Inferred attention is too uncertain to serve as a decision criterion

What it is

Attention inference uncertainty arises when cursor, gaze, application focus, or input activity is used to infer latent states such as focused, idle, or understood. Proxy and target are not one-to-one, so the result is an error-prone estimate. It may inform judgment but should not determine access, performance, discipline, or forced interruption.

Why it happens

One signal has many causes: long gaze can indicate comprehension, confusion, or absence; switching can mean distraction or legitimate integration. Models learned in limited settings shift across devices, tasks, disability, and culture. Collapsing probability into a categorical label hides error, and automated action transfers its consequences to the classified person.

Studying it

Define the target and an independent reference label in advance. Report confusion matrices, calibration, uncertainty, and errors across tasks and groups rather than accuracy alone. Experience sampling, replay, and outcomes can triangulate labels but none is perfect. Monitor distribution drift and correction after deployment.

Where it stops holding

Low-risk reversible suggestions tolerate more error; consequential decisions require direct evidence and appeal. Self-report can be stale or pressured, yet personal preference has greater authority than a prediction for interruptibility. Better sensors do not remove ambiguity in the construct.

Applying it

• Display probability language and evidence age rather than a factual mental-state label.
• Permit correction, override, and opt-out without penalty.
• Prohibit direct use of attention proxies for performance, access, or discipline.
• Audit calibration and harm by task and group; degrade to unknown under drift.

Related

• Same group: V2.07.1 Gaze and scroll indicate attended regions · V2.07.2 Attention cues inform interruptibility · V2.07.3 Gaze sharing requires explicit opt-in · V2.07.5 Value rises with work coupling
• Adjacent: V2.08 Availability and willingness to be interrupted · V2.09 Boundary between awareness and surveillance
• Search terms: attention inference · model calibration · proxy label