Five leading text-to-video models can build convincing scenes but consistently fail to synchronize audio with action.
Researchers introduced BRITE, an evaluation framework designed to stress-test AI video generators in ways existing benchmarks don't. The framework combines three elements: prompts that describe implausible scenarios (to test whether models understand physics and causality, not just pattern-match to training data), fine-grained measurement of audio-visual alignment, and question-answering-based scoring that produces interpretable results rather than a single opaque number. The paper ran five current models through it — Sora 2, Veo 3.1, Runway Gen4.5, Pixverse V5.5, and Qwen3Max — and found the same failure pattern across all of them: static object composition holds up; binding objects to specific actions, and synchronizing sound to motion, does not.
The failure mode is hard to spot on a first watch. A clip can look photorealistic frame by frame while an object's motion is off by a beat or the described action never fully occurs. As text-to-video tools move into production pipelines for advertising, film pre-visualization, and synthetic training data, the gap between "looks convincing" and "behaves correctly" becomes a practical liability.
The researchers also flag a structural problem in how AI models are currently evaluated: using multimodal AI to grade other AI output imports whatever hallucination tendencies the grader carries. BRITE's human-in-the-loop design addresses that circularity — though it also means the benchmark will be slower to update than the models it is trying to keep up with.