Do AI Report Generators Actually Save Doctors Time? The 18.29-Second Reality Check

Sam: Wait, so let me get this straight. A hospital brings in an AI tool to draft radiological impressions, assuming it is going to save their exhausted doctors a massive amount of time. But when they actually measure the workflow, the AI drafts increased the editing time from 12.20 s to 18.29 s? And the number of words the doctor had to change jumped from 5.74 words to 12.32 words? If the goal is to reduce clinician burnout, how is adding more editing time a win?

Maya: Exactly. That paradox is the single most interesting finding we are unpacking today. Welcome back to the podcast. Today we are looking at a systematic review published in the Journal of medical Internet research. It assesses the effectiveness, safety, and workflow burden of large language model-based medical report generation. And as always, a quick reminder in our first few lines: our voices are AI-generated, and this is a summary of a public document. But Sam, you hit the nail on the head. We are constantly told that AI will solve the documentation crisis, but this review asks a much harder question. Do these systems actually improve clinical reporting, or do they generate plausible but unsafe text that takes even longer to fix?

Sam: That is the central translational question for every health system leader right now. Everyone is buying or building these tools because clinical documentation is a massive bottleneck. The source document points out that in fields like radiology, pathology, and endoscopy, doctors are under rising volume, workforce pressure, fatigue, and intense time constraints. The review notes that reporting work is a major part of the information-transfer burden in image-based specialties. And these reports are not just paperwork, they are "safety-critical communication artifacts". If an AI makes a mistake, the downstream clinical decisions are compromised. So, how did the researchers set out to measure this?

Maya: They conducted a highly comprehensive systematic review. They searched PubMed/MEDLINE, Embase, Web of Science Core Collection, Scopus, and the Cochrane Library for studies published from January 1, 2016, through May 15, 2026. They started with 6936 records before deduplication. After the removal of 3008 duplicate records and 499 records already represented in the search corpus, 3429 database records proceeded to title and abstract screening. Ultimately, 101 studies met the inclusion criteria for this systematic review. They were looking specifically for studies evaluating large language models, multimodal models, or vision-language foundation models.

Sam: That is a massive sweep of the literature. But with 101 studies, I imagine there is a huge variety in how these AI tools are actually being used. An AI that writes a whole report from scratch before a human even sees it carries a very different risk profile than an AI that just helps format a draft that a doctor already dictated. If you treat those as the exact same intervention, you are going to get completely meaningless data. How did they categorize the different ways AI is being deployed?

Maya: You are exactly right, and the researchers felt the same way. Treating these workflows as a single intervention can obscure important differences in risk and human oversight. So they categorized the AI role into three distinct levels of oversight. First, "autonomous generation" referred to systems that generated reports or outputs from images or clinical inputs before human evaluation. Out of the 101 studies, 66 were classified as autonomous generation, which is 65.3%. Second, "human-supervised drafting" referred to workflows in which clinicians actively used, or iteratively collaborated on, AI-generated text during drafting. There were 15 studies, or 14.9%, in this bucket. Finally, "report editing or structuring support" referred to settings where the AI was applied after the source report text, findings, or draft content was already available. That accounted for 20 studies, or 19.8%.

Sam: So the majority, 66 studies, are testing autonomous generation. That feels very ambitious. Did they also break down what kind of medical data these models were looking at? I assume radiology is the biggest player here?

Maya: Yes, radiology dominates, but it is highly concentrated even within that field. Chest x-ray was the largest modality group with 36 studies, making up 35.6%. Then you have 20 studies, or 19.8%, in mixed-modality settings, and 11 studies, or 10.9%, in computed tomography, computed tomography angiography, or positron emission tomography/CT. There were 10 studies for magnetic resonance imaging, which is 9.9%. Beyond that, they found 6 studies for ultrasound, 5 for endoscopy, 5 in pathology reporting, 5 for other x-ray or radiography, 2 in ophthalmic imaging, and exactly 1 in electrocardiogram.

Sam: It is interesting to see the spread, but chest x-rays clearly dominate the AI conversation. Why is that? Is it just easier for an AI to read a chest x-ray than a pathology slide?

Maya: It is more about the data that is publicly available. The review points to public corpora like MIMIC-CXR. These massive datasets have accelerated model development and benchmarking because developers have easy access to them. But the authors offer a strong warning here. They state that repeated use of the same datasets can increase the risk of overlapping test sets, as well as model and data leakage. Therefore, evidence from templated chest radiography should not be generalized uncritically to more complex reporting domains.

Sam: That makes total sense. A chest x-ray report is very structured, whereas a complex endoscopy or whole-slide pathology report might be far more nuanced. Okay, so we have 101 studies. Before we get into the safety numbers, how were these AI companies and researchers actually proving that their tools worked? What metrics were they using?

