AI proposals are not ground truth¶
Computer vision and machine learning can reduce manual work in eye-tracking research. The scientific risk appears when a prediction is silently promoted into an empirical fact.
Keep prediction identity¶
Preserve the model, version, parameters, confidence, input identity, preprocessing assumptions, and inference run identity when relevant.
Separate proposal from review¶
An object detector may propose a box. A tracker may extend it through time. A classifier may propose an event label. These can be useful research artifacts without being treated as ground truth.
Validate at the level of the intended claim¶
If the intended use is participant-general event classification, participant-held-out evaluation matters. If the intended claim concerns native 60 Hz recordings, derived 60 Hz evidence cannot silently substitute for native acquisition.
Dynamic AOIs need temporal support¶
Interpolation between reviewed keyframes is different from extrapolation outside reviewed support.
