanks for commenting!
A few years ago I was contacted by the NYPD Cold Case Unit, about a murder victim who I had pronounced dead on arrival at a now defunct ED I was working in. As part of the process they sent over my old, handwritten chart, and, my goodness, I could not believe how cogent and well-written it was! We have lost a lot with the adoption of click-based EMRs, and AI charting is just an extension of this process. I have never used an LLM-based scribe, but I have been using human scribes for over 10 years, with mixed results. They are necessary in my work because of the high volume of patients I have to see, but they are, let's just say, not 100% reliable. The bottom line is that we are burnt out because we no longer have the time we need to focus on the primary tasks we are trained to do: see patients at the bedside, perform procedures, order and interpret studies, and then write about our findings and plans of care. AI scribes are just a symptom of a deeper malaise, which is that the modern health care system does not give us the time we need to do what we do. The solution? Just staff up with an adequate number of docs and ancillary providers, and the budget be damned!
So well said! Thank you so much.
You might be interested in this new "Perspectives" in today's NEJM, "Can I Say I don't Know?" https://www.nejm.org/doi/pdf/10.1056/NEJMp2517624
That was spectacular! What incredibly insightful writers.
I thought you would appreciate it!
Just saw that and I’m looking forward to reading it! Thanks!
Lots of punchy lines in this piece that really get at the core of corrupting influences in medicine.
Thanks!
In medical school and residency, writing an H&P is what helped me actually process and synthesize the HPI, exam, labs and imaging into a coherent plan. Now as a new attending, I still find significant personal fulfillment from a well written A&P. I hope we can find a way to utilize these AI scribes to improve note bloat and tackle the billing/coding nightmares we currently have in our EMR, but I also find myself concerned that part of what makes internal medicine a cognitively stimulating field may be disappearing. Appreciate your article!
Thank you!
The problem I have with AI scribes is simply that they’re not very good. I have now used two different products and they both suffer from the same problems - namely, they cannot take a patient’s meandering, non-linear story and organize it into a concise, coherent narrative that is relevant to the deductions that the doctor is striving towards. The Assessment and Plan are even worse. This then leads to extensive editing which defeats the purpose of the scribe in the first place. When I brought this up to our organization’s MD “AI champion”, the response was that the note just needs to be “good enough”, because doctors are burnt out. This may be sufficient for some doctors, but I think the standard for the medical profession needs to be higher. When I read a note generated by an AI scribe, it’s obvious. It takes much more cognitive load to decipher. Until the technology improves, for the quality level that I’m comfortable with, the older method of taking a few notes during the visit and dictating is much faster. For all the noise made about AI in medicine, I think scribing is the weakest and least useful application because the notes generated are not just bad, they are counterproductive. Even within just the same area of patient encounters and day to day clinical decision making there are much better uses for AI.
We have a pretty heavily promoted AI program for our visit notes with our patients. Our residencies are not allowing the first year family medicine residents or I believe the second years to use the AI product. I think this is reasonable as they definitely do need to learn how to take histories and make diagnoses. For me as an experienced internist the product has definitely shortened my charting time. I generally scan each note and often do a little bit of typing to edit it. And I do find errors occasionally. I have had a couple of patients message me about errors in the notes and I have made the changes they requested. We do request consent from every patient at every visit.
I have had several recent experiences with excellent physicians using AI scribes. I have found that the AI scribe jumbles the sequence of information and even though I came into my sessions with a written document (given to the physician) of the timeline of development of symptoms, and associated timeline of changes in medications--the AI scribe jumbled up the history. I think this was because we talked about some of these events outside of the historical sequence. In my case this did not lead to any devastating consequences, but I do not trust these systems to do the complex cognitive processing that a human clinician does in creating a timeline from a discussion.
An interesting exposition. I’ve read another on Substack, where the physician tried using a scribe but found his editing took almost as much time as writing the notes, so he dropped it.
My oncologist has just begun using a scribe and had a new user’s enthusiasm for it. I will be interested to hear how it goes in a few months.
Thanks again for a thoughtful and thought-provoking piece.
You raise several interesting and important points.
I note that my HMO is being sued by a group of patients over the fact that they were not informed that a "scribe" note program was used to record their session with their physician on his/her iPhone. Apparently, their action has to do with the fact that they are claiming that their privacy was breached in the summarization process when the recording was uploaded to the AI provider's server. The provider is claiming that the summarization happens without any "human" intervention whatsoever i.e. the summary is written entirely by computer algorithm, and then the voice recording is promptly erased. This morning, bright & early at 0730, my specialist began the session by asking my specific permission to record our ENT session; and by 1315, the "summary was available for my "consumption." My wife - who is an assistant program director of mental health services of a large county health provider - made me aware of this situation as they were developing a consent form in their own agency where their IT Department issues clinicians who ask old phones to record patient sessions that automatically erases recordings after three hours. All of this is to say that, while your comments regarding the need for trainees to learn the skill, subtleties, and nuance of writing clinical notes are well taken, it seems to be the goal of many AI note providers to directly "subvert" the process by providing session summaries that are written purely by AI for consumption by the patient & purely for the EHR, all from the same process recording. Personally, I object to being referred to as "her." And so it goes...
