Engineers evaluate earthquake early warning systems by detection accuracy: did the system correctly identify the earthquake's magnitude, location, and expected shaking intensity? Users evaluate the same systems by a different metric entirely. Did the warning arrive in time to do something?
Analysis of over 500 social media posts revealed a strong correlation between user trust and alert timeliness. Users who received warnings more than a minute before strong shaking arrived praised the system regardless of technical accuracy. Users who received warnings during or after shaking condemned it regardless of technical accuracy. Timeliness is accuracy in the user's mind.
Note (2026-11-08): the abstract analyzes more than 500 public X posts about one event, Android's Earthquake Alert during the Mw 6.2 Marmara Ereglisi earthquake of April 23, 2025, where the system 'successfully delivered alerts to users with high precision' with over a minute of warning, extracts 42 attributes with LLMs, and finds a strong correlation between user trust and alert timeliness, concluding that 'timeliness is accuracy in the user's mind'. The essay states the findings for earthquake early warning systems in general, says users praised or condemned alerts 'regardless of technical accuracy', and describes a design tension between accuracy and speed; the abstract reports one event in which the alerts were precise, and the tension is not in it.
The engineering and user definitions of success are measuring different things. A technically perfect alert that arrives two seconds before shaking is instrumentally correct but practically useless — there's nothing to do in two seconds. A technically approximate alert that arrives ninety seconds early, even if it overestimates the magnitude, gives time to stop driving, move away from windows, protect children. The user's metric is actionable lead time, not sensor precision.
This creates a design tension. More accurate detection requires more seismic data, which takes more time to accumulate. Faster alerts require less data, which means less accuracy. The engineering community naturally optimizes for accuracy because that's what they measure. The user community naturally selects for speed because that's what they experience.
The system succeeds when the user has time to act. Everything else is instrumentation.