Copilot Returning Model Grounded Answers for Reports in Shared Capacity (Unexpected Access Scope)

We are observing Copilot in Power BI returning semantic‑model‑grounded responses (including DAX calculated results) for reports and models hosted entirely in shared capacity, even when the user has View‑only access. This appears to conflict with current Microsoft documentation around Copilot capacity and  behaviour.

Environment / Setup

• Fabric F4 capacity enabled for a Copilot PoC (all existing reporting is on a shared capacity)
• Test report and semantic model:
  • Both saved in a shared capacity workspace
  • Workspace is not Fabric‑backed
• User permissions:
  • Viewer access to the report
  • No Build permissions on the semantic model
• Copilot accessed via:
  • Copilot pane in the Power BI Service (within the report)

We expected that copilot will only provide responses at the visual level for any artefacts in shared capacity workspaces and that model grounded responses would only be provided where the model is in the fabric workspace.

Believe this is unexpected behaviour or a bug? If it is expected behaviour then the issue of preventing copilot accessing existing models is then our very big issue as there seems to be no way to control access to existing models that are published on the service is we enable copilot for our users..