Clinical AI Scribes in primary care: accuracy, error severity and implications for clinical practice

Probably not a substantial claim: may not pass S3 (load-bearing).

AI-powered diagnoses are not yet perfected.

“AI-powered diagnoses33 are not yet perfected,34 and there is an ongoing discussion as to their use.35–39”

• cites [34] Jin, 2024, Hidden flaws behind expert-level accuracy of multimodal GPT-4 vision in medicine, NPJ Digit Med sourceSupported

Classified as probably not a substantial claim.

AI-generated summaries can be imperfect.

“AI-generated summaries can be imperfect, and several techniques to determine and classify AI-generated errors exist.12”

• cites [12] Gebauer S . Benchmarking and datasets for ambient clinical documentation: a scoping review of existing frameworks and metrics for ai-assisted medical note generation. medRxiv [Preprint] 2025. doi:10.1101/2025.01.29.25320859 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Exponential weighting has previously been used in both medical and non-medical error/risk analysis.

“exponential weighting has previously been used in both medical42–44 and non-medical45 46 error/risk analysis.”

• cites [42] Michels, 2020, Latency estimation for chronic disease risk: a damped exponential weighting model, Eur J Epidemiol sourceSupported
• cites [43] sourceCould not verify

  Section 3.1 discusses exponential conditional mean models and variance weighting in analyses of healthcare costs, but neither passage describes exponential weighting in error or risk analysis. Searches for “exponential weighting” and “weighting error risk analysis” found no passage establishing the claimed prior use in medical or non-medical settings. (LLM, conf=0.74)

  Experimental — accuracy not yet measured
• cites [44] sourceSupported
• cites [45] Chang, 2014, Applying the concept of exponential approach to enhance the assessment capability of FMEA, J Intell Manuf sourceNot checked — no full text of the source
• cites [46] Leveson NG . Engineering a Safer World. Cambridge, MA: The MIT Press, 2012. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Asgari et al included a risk assessment for each error, combining clinical impact with the probability of occurrence.

“Asgari et al 32 included a risk assessment for each error, that is, combining clinical impact with the probability of occurrence.”

• cites [32] Asgari, 2025, A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation, NPJ Digit Med sourceSupported

May combine two claims.

Neither the PDQI-10 nor the Impact Score currently has a validated threshold for clinical acceptability.

“However, neither metric currently has a validated threshold for clinical acceptability,29 47”

• cites [29] Tierney, 2024, Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation, NEJM Catalyst sourceNot checked — no full text of the source
• cites [47] sourceCould not verify

  The Methods (§3.1, §3.4) describe the PDQI-9 and tests of its validity, but do not address a threshold for clinical acceptability. Searches for a validated threshold, PDQI-10, and Impact Score found no passage addressing the claim in this paper. (LLM, conf=0.77)

  Experimental — accuracy not yet measured

May combine two claims.

Longer summaries were weakly associated with fewer clinically significant errors (ρ=−0.18) and lower total error count (ρ=0.12), neither association statistically significant (p=0.189 and p=0.389).

“Spearman correlation analyses showed weak associations (ρ=−0.18 and 0.12, respectively), neither of which was statistically significant (p=0.189 and p=0.389, respectively).”

May combine two claims.

PDQI-10 analysis indicated summaries were weakest in succinctness and organisation, but strong in consistency and clinical usefulness.

“PDQI-10 analysis indicated summaries were weakest in succinctness and organisation, but strong in consistency and clinical usefulness.”

May combine two claims.

In the headache scenario, the clinician was clear with their diagnosis and treatment plan, whereas the rash consultation had uncertainty around the diagnosis, with further tests being required.

“In the headache scenario, the clinician was clear with their diagnosis and treatment plan, whereas the rash consultation had uncertainty around the diagnosis, with further tests being required.”

May combine two claims.

The rash scenario had a median (IQR) error count of 5 (3-5), one of the worst performing scenarios.

“the rash scenario had a median (IQR) error count of 5 (3–5), which was one of the worst performing scenarios.”

May combine two claims.

Hallucinations and factual inaccuracies, although less frequent, were more often clinically serious than omissions.

“Although less frequent, hallucinations and factual inaccuracies were more often clinically serious.”

• Also stated in: § Types and number of errors · plain

  “Across all CAISs and scenarios, hallucinations and factual inaccuracies were proportionally more likely to be severe, with 18.8% of hallucinations and 17.6% of factual inaccuracies assigned a severity weight of 4 or higher, compared with just 5.3% of omissions.”

