Medical records and imaging reports were originally written by physicians for physicians. Patients reading them directly is a relatively new phenomenon—enabled by the widespread adoption of patient portals and 'open notes'—yet the writing style of these documents has largely remained unchanged to accommodate this shift.
A randomized controlled study published in Neurosurgery in October 2026 bluntly states in its background section: many spinal imaging reports exceed recommended health literacy levels, while patients’ electronic access to results grows ever faster and more direct.
This autumn of 2026, patient-facing evidence suddenly became dense: the spinal imaging report randomized trial on October 22; the multicenter breast imaging study on November 2; and the patient-facing tool evaluation on November 5—three major studies landing within two weeks, adding to prior randomized trials. For the first time, there is now substantial human-subject evidence addressing the question: 'Can AI transform incomprehensible reports into comprehensible ones?'
Laid out side-by-side, the evidence falls into three distinct accounts: comprehension, preference, and trust. The answers across these three accounts are not identical.
## Account One: Comprehension—Real Gains, with Clear Boundaries
The strongest evidence comes from two randomized trials enrolling real patients. In a trial published in Arthroscopy in November 2025, an NYU orthopedic team randomized 64 outpatient patients to receive either knee MRI reports rewritten by ChatGPT-4 or original radiologist-written reports. The AI group scored 20 percentage points higher on comprehension tests (74.7% vs. 54.7%, P < .001). A Columbia University spinal imaging trial (52 enrolled, 42 completed surveys) used ChatGPT-4.0 to rewrite spinal imaging reports, with researchers verifying accuracy before releasing the simplified versions as supplemental material alongside original portal reports. The rewritten versions reduced the Flesch-Kincaid Grade Level from 11.0 to 5.9; patients rated the rewritten version as 'easy to understand' at 4.2 versus 2.6 for originals (Bonferroni-corrected P = .02).
Larger in scale is the multicenter prospective study published in Insights Imaging on November 2, 2026: 314 real patients across Italy, Switzerland, and San Marino performed 2,965 blinded paired evaluations of 10 breast imaging reports—each rewritten by ChatGPT-4 into ≤50 words and pre-reviewed by a panel of 20 physicians. The AI-rewritten versions were selected as 'preferred' in 63.3% of evaluations; high scores (4–5) for 'simplicity' and 'understandability' were 70.0% vs. 43.4% and 68.1% vs. 40.4%, respectively.
But a necessary reality check came from a prospective randomized crossover study at the University of Tübingen in February 2026: 48 trained standardized patients read discharge letters generated by GPT-4 for patients. Accuracy in recalling safety-critical information rose from 38.5% to 45.7%; omission rates fell from 34.8% to 24.3% (OR 1.74). Gains were concentrated in 'medication' (66.9% vs. 50.8%) and 'organizational arrangements', but improvements were minimal for higher-order concepts like 'complication prevention' and 'lifestyle management'. Even after reading the patient-friendly letter, 24.3% of critical information remained unrecognized. In other words: AI rewriting excels at making information easier to *remember*, but still struggles to make it easier to *understand*.
## Account Two: Preference—Patients Welcome Clarity, Not Machines
In the multicenter breast study, the AI version was preferred in 63.3% of cases; multivariable analysis showed 'ease of understanding' was the strongest predictor of preference. In the NY MRI study, 87.5% of patients preferred the AI version (only 4.7% preferred the original), with clarity scores of 4.5 vs. 2.2. What patients clearly want is clarity.
Yet a single-blind, within-subject trial published in Clinical Practice by a Croatian team in November 2025 drew boundaries around preference: 129 emergency ophthalmology patients blindly evaluated GPT-5–generated and physician-handwritten versions of the same clinical facts. GPT-5 texts received significantly lower empathy scores (3.97 vs. 4.30; difference −0.33, 95% CI −0.44 to −0.22, p < 0.001); no other dimensions differed significantly after multiple-comparison correction; overall preference split nearly evenly (47.8% chose GPT-5 among those expressing preference). An earlier Yale study (August 2024, 156 crowd-sourced participants representing the general patient population, five simulated emergency scenarios) similarly found GPT-4–generated discharge instructions were significantly superior only for 'return-to-hospital instructions', with all other sections non-inferior to standard versions. Taken together, the picture is clear: linguistic clarity can be outsourced to models—but empathy and trust cannot. This is precisely why 'human-in-the-loop' (physician review and contextual tailoring of AI drafts) repeatedly emerges as a key conclusion across studies.