Maya: This is where the review gets really critical of the current research landscape. Many studies still emphasize benchmark language metrics. You will see acronyms like BLEU, ROUGE, CIDEr, and BERTScore. These are essentially semantic similarity metrics. They measure how closely the AI's language matches a reference text. But the authors explicitly state that benchmark similarity metrics help technical development but do not show whether a generated report missed a critical lesion, introduced a misleading statement, or reduced physicians’ finalization effort. For this review, the authors demanded clinically interpretable endpoints.

Sam: So basically, a high BLEU score just means the AI sounds like a doctor, not that it is actually practicing safe medicine. How did this systematic review define their clinical endpoints then?

Maya: They broke it down into three clinical pillars: effectiveness, safety, and workflow burden. Effectiveness was defined strictly as expert acceptance and blinded expert preference. Safety was defined as clinically significant error rate, omission error rate, and commission error rate. And workflow burden was defined as reporting time, number of corrections, edit distance, and editing burden.

Sam: Let's pause on omission and commission errors. Just so we are completely clear on the definitions, what exactly is the difference in this context?

Maya: The review defines omission errors as "missed clinically relevant findings". So, the patient has a mass, and the AI report completely leaves it out. Commission errors are defined as "unsupported or hallucinated findings introduced into the generated report". So, the AI invents a finding that is not actually in the imaging. Both "can affect downstream clinical decisions".

Sam: Both of those sound terrifying if you are a patient. So, with those strict definitions in place, how did the 101 studies actually perform? I want to hear about the quality of this evidence. Systematic reviews usually have some kind of grading scale for the risk of bias. How did these stack up?

Maya: This is a major red flag for anyone looking to buy or implement these systems right now. Out of the 101 studies included, 0 were judged at low risk of bias. Zero. There were 15 moderate, 72 high, and 14 serious.

Sam: Zero? Out of 101 published papers on AI report generation, not a single one had a low risk of bias? What is driving that?

Maya: It comes down to study design and transparency. Of the 101 included studies, 68, or 67.3%, were retrospective or technical validation studies. 28, which is 27.7%, were nonrandomized comparative, reader, or workflow studies. And only 5, exactly 5.0%, were prospective studies or real-world workflow or validation studies. When applying their custom AI validation framework, they found recurrent concerns involving validation independence or leakage risk, absent or unclear blinding, heterogeneous outcome definitions, and poor accounting for failed or unusable reports.

Sam: Wait, poor accounting for failed reports? Does that mean if the AI just broke down, hallucinated wildly, or produced total garbage, the researchers might have just tossed that data point out of their final metrics?

Maya: Yes, exactly. The authors noted that serious concern was assigned when a limitation could directly undermine clinical validity, such as "the absence of handling of failed generations when failures could alter outcome interpretation". If you only measure the reports where the AI succeeded, your effectiveness and safety numbers are going to look falsely optimistic. This is why the authors warn that individual positive studies do not yet establish whether the field is approaching clinical readiness or overstating progress through benchmark-dominated evaluation.

Sam: That is incredible. Okay, despite the bias, let's look at the actual safety and effectiveness numbers they did extract. Chest x-rays were the biggest group, with 36 studies. What did the data show there? Can the AI write a safe chest x-ray report?

Maya: In one large clustered reader evaluation of autonomous AI preliminary reports for chest radiographs, the AI looked very competitive on the surface. For the AI-generated reports, acceptance was 70.5%, which represents 6047/8580. That is very close to the radiologist reports, which had an acceptance of 73.3%, or 6288/8580.

Sam: Okay, so the AI is basically tying the humans on acceptability. 70.5% versus 73.3% is a very small gap.

Maya: Close in acceptance, but you have to look at the safety outcomes, not just whether the doctor accepted the draft. For false-negative findings, the AI was slightly higher at 18.5%, which is 1584/8580. The human radiologists were at 17.8%, or 1527/8580. And for false-positive findings, the AI reached 11.3%, which is 971/8580, while the human baseline was 9.7%, or 831/8580. So the AI looks acceptable but introduces slightly more errors across the board.

Sam: That perfectly highlights the danger. A doctor might read an AI draft, think it sounds incredibly fluent and professional, and accept it, without realizing it missed a subtle false-negative. What about the studies where clinicians actively collaborated with the AI, rather than just reviewing an autonomous draft?

Maya: In a clinician-collaboration chest x-ray study, AI reports were equivalent to or preferred over clinician reports in 233 of 300 cases across one dataset. That is 77.7%. In a second dataset, they were equivalent or preferred in 170 of 303 cases, which is 56.1%.

Sam: That sounds like a massive win for the AI. Being preferred in 77.7% of cases is a great sales pitch.

Maya: It is a great sales pitch, yet, as the authors explicitly note, "clinically significant errors persisted" in both the AI-generated and clinician-generated reports. And this brings us to a major clinical risk mentioned in the text: automation bias. Automation bias can reduce human verification of decision-support outputs when users overrely on automated suggestions, especially when verification is complex. The authors state firmly that human preference or acceptance of generated text should not be interpreted as equivalent to safety unless omissions, commissions, clinically significant discrepancies, and failed generations are measured explicitly.

