> For the complete documentation index, see [llms.txt](https://doc.thordata.com/doc/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://doc.thordata.com/doc/faq/product-problem/multimodal-dataset/why-do-ai-models-need-multimodal-training-data-instead-of-just-text.md).

# Why do AI models need multimodal training data instead of just text?

Because most real-world information is not text-only: models that learn from a single modality cannot interpret the other formats their users actually work with, while multimodal training data teaches them to connect what they see, hear, and read.

An ASR engine trained only on transcripts cannot decode an accent; a video-understanding LLM needs clips paired with subtitles and metadata, not text descriptions alone; an embodied AI agent must link camera frames, sensor readings, and action sequences. Multimodal pairs — audio aligned word-by-word with transcripts, video with captions, images with OCR-labeled text — are what enable cross-modal reasoning, and the alignment quality between modalities matters more than raw volume. This is why frontier labs increasingly source paired multimodal data rather than merging separate single-modality collections.