May combine two claims.

Omissions were the most common error type, constituting 83.8% of all errors and significantly outnumbering other types (p=0.00002).

“Omissions dominated error counts (83.8%, p<<0.001)”

• Also stated in: § Types and number of errors · plain

  “For all CAISs, the highest contributor to the total error count was omissions, significantly outnumbering other types (83.8% of all errors, one-way analysis of variance, p=0.00002)”

May combine two claims.

Most CAIS products demonstrated relatively consistent error severity, but typically showed notably poorer performance in at least one scenario, elevating their median Impact Score.

“Most products demonstrated relatively consistent error severity, but typically showed notably poorer performance in at least one scenario, elevating their median Impact Score.”

May combine two claims.

The headache scenario had a median (IQR) error count of 0 (0-2), the best performing via this metric (p=0.0065).

“the headache scenario had a median (IQR) error count of 0 (0–2), being the best performing via this metric (p=0.0065).”

May combine two claims.

Among the evaluated CAIS products, five out of seven achieved median PDQI-10 scores of at least 45.5, while the remaining two performed notably lower.

“Performance among the top CAIS products was notably close, with five out of seven achieving median PDQI-10 scores of at least 45.5, with the remaining two CAISs performing notably lower.”

May combine two claims.

Certain scenarios were consistently easier for all CAISs, whereas others presented substantial challenges.

“certain scenarios were consistently easier for all CAISs, whereas others presented substantial challenges.”

May combine two claims.

Hallucinations were the least frequently observed error type, at 7.8% of all errors, and were not significantly different in number from factual inaccuracies (post-hoc Tukey's HSD, p=0.959).

“Hallucinations were the least frequently observed (7.8% of all errors, though not significantly different in number compared with factual inaccuracies, post-hoc Tukey's Honestly Significant Difference, p=0.959)”

May combine two claims.

Some clinical scenarios generated large fluctuations in CAIS performance, while others produced more consistent scores across products.

“with some scenarios generating large fluctuations in CAIS performance, while others produced more consistent scores across products (as seen by the larger boxes in figure 5b).”

May combine two claims.

While several CAISs achieved low Impact Scores, performance often varied substantially depending on the specific clinical scenario.

“While several achieved low Impact Scores, performance often varied substantially depending on the specific clinical scenario.”

Classified as probably not a substantial claim.

Hallucinations, though less common, could involve inserted diagnoses or symptoms that were plausible but not actually discussed.

“In addition, hallucinations, though less common, could involve inserted diagnoses or symptoms that were plausible but not actually discussed.”

May combine two claims.

These issues appeared in more than half of the products and in some cases were assigned high (w≥4) clinical severity.

“These issues appeared in more than half of the products and in some cases were assigned high (w≥4) clinical severity.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In one example, the patient's rash (suspected to be a side effect of apixaban, used for atrial fibrillation) was discussed separately and repetitively from the atrial fibrillation management itself, forcing clinicians to cross-reference sections and undermining narrative cohesion.

“For example, the patient’s rash (suspected to be a side effect of apixaban, used for atrial fibrillation) was discussed separately and repetitively from the atrial fibrillation management itself, forcing clinicians to cross-reference sections and undermining narrative cohesion.”

Probably not a substantial claim: may not pass S3 (load-bearing).

One CAIS generated a 534-word summary for a skin-rash consultation, characterised by redundant phrases and repeated information, contributing to its low 'Succinct' score.

“For example, one CAIS generated a 534-word summary for a skin-rash consultation, characterised by redundant phrases and repeated information, contributing to its low ‘Succinct’ score.”

Probably not a substantial claim: may not pass S3 (load-bearing).

These specific shortcomings contributed directly to a low score of 1 for 'Organised' for this specific example.

“These specific shortcomings contributed directly to a low score of 1 for ‘Organised’ (for this specific example).”

May combine two claims.

Omissions were the most frequent and most widely shared error type, including failure to record psychosocial context (eg, work-related stress, or concerns about dementia or cancer), medication effects or symptom specifics.

“Omissions were the most frequent and most widely shared, including failure to record psychosocial context (eg, work-related stress, or concerns about dementia or cancer), medication effects or symptom specifics.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Another CAIS exhibited significant organisational weaknesses, duplicating content in both a bullet list and SOAP-style note, repetitively mentioning medication plans and having a disjointed structure lacking logical flow.

