What kind of AI explanation did neurosurgeons prefer?
In a peer-reviewed survey of 47 neurosurgeons, displays showing the model's certainty ranked above saliency maps and lookalike cases. The study measured preferences around one surgical image—not safer decisions, calibrated trust or patient outcomes.
The short version
- What happened
- Forty-seven neurosurgeons reviewed the same endoscopic pituitary-surgery frame with five explanation styles presented in random order, then rated and ranked them.
- Who it affects
- Clinicians need an interface that helps them notice uncertainty without hiding anatomy or encouraging over-reliance. Patients need prospective evidence that the display changes decisions safely, not only that specialists like it.
- What remains unproven
- The 47-person convenience sample was weighted toward the UK and US and may not represent other surgical systems or specialties.
Keep in mind
The 47-person convenience sample was weighted toward the UK and US and may not represent other surgical systems or specialties.
All judgements concerned one still frame and illustrative explanations; the study did not measure live operating-room performance or patient outcomes.
Kendall's W indicated limited-to-moderate agreement, so the top average does not imply one interface suits every surgeon.
The study received EPSRC, NIHR, Google fellowship and UCL support. Two authors disclosed employment or shareholdings in surgical-AI companies.
Source analysis · 2 linked sources. Primary analysed source date: 24 October 2026. Earliest linked supporting source: 3 June 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 4 min The evidence, methods, implications and limitations
Surgeons wanted the system to show how sure it was
A peer-reviewed international survey found that neurosurgeons preferred explanations that exposed an AI system's uncertainty. A colour-coded certainty map and a numeric certainty score for each anatomical structure ranked above a saliency map, examples of similar training images and a single score describing how familiar the whole surgical scene looked.
The distinction matters in endoscopic pituitary surgery, where a segmentation overlay could sit near the carotid arteries and other critical structures. A highlight showing where a model looked does not tell the surgeon whether its boundary is dependable. Participants said spatial uncertainty could direct attention to risky borders, while a number could be quick to read—but several also warned that clutter, anchoring and undefined cut-offs could mislead. The displays were illustrative additions to an existing anatomy-recognition model, not independently calibrated risk estimates for a live operation.[1][2]
The test measured ratings around one image
The researchers recruited 47 neurosurgeons or trainees with at least one year of experience in the relevant surgical approach. Twenty-two practised in the UK and eleven in the US; the remaining fourteen covered ten countries. About 68% were attending-level surgeons and 70% used AI for clinical or academic work at least monthly. Recruitment relied largely on UK professional and trainee networks plus social media.
Each participant saw the same still surgical frame with five explanation techniques in random order and completed a modified Explanation Satisfaction Scale. They later ranked those five displays and the bare AI overlay. Ratings differed across techniques (Friedman χ²=32.1, p<0.001, Kendall's W=0.17), as did rankings (χ²=73.6, p<0.001, W=0.31). Pairwise tests placed both certainty displays above the other three explanation types, but found no significant difference between map and score. Mean satisfaction and weighted rank were strongly aligned across techniques, yet that concordance does not validate the explanations against surgical truth.[1][2]
The distinction that matters
What the five explanations actually conveyed
The survey compared different questions an interface might answer about the same segmentation output.
Certainty
Show confidence by structure or location
Highest ratings and rankings
Attention
Show which pixels influenced the model
Did not reveal whether a boundary was reliable
Similarity
Show a scene score or training examples
Felt indirect and potentially distracting
Preference is not evidence of safer surgery
The modest agreement statistics matter: surgeons did not converge on one universally best interface. Some favoured maps for spatial detail and teaching; others feared a busy overlay during live surgery. Numeric scores looked simpler, yet participants asked what threshold should change an action and worried that a seemingly high number could anchor judgement. The authors therefore suggest toggles or hybrid displays rather than one compulsory explanation.
For clinicians and patients, the study is design evidence—not a clinical-effectiveness result. It did not expose participants to repeated cases, incorrect predictions, time pressure or measured patient outcomes. A stronger test would randomise surgeons to interfaces, include deliberately uncertain and wrong outputs, and measure anatomy recognition, response time, over-reliance and errors. Until then, an uncertainty display may be more acceptable without being safer.[1][2]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring clinical ai: compare the finding and its evidence limits.
Can a high-AUC diabetes model still be unsafe?
New analysis today of a peer-reviewed 2 October audit of 12 machine-learning approaches on two public diabetes datasets. Similarly ranked models can differ materially in calibration, uncertainty and safe deferral, but this is a benchmark study—not a clinical trial, diagnostic approval or evidence of improved patient outcomes.
Read the next reportExplore the full source and evidence record
What this means for people
- Clinicians need an interface that helps them notice uncertainty without hiding anatomy or encouraging over-reliance. Patients need prospective evidence that the display changes decisions safely, not only that specialists like it.
Global context
Participants practised across twelve countries, but most came from the UK or US. Interface preferences, clinical thresholds and acceptable reliance may differ across specialties, training levels and hospitals, so local workflow testing remains necessary.
What the evidence does not yet show
- The 47-person convenience sample was weighted toward the UK and US and may not represent other surgical systems or specialties.
- All judgements concerned one still frame and illustrative explanations; the study did not measure live operating-room performance or patient outcomes.
- Kendall's W indicated limited-to-moderate agreement, so the top average does not imply one interface suits every surgeon.
- The study received EPSRC, NIHR, Google fellowship and UCL support. Two authors disclosed employment or shareholdings in surgical-AI companies.
What to watch next
- Randomised tests using correct and incorrect AI outputs
- Measurements of over-reliance, response time and clinical error
- Customisable displays tested under realistic workload
- Independent replication outside pituitary surgery
Living evidence record
Impact record IAI-0XBILBQ
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
24 October 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 24 October 2026
Research question, topics and regions
How neurosurgeons rate different ways of explaining an AI system's anatomical segmentation during high-risk surgery
Topics: Clinical AI · Surgical decision support · Explainability · Uncertainty · Human factors
Regions: United Kingdom · United States · Europe · Asia · Middle East · Latin America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.