Automatically inferred status is often wrong
Aliases: automatic presence inference · status misclassification · interruptibility prediction error
What it is
Availability inference error is disagreement between a status automatically derived from keyboard activity, calendar, location, or application focus and the person's actual willingness to be interrupted. Automation reduces maintenance but estimates behavioral proxies, not intent. It can mark a focused person free or leave a reachable person busy.
Why it happens
Sensors miss paper work, conversation, caregiving, and rest and do not know the sender, message, or urgency. Rules shift across devices, roles, and habits. Predictions also alter contact behavior, removing counterfactual evidence: nobody contacting a person marked busy does not prove the label correct.
Studying it
Across real work cycles, collect prediction, contemporaneous self-report, contact, and retrospective appraisal. Report false positives, false negatives, calibration, subgroup error, and consequences. Sampling some message opportunities independently of the model can reduce selection bias but needs strict risk control. Accuracy alone does not capture harm.
Where it stops holding
Self-report also becomes stale and is not perfect ground truth; availability varies by sender and content. Imperfect prediction can support low-risk suggestions, not automatic blocking, escalation, or performance decisions. Mean accuracy can hide persistent harm to particular roles or assistive-technology users.
Applying it
- Mark automatic status as a suggestion with evidence and age, not a user declaration.
- Provide one-step correction, temporary override, and opt-out without penalty.
- Default uncertainty to unknown and offer quiet or delayed delivery.
- Audit false positives, misses, and interruption harm by group; stop publication under persistent bias.