## Account Three: Trust—Least Discussed, Yet Decisive for Correct Use
A disquieting detail hides in the NY study: patients rated the AI version as clearer and preferred it—but trusted it less (4.1 vs. 4.5, P = .003). Greater comprehensibility and lower trust coexist.
Meanwhile, patients are already using AI themselves. An April 2025 proof-of-concept study by the OpenNotes team (Beth Israel Deaconess Medical Center) asked patients to query a real neuro-oncology outpatient note using commercially available large language models of their choice (ChatGPT-4o, Claude 3 Opus, Gemini 1.5). 'Persona'-style prompting (instructing the model to first assume the role of someone who understands the patient’s situation) yielded the highest-quality responses—but all models scored poorly on 'evidence use': patient-initiated AI consultations produced fluent, on-topic answers lacking evidence anchoring. A November 5, 2026 evaluation published in Scientific Reports by an MIT team offered another form: a GPT-4–based tool generating an 'enhancement layer' for breast oncology notes—simplified text, term explanations, FAQs, key takeaways, and action items. Among 200 healthy female participants, self-reported confidence (2.83 vs. 2.18) and comprehension test scores (0.63 vs. 0.54) both increased significantly (both p < 0.001). In interviews with seven participants with prior breast cancer history testing the tool on their *own* real medical records, improved understanding co-occurred with heightened concern.
Patients are not passive endpoints receiving rewritten documents. They are active agents already using AI—and forming their own judgments about it. The implication of the 'trust account' is therefore clear: systems should provide patients with *vetted, enhanced materials*, not push them toward unsupervised, unfiltered AI use.
## A Methodological Warning: Simulated Patients Cannot Replace Real Patients
A multi-model evaluation published in JMIR in October 2026 compared four contemporary LLMs (GPT-4.1, GPT-5.2, Claude Sonnet 4.5, Opus 4.6) simulating patient evaluations of discharge summaries using persona prompting, against 99 real patients. Models approximated human performance on factual comprehension questions—but performed unreliably on 'perceptual' items (subjective understanding, uncertainty perception). On factual items, each model aligned more closely with the high-education subgroup; LLMs also systematically underused 'I don’t know'—tending to supply confident answers where humans expressed uncertainty. The authors concluded: persona-based simulation can serve only as an auxiliary, preliminary screen for patient-facing materials—and must always be complemented by, never substituted for, real-patient evaluation.
## Integrating the Three Accounts
For healthcare systems: AI rewriting should launch as 'supplemental material + physician review' (the design adopted in both the spinal trial and the multicenter breast study—where physician review serves explicitly as a safeguard against AI-introduced errors). The goal is comprehension—not replacement of the original report. Before deployment, test with real patients—not LLM-simulated ones.
For clinicians: Outsource clarity; retain empathy. AI drafts lower the reading barrier; explanation, reassurance, and shared decision-making remain fundamentally human work.
For policymakers and procurement officers: Attend to educational gradients. The NY study found education level predicted comprehension scores; the breast study found higher education correlated with greater AI-version preference—meaning the populations most in need of plain-language materials may benefit least. Equity must be embedded in acceptance criteria.
For patients: You have the right to a report you can understand. If you’re already using ChatGPT to read your own medical records, treat the model’s output as 'a better list of questions to ask'—not as the final answer.
Uncertainties must also be transparent: existing randomized studies have small samples (64, 52, 129 participants), mostly single-center; the finding of 'lower trust' currently stems from only the NY randomized trial and warrants replication; the Tübingen study used trained standardized patients—not real patients; the MIT tool’s 200 participants were healthy women; no study has measured downstream outcomes—medication errors, re-presentations, adherence—to determine whether rewriting actually improves them; and rapid model iteration means conclusions based on GPT-4/GPT-5 today have limited shelf life. Yet the direction is unmistakable: making reports understandable to patients has shifted, for the first time, from 'something we ought to do' to 'something we now have evidence for—and can operationalize.'
One final question worth debating: When 'more understandable' and 'less trusted' coexist, should hospital portals explicitly label summaries as 'AI-generated and physician-reviewed'? Might transparency erode the gains in clarity—or might it instead be the first step toward rebuilding trust?
Source: Magni V, et al. Large language model-based simplification of breast imaging reports: a prospective multicentre study with 2965 patient-based evaluations. Insights Imaging, 2026. https://doi.org/10.1186/s13244-026-02423-0