Who owns an AI-assisted report?

A request to give another radiologist access to reporting initially looked like a permissions change.

It was actually a question of clinical ownership.

The situation

VoxelBox involved multiple possible radiologist roles: an internal radiologist supporting BrainSightAI, a hospital radiologist responsible for the patient, a reporting radiologist, and a neurosurgeon consuming the result.

At the time, each report was tied to an individual user. Giving multiple radiologists the same reporting access would create unclear and potentially conflicting ownership.

Two people could act on the same report without a defined model for authorship, authority or sign-off.

The risk

This could create uncertainty around who owned the report, whose version was authoritative, whether one radiologist could overwrite another, whether review notes were clinical documentation, whether a report was draft, preliminary or final, and what the neurosurgeon should treat as the signed interpretation.

The interface could not solve this through role labels alone.

My decision

I recommended that we not give the internal radiologist reporting permission in that release.

The internal team was not actively reporting on behalf of customers, and the larger reporting model had not yet been resolved. Adding the permission would create complexity before there was a real clinical need for it.

The safer release was to preserve a single reporting owner and define the broader workflow separately.

The larger model

I mapped reporting as a sequence of responsibilities:

1. The result is generated.
2. A radiologist reviews the imaging and model output.
3. Findings may be captured as working notes.
4. An authorised reporting radiologist prepares the report.
5. The report is signed or finalised.
6. Other users can view it without creating conflicting ownership.

This separated collaboration from clinical authorship.

[ image — Responsibility diagram ] Six-step responsibility sequence, editorial diagram.

What this demonstrates

The work was not about making the permission system more flexible.

It was about recognising when flexibility would make the clinical model less safe.

Sometimes the correct product decision is to withhold a capability until responsibility is clear.