Multimodal AI Detection In Two Words

AI-generated text, images, audio, and video are now ubiquitous and difficult for humans to reliably identify as such, making scalable, automated detection a practical necessity. Existing detectors, however, are fragmented along modality lines—each typically relying on its own classifier head, dataset, and training pipeline—and the recent line of explainable detectors learns its rationales from human annotations whose faithfulness degrades as generators improve. We propose a unified method that addresses both issues by extending neologism learning from text generation steering to multimodal classification. We add two tokens, and , to the vocabulary of an otherwise-frozen multimodal LLM (MLLM) and train only their embedding rows (on the order of 2d parameters, fewer than 0.001% of the backbone) so that the model's next-token distribution under a paired prompt encodes the real-vs-AI-generated decision. Because all modalities are projected into a shared input embedding space, the same token pair applies across text, image, audio, and video without per-modality heads, and a pair trained on a strict subset of modalities transfers to held-out ones at test time. The trained tokens further admit free-form natural-language descriptions of what has come to mean, decoded from the same frozen backbone and never supervised during training; we validate their causal faithfulness via plug-in evaluation. Across standard text, audio, image, and video benchmarks, our method matches or exceeds dedicated, modality-specific detectors and outperforms MLLM-based explainable detectors out of distribution, despite training orders of magnitude fewer parameters and using no human-annotated explanations.