Referential Drift

Referential drift is a failure mode in AI-generated sign language where spatial loci — the established positions in signing space used to refer to people, objects, or locations — shift or are not maintained consistently across a sentence. Because signed languages use spatial persistence to track referents (a signer points to a location, then later verbs and pronouns agree with that location), drifting pointing gestures and unstable classifier placement cause viewers to lose track of who or what is being discussed. Frame-level similarity metrics such as SSIM do not detect referential drift because individual frames may still look plausible; detecting it requires evaluating spatial coherence across the full utterance.