Sam: Right, because if the AI generates a beautifully formatted report that says everything is normal, and you are a fatigued doctor on hour 10 of your shift, you might just hit approve. That is automation bias in action. Now, I want to circle back to the hook of our episode. The workflow burden. Does this technology actually save doctors time?

Maya: The workflow effects were totally mixed depending on the task. In one brain MRI reader study, AI assistance actually improved diagnostic performance and reduced reading time from 61 to 53 seconds.

Sam: Okay, dropping from 61 to 53 seconds per scan might not sound like much, but over hundreds of scans a day, that adds up. So there is some time saved there.

Maya: Yes, but contrast that with the study we mentioned at the top of the show. In a study of radiological impression drafting with multiple readers, the drafts generated by the model required more editing time than the radiologist baseline. Specifically, it took 18.29 versus 12.20 s. And the edit distance was greater too, at 12.32 versus 5.74 words changed.

Sam: For someone not building these tools, what exactly does "edit distance" measure in this context? Why is jumping from 5.74 to 12.32 words a bad thing?

Maya: Edit distance basically measures the physical workload for the physician. It represents the volume of text a clinician has to change to make the AI draft accurate and clinically usable. If the edit distance goes up, it means the doctor is spending more time deleting, retyping, and fixing hallucinations than they would have spent just dictating the impression from scratch. As the source notes, a fluent report can still increase the review burden for the clinician who must verify it.

Sam: That is fascinating. The AI might write three paragraphs in a second, but if the doctor has to spend 18.29 seconds hunting for the one hallucinated word, it is a time tax, not a time saver. Usually, a systematic review pools all of this data together into a meta-analysis to give us a definitive statistical answer. Did they do that here?

Maya: No, they could not. The authors state that meta-analysis was not performed because no comparable outcome had at least 2 studies with compatible task structure and analyzable data. The safety and workflow evidence remained heterogeneous and largely nonpoolable. The evidence is simply too heterogeneous, biased, and sparse on case-level endpoints to support a pooled meta-analysis or any autonomous clinical-readiness claims.

Sam: That is a bit disappointing for anyone hoping for a clear thumbs up or thumbs down. Does the review acknowledge its own limitations in not being able to do a meta-analysis?

Maya: Yes. They acknowledge that structured narrative synthesis provides less statistical compression than meta-analysis, but they argue it better reflects the current nonpoolable evidence structure. They also note that their custom AI validation framework was prespecified and transparently domain-based, but it was not a formally validated measurement instrument. Additionally, they excluded preprints. Because preprints were excluded, this review was designed to evaluate peer-reviewed clinical evidence rather than the absolute frontier of technical model capability.

Sam: That is a fair boundary to set. We want to know what is proven in peer review, not just what a tech company claims on a preprint server. So, Maya, you are the clinician-turned-analyst. We always ask: what does this change on the ward or in the boardroom? What is the practical takeaway for a hospital executive looking to buy one of these LLM report generators?

Maya: For the boardroom, it changes how you evaluate these tools entirely. You cannot accept benchmark metrics like BLEU or BERTScore as proof of safety. The review states clearly that adoption should remain locally validated, clinician-supervised, and accompanied by standardized reporting of acceptance, preference, omissions, commissions, failed generations, reporting time, corrections, and editing burden. You must evaluate the system in your local modality, using your report templates and your patient mix.

Sam: And for the clinicians on the ward? The ones actually using the software?

Maya: For clinicians, the takeaway is that LLM-based reporting systems should be evaluated as supervised assistive tools, not as unsupervised replacements. The current evidence does not justify autonomous replacement of clinicians. Benefit and risk differ by task. An AI tool that drafts an impression from existing findings has a totally different safety profile from an autonomous image-to-report model. Clinicians must remain hyper-vigilant for omission and commission errors, because even if a report looks perfectly fluent, it might just take you 18.29 seconds of editing to make it safe.

Sam: That is incredibly clarifying. The barrier is no longer simply whether models can produce plausible reports. It is whether they are actually safe and actually save time in practice. To quickly recap: we explored a systematic review of 101 studies on LLM-based medical report generation. 0 studies had a low risk of bias. While AI acceptability is high, sometimes hitting 70.5% in chest x-rays, the AI still introduces slightly more false negatives, at 18.5%, compared to human baselines. And crucially, generating a draft does not always save time; it can increase the edit distance and editing burden.

Maya: Perfectly summarized. As always, you can find a link to the full systematic review from the Journal of medical Internet research in our show notes. Please remember that this podcast is for informational purposes only and is not medical advice. Thanks for listening to AI in Medicine - Smart Summaries, and we will catch you next time.