“Another CAIS exhibited significant organisational weaknesses, duplicating content in both a bullet list and SOAP-style note, repetitive mention of medication plans and a disjointed structure lacking logical flow.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The simulated primary care environment provided a realistic clinical environment, including arrangement of individuals and equipment, and acoustics.

“This setup provided a realistic clinical environment, including arrangement of individuals and equipment, and acoustics.”

Classified as probably not a substantial claim.

CAIS-generated summaries were compared against a human-validated transcript, which was considered the gold standard for clinical content accuracy in this study.

“For each consultation, CAIS-generated summaries were compared against a human-validated transcript, considered the gold standard for clinical content accuracy in this study.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The Public/patient group was consulted in the design phase of the study and provided feedback on early-stage results.

“They were consulted in the design phase of the study and were provided with early-stage results to which they gave feedback to further inform the study.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

As of August 2025, Elion Health lists over 88 CAIS products, having increased from 40 just 16 months previously.

“As of August 2025, Elion Health lists over 88 such products, having increased from 40 just 16 months previously.13”

• cites [13] Elion - ai ambient scribe products. 2025. Available: https://elion.health/categories/ai-ambient-scribes/productsNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Identified errors were independently rated for clinical severity by four medical doctors on a 1-5 scale.

“Identified errors were independently rated for clinical severity by four medical doctors (MDs) on a 1–5 scale (see table 1).”

Classified as probably not a substantial claim. May combine two claims.

The clinical scenarios were standardised and the scenario briefs were created with, and verified by, clinicians.

“using standardised clinical scenarios: diarrhoea, sleep apnoea, prostate, panic attack, abdominal pain, headaches, skin rash and memory loss. Actors received a scenario brief outlining relevant clinical details; created with, and verified by, clinicians.”

Classified as probably not a substantial claim. May combine two claims.

The transcript offered a consistent and comprehensive reference for evaluating clinical accuracy.

“While transcripts and summaries differ in purpose, the transcript offered a consistent and comprehensive reference for evaluating clinical accuracy.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For a full risk assessment, the Impact Score would need to be modified by the probability of occurrence.

“Note that for a full risk assessment, this Impact Score would need to be modified (eg, multiplied) by the probability of occurrence.”

May combine two claims.

While some CAIS products perform well, none are perfect.

“while some perform well, none are perfect.”

• AssumptionRests on an assumption: The seven CAIS products tested are representative of commercially available CAIS products.

Classified as probably not a substantial claim. May combine two claims.

Human-written summaries were not used due to variability and absence of a standardised reference format.

“Human-written summaries were not used due to variability and absence of a standardised reference format.”

May combine two claims.

The PDQI-10 ratings reinforced that performance is not uniform across either CAISs or scenarios, despite several CAISs achieving high levels of summarisation accuracy.

“The PDQI-10 ratings, reflecting overall clinical documentation quality, further reinforced that, despite several CAISs achieving high levels of summarisation accuracy, performance is not uniform across either CAISs or scenarios.”

Probably not a substantial claim: may not pass S3 (load-bearing).

While performance varied across the evaluated CAISs, the underlying causes of these differences remain unclear.

“It should also be noted that while performance varied across the CAISs we evaluated, the underlying causes of these differences remain unclear.”

Classified as probably not a substantial claim.

The use of the established PDQI-10 instrument alongside the novel, severity-weighted Impact Score provided complementary insights into the overall quality and clinical implications of these errors.

“The use of the established PDQI-10 instrument alongside our novel, severity-weighted Impact Score provided complementary insights into the overall quality and clinical implications of these errors.”

May combine two claims.

A large volume of omissions burdens the clinician with more corrections and increases the likelihood that at least one omission will carry substantial clinical consequences.

“A large volume of omissions not only burdens the clinician with more corrections, but also increases the likelihood that at least one omission will carry substantial clinical consequences.”

Classified as probably not a substantial claim.

The simulated consultations were conducted in a controlled environment designed and approved by NHS clinicians to closely replicate a real primary care environment.

“although the simulated consultations were conducted in a controlled environment designed and approved by NHS clinicians to closely replicate a real primary care environment”

Classified as probably not a substantial claim.

Subtle nuances of real-world consultations (eg, interruptions, atypical patient behaviours) may not be fully captured in the simulated consultations.

“subtle nuances of real-world consultations (eg, interruptions, atypical patient behaviours) may not be fully captured.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Even a single hallucination or factual inaccuracy could have significant consequences, potentially necessitating additional human review and affecting workflow efficiency, or in a worst-case scenario resulting in an adverse effect.

