Mayo Clinic says an AI model identified 73 per cent of prediagnostic pancreatic cancers in CT scans taken before the patients were diagnosed.

In its 29 April account, the clinic reported a median lead time of about 16 months. The scans had originally been interpreted as normal and came from multiple institutions, imaging systems and protocols.

That is a promising detection result. It is not yet evidence of a screening programme improving outcomes in routine care. Mayo said the next step was prospective clinical testing to examine integration into care, false positives and clinical outcomes.

The distinction is where trust becomes operational. A useful model can create an alert earlier. However, someone still has to decide what that alert means, whether that's a system alert or something more important, such as detecting cancer cells. Either way, a human must review the evidence to decide what justifies further investigation and balance that against the harm of acting or waiting.

REDMOD Looked for Changes before a Visible Mass

Mayo's Radiomics-based Early Detection Model, REDMOD, analyses quantitative features of tissue texture and structure in abdominal CT scans. It is intended to find subtle changes associated with pancreatic cancer before a visible tumour appears.

The validation included nearly 2,000 CT scans, among them scans from people later diagnosed with pancreatic cancer. Mayo said REDMOD found 73 per cent of the prediagnostic cancers at a median of about 16 months before diagnosis. In scans taken more than two years before diagnosis, it found nearly three times as many early cancers as specialists reviewing the same scans without AI assistance.

Those numbers describe sensitivity in the studied material. They do not tell a future clinic how many alerts would be false, which patients should receive more invasive testing or whether earlier investigation would improve survival. Mayo's prospective AI-PACED study aims to answer those questions in patients at elevated risk.

This is why “the clinician remains responsible” is too thin as a control statement. In prospective use, responsibility would depend on the flag's basis being visible and on the clinician being able to reject it without the case disappearing from the workflow.

Better Assisted Scores Took More time

A Stanford-led randomised trial published in Nature Medicine in February 2025 tested a different problem: whether GPT-4 could assist physicians working through complex patient-management cases.

Ninety-two practising physicians were assigned either conventional resources or those resources plus GPT-4. They worked on five simulated clinical cases derived from de-identified encounters. The AI-assisted group scored 6.5 percentage points higher on expert-developed rubrics.

They also spent an average of 119.3 seconds longer per case.

That result does not show that extra time caused the improvement. It does show that this trial cannot support the assumption that the measured benefit came with faster review. The study assessed reasoning in simulated cases.

The order of Opinions Changed the Interaction

A second randomised trial, published in npj Digital Medicine on 18 March 2026, involved 70 clinicians and up to six clinical vignettes. Clinicians were randomised between two assisted workflows: AI as the first opinion and AI as a second opinion after the clinician had made an initial assessment.

Both assisted workflows scored higher than the conventional-resource baseline, based on initial responses in the second-opinion arm before AI assistance. The researchers found no statistically significant difference in overall performance between the two AI workflows. Their qualitative and post hoc analysis did, however, show that sequence affected the exchange.

When the AI was meant to provide a second opinion, its supposedly independent output sometimes overlapped heavily with diagnoses or next steps the clinician had already provided. The paper also records a case in which a clinician challenged an incorrect AI statement by checking it against the vignette.

These observations do not prove that one sequence is generally superior. The authors described the work as exploratory and used structured vignettes rather than real clinical encounters. They do show why agreement is not enough. Two aligned answers may reflect two sound assessments, one assessment influencing the other, or a person accepting an attractive mistake.

The observed overlap makes provenance consequential. If a supposed second opinion has incorporated the clinician's first assessment, agreement between them is not independent corroboration. Another fluent summary does not become an independent check merely because it appears in a second panel.

Final Perspective

I'd ask the service owner to defend the time allowed for disagreement. The Stanford trial does not prove that extra time caused the higher scores, but it gives no basis for assuming that the measured benefit came with faster review.

If the operating target turns review into instant confirmation, the expert's name may remain beside “approved” while their judgement has lost the power to change the outcome.