Foundation Models Do Not Understand Biology

Imagine an AI medical tool that can write a beautiful diagnostic report in seconds, which is clean, confident, perfectly structured, and then say something completely impossible, like a blood parasite living inside a tissue sample where it simply could not exist. That’s the strange thing about modern multimodal LLMs. On one hand, they can do genuinely impressive work: pull together clinical details, write clear summaries, spot patterns, and help overstretched healthcare workers move faster. In places where training or specialist support is limited, that kind of help could matter a lot. But there’s a catch. A serious one. Image Generated using Gemini These models don’t actually understand biology. They don’t reason from anatomy, physics, or disease mechanisms. They generate answers by predicting which words are most likely to come next. So when a model talks about a cell, it doesn’t know what a cell is in any real-