Yes, but to clarify, my point is not about the note as the goal, it’s about what we accomplish when we write the note: synthesis, deeper thinking, and ownership. I sadly suspect that what you described is playing out all over our country.
This really resonates! The “documentation time saved” story is seductive, but you’re naming the deeper variable: ownership of the clinical narrative. For many of us, the note isn’t (only) clerical output; it’s where messy signals get integrated into a coherent model of the patient, and where our memory of the case gets consolidated. If AI scribes convert that moment from synthesis → supervision, we may relieve exhaustion while quietly eroding the parts of medicine that create meaning: judgment, mastery, and longitudinal knowing. Two things I hope health systems don’t miss:
1. Efficiency gains are not benign unless they’re protected. If saved minutes are immediately “re-monetized” into more throughput, we’ll get faster clinics, not better care (and not less burnout).
2. Training risk is real. Crafting a crisp HPI and a thoughtful A/P is a learned cognitive skill. If we outsource the scaffolding too early, we may create a generation that’s great at “reviewing plausible text” but less practiced at clinical reasoning in writing.
A practical middle path I’ve found compelling: let AI help with capture/structure (HPI, ROS scaffolding, meds, vitals), but keep the A/P as the clinician’s authored synthesis and make “protected time” an explicit implementation requirement (e.g., a schedule buffer that cannot be backfilled). Otherwise the technology will succeed at what it’s best at (speed), while we lose what patients actually come for: a mind taking responsibility for their story.
Curious how you’d measure “success” beyond minutes saved; continuity, diagnostic calibration, and patient trust feel like the outcomes that matter most here. 🙂
Very well stated. Being a doctor is partially being a craftsman, and the clinical note is your craft. I am old and sometimes crotchety and take a lot of pride in my notes. Thus, I became a slow convert to the AI-generated note. Part of the reason that I am a convert is that computer-generated text is now the norm, so most physicians are not viewing the clinical note as part of the craft and the note bloat and acronyms make the notes useless. I wrote an essay on the AI-generated note, which is also an homage to Lawrence Weed (https://roberteidus.substack.com/p/the-demise-of-the-clinical-encounter). For me personally, the biggest drawback of the AI-generated note is that the words are not my words (craftsman). The biggest advantage is that it takes me away from the computer and gives me more eyeball to eyeball contact. You did hit the nail on the head however, that the organizations that physicians work for (virtually all are employed who use AI-generated notes) are using it to increase throughput and not to reduce stress
You make me think, is the note necessary? I've always taken it for granted that the note is sacrosanct but that may not be so. Certainly notes have morphed into distorted records with the administrative takeover of medicine.
I certainly think so. As I noted later in the piece, "Documentation has become distorted by billing and compliance, often failing to reflect meaningful clinical thinking. If notes are bloated and templated, automating them may simply acknowledge reality ... but automating hollow documentation does not restore its meaning; it codifies the hollowness. The solution is not to remove clinicians further from the narrative, but to reclaim its purpose."
I share your concerns, especially in regards to trainees. Mastering formulating in writing a concise but clinically comprehensive/relevant HPI, and even more so an A/P takes years from intern to resident to fellow, then junior attending then senior attending. It seems the proverbial cat is already out, now time will tell…personally I use the AI scribe to gather the HPI (then proofread), while the A/P is dictated with thought and nuance…
Well said and agree a 100%.
I abhor templates and don’t use them much.
When I read a heavily templated note it’s the intellectual equivalent to an eating junk food.
I am not optimistic. Those who make the decisions re these technologies have a different agenda. This is where our loss of autonomy most affects us.
Maybe as patients we can have our own personal health record and AI navigator and then “they” consult each other and then the doc and patient can have a nice conversation and mutually work through what the AI generates. I am only half joking. Then the doc writes a “real” note limited to the pertinent issues ie a true synthesis of the clinical encounter that when read later provides the narrative thread to the patients clinical “story”.
It would be very nice if I as a patient had an opportunity to correct the incorrect information in the AI generated notes. In my experience I have not thought that it would be useful to even attempt that, it would just be seen as a nuisance by my excellent physician.
I personally hate all chart errors, but it is a huge nuisance to correct them unless it is really important.
Thank you. I think “skilling“ is a major risk of all AI-assisted use cases, in medicine and outside of it. It’s important to be careful not to lose skills.
Thank you - well said.