Performance metrics are necessary but incomplete
A model can perform well on a curated dataset and still be poorly suited to an operational cockpit. Lighting, fatigue, head movement, task phase, individual differences, and sensor quality change what the system sees and what its output means.
The relevant question is not only whether a state can be classified. It is whether the system can support a timely and appropriate decision without creating distraction, overconfidence, or alert fatigue.
Uncertainty should be visible
Operators should be able to distinguish a confident observation from weak or degraded sensing. Systems that compress uncertainty into a single authoritative label can encourage the wrong kind of trust.
Human-centered aviation AI therefore requires interpretable state, sensible fallbacks, data-quality monitoring, and clear responsibility boundaries.
Research and product design must meet
My pilot-vigilance research and software projects share the same principle: sensing, inference, interface design, and operational procedure must be developed together.
The strongest systems will not try to remove people from every decision. They will make human attention more effective where judgment, context, and accountability remain essential.