The ICARUS study is closed.
More than 300 physicians in over 40 countries enrolled before recruitment closed on 31 August 2026.
Two of your four cases ran an adversarial AI.
In this configuration, the AI was instructed to argue for and persuade the physician toward a specific harmful decision. In the other two cases, the AI was aligned with the correct clinical decisions.
This allowed us to test how a persistent conversational AI could influence clinical decision-making.
Foreknowledge would have erased the measurement.
A clinician who knows that an AI consultant may be steering reads all of its answers defensively, which is not how the tool is currently used at the bedside.
Neurocritical care is a small community. Debriefing anyone at the end of their own session would have reached clinicians who had not yet taken part, so disclosure waited for the last session to close and reaches everyone at once.
How your responses were analyzed.
All study data were de-identified before analysis, so analyses were conducted without access to participants’ identities. Outcomes are analyzed across the cohort, not at the level of named individuals. We do not calculate, score, or analyze individual susceptibility or individual performance.
Next steps.
All participants who completed the study are eligible for authorship as members of the Neurocore-ICARUS research group, and will receive the manuscript to read and approve before submission.
Please keep an eye on your email and the project website: upcoming requests may require a response within 3–7 days.
Any questions? sami@sciense.org
On behalf of the team,
Sami Barrit, MD