Video AI chatbots love slapping a timestamp on their answers, as if pointing to a frame proves the claim. A new arXiv paper shows that proof is often theater.
Researchers built a self-verification pipeline for video question-answering systems. A retrieval-augmented language model drafts an answer with a timestamp citing the frame that supposedly backs it up, then a second stage independently checks whether that frame actually supports the claim. The team tested three ways to do that checking: asking the vision model directly whether a frame backs its own claim caught zero of 40 fabricated claims, because the model just agrees with itself. A blind re-captioning step paired with a general-purpose LLM judge did better but was erratic, swinging from flagging nothing to flagging everything depending on how the prompt was worded. Swapping that judge for a small natural language inference model produced a stable verifier that caught 79% of fabricated claims in adversarial tests while leaving correct claims alone. The team ran the pipeline on both Apple Silicon and Google Colab and released the code.
This matters because a timestamp is a trust signal, not a truth signal. Users are more likely to believe a claim when it comes with a citation, even when that citation is wrong, and video-LLMs currently exploit that gap for free. The fix here is refreshingly cheap: instead of building a bigger judge model, a small, boring NLI classifier outperforms a general LLM asked to grade its own homework.
Text-based LLMs have had a citation-hallucination problem for years; this is the video equivalent showing up on schedule, and a 79% catch rate means one in five made-up citations still gets through.