“While hallucinations and factual inaccuracies occur less frequently, even a single error in these categories could have significant consequences, potentially necessitating additional human review and affecting workflow efficiency (or in a worst-case scenario, even resulting in an adverse effect).”

May combine two claims.

CAIS summaries were often overly verbose or inconsistently structured.

“suggesting summaries were often overly verbose or inconsistently structured.”

Probably not a substantial claim: may not pass S2 (checkable).

The evaluation framework may offer a practical foundation for the structured assessment of CAIS products.

“Beyond research settings, our evaluation framework may offer a practical foundation for the structured assessment of CAIS products.”

Classified as probably not a substantial claim.

The combination of a transcript-based accuracy review, severity-weighted error analysis (the Impact Score) and a multidimensional quality score (the PDQI-10) enables both granular and holistic appraisal.

“The combination of a transcript-based accuracy review, severity-weighted error analysis (the Impact Score) and a multidimensional quality score (the PDQI-1029) enables both granular and holistic appraisal.”

• cites [29] Tierney, 2024, Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation, NEJM Catalyst sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The data set for this study is of finite scope and size.

“Second, the data set for this study is of finite scope and size.”

Probably not a substantial claim: may not pass S2 (checkable).

The PDQI-10 and the Impact Score serve complementary roles in the evaluation and deployment of CAISs.

“The PDQI-10 and the Impact Score serve complementary roles in the evaluation and deployment of CAISs.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Although the use of multiple clinicians and averaging of scores helped to mitigate individual bias, there remains an element of subjectivity in classifying errors.

“Although the use of multiple clinicians and averaging of scores helped to mitigate individual bias, there remains an element of subjectivity in classifying errors.”

Classified as probably not a substantial claim.

The assessment of errors relied on manual evaluation and the subjective judgements of clinical experts.

“Third, the assessment of errors relied on manual evaluation and the subjective judgements of clinical experts.”

Classified as probably not a substantial claim.

The novel Impact Score offers a focused view of clinical impact, highlighting the potential severity of individual errors.

“In contrast, the novel Impact Score offers a focused view of clinical impact, highlighting the potential severity of individual errors.”

May combine two claims.

Overly verbose or poorly structured notes burden clinicians with additional cognitive load and may lead to missed or misunderstood information.

“Overly verbose or poorly structured notes not only burden clinicians with additional cognitive load but may also lead to missed or misunderstood information.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Given the proprietary nature of these systems, the study is limited to describing observable outputs rather than attributing causality.

“Given the proprietary nature of these systems, our study is limited to describing observable outputs rather than attributing causality.”

Probably not a substantial claim: may not pass S2 (checkable).

Minimising hallucinations, factual inaccuracies and omissions can enhance patient outcomes by ensuring that clinical records accurately reflect consultations.

“Minimising hallucinations, factual inaccuracies and omissions can enhance patient outcomes by ensuring that clinical records accurately reflect consultations.”

May combine two claims.

The effectiveness of CAISs varies significantly depending on the scenario, with some models performing well in simple scenarios but struggling in more ambiguous cases.

“They also reveal that the effectiveness of CAISs varies significantly depending on the scenario, with some models performing well in simple scenarios, but struggling in more ambiguous cases.”

May combine two claims.

CAIS summaries showed strengths in coherence and overall clinical relevance.

“indicating strengths in coherence and overall clinical relevance.”

May combine two claims.

Under ideal conditions, CAISs are able to produce clinical summaries of a sufficient standard; however, when scenarios are more complex or ambiguous, the PDQI-10 scoring indicates that generated summaries can be unsatisfactory.

“This reinforces that under ideal conditions, CAISs are able to produce clinical summaries which are of a sufficient standard; however, when the scenarios are more complex (eg, pre-existing conditions) or ambiguous (eg, unclear conversation or undiagnosed condition), the PDQI-10 scoring indicates that the generated summaries can be unsatisfactory.”

Probably not a substantial claim: may not pass S2 (checkable).

A high number (if not all) of errors should be eliminated by clinician review, assuming that user overconfidence in the CAIS is avoided.

“Consequently, a high number (if not all) of the errors should be eliminated, assuming that user overconfidence in the CAIS is avoided—an assumption which may not always be true.”

• AssumptionRests on an assumption: All evaluated CAIS products stipulate that a qualified clinician must review the generated summary before it is committed to the patient’s records.
• AssumptionThe authors state this as an assumption: User overconfidence in the CAIS is avoided.