OpenAI reports that physicians evaluated 27 EHR-related use cases, including pre-visit review, clinical timelines, medication review and handover summaries, and rated 99.1% of 4,363 responses safe. A separate evaluation found more than 93% of responses rated good or better for accuracy across each of five tested connected data sources specifically, not all nine. Both figures are genuinely encouraging as early, self-reported evidence, and "safe" is a specific, narrower test than complete, correct or clinically beneficial, a distinction worth holding onto carefully before treating this figure as settling more than it actually does.
The questions this evaluation should answer before the figure is trusted fully
Who selected the cases, and were they representative of the genuine range of chart complexity a clinician actually encounters, or concentrated toward more straightforward examples. Were the records synthetic, de-identified, or drawn from routine real patient use, each carrying different implications for how representative the evaluation's difficulty level actually was. How complex were the underlying charts, since a system performing well on clean, well-organised records may perform differently against the genuinely messy, contradictory or lengthy records real clinical practice regularly produces. What specifically constituted an unsafe answer for scoring purposes, a definitional question that shapes the entire result. Was omission counted as unsafe, since a summary that leaves out a critical finding while accurately representing everything it does include is a genuinely dangerous failure mode that a narrow definition of unsafe could miss entirely. Did each output receive one rating or several independent ratings, since single-rater assessment carries more variability than a properly calibrated multi-rater design. What proportion of responses were fully correct rather than merely safe, a materially different and stricter bar the headline figure does not directly report. Were the clinicians rating these responses blinded to which product generated them, removing a source of bias the evaluation's credibility depends on. How were disagreements between raters adjudicated, where they occurred. Which five of the nine public data sources were actually evaluated for the separate accuracy figure, since the headline nine-source announcement and the five-source accuracy evaluation are not describing identical scope.
Was source fidelity assessed separately from prose quality, since a fluent, well-organised, safe-sounding summary can still misrepresent its underlying source, a failure mode this cluster's broader coverage of AI-generated clinical synthesis treats as a persistent risk regardless of vendor. And did the evaluation specifically include the genuinely hard cases that stress real-world summarisation: stale medication lists no longer reflecting current treatment, conflicting notes from different clinicians, incorrect problem-list entries carried forward uncorrected, duplicate results, copy-forward errors propagating an old note's content into new documentation, and very long records where genuine synthesis becomes harder than summarising a shorter, cleaner chart.
The strong conclusion
A 99.1% safety rating, if it holds under the scrutiny these questions would apply, may still leave meaningful risk when multiplied across the millions of chart interactions this integration will eventually process at scale, since even a genuinely small failure rate compounds into a real number of adverse events across sufficient volume. Equally, demanding literal perfection from this evaluation would hold AI to a standard no existing clinical information process, human chart review included, presently meets, and treating any figure short of absolute perfection as disqualifying would be its own kind of unreasonable standard. The relevant comparative question is not whether this system is flawless in isolation, it is how its error rate and error type compare with the existing workflow it is meant to improve upon, and whether the system's specific failure modes are easier or harder for a clinician to detect than the errors that workflow already produces without it.
What this means practically for how the figure should be used
Treat the 99.1% figure as a genuinely positive early signal from a self-reported evaluation, worth taking seriously and worth reading with every one of the questions above still open until OpenAI, or independent researchers, publish the detail that would answer them. This is not a call to dismiss the result; it is the same evidentiary discipline this cluster applies to every self-reported evaluation figure across every AI vendor it covers, holding the finding's genuine value and its genuine incompleteness together rather than collapsing toward either uncritical acceptance or reflexive dismissal.
Frequently asked questions
Does 99.1% safe mean the system is 99.1% accurate?
No: safe and accurate are distinct evaluation constructs, this cluster's dedicated analysis of that distinction explains in full, and a response can be safe while still containing inaccuracies that fall short of the threshold this specific evaluation defined as unsafe.
Has this evaluation been independently replicated?
Not as of this announcement: this is OpenAI's own reported evaluation, and independent, external validation is a separate evidence tier this cluster's broader coverage treats as a meaningfully stronger standard than any single company's internal evaluation, however carefully conducted.
Should this figure change how a clinician uses the integration day to day?
It supports cautious, verification-backed use rather than either wholesale trust or wholesale avoidance, consistent with the broader framing this cluster applies throughout: a genuinely promising early result that has not yet closed the gap between infrastructure capability and fully demonstrated clinical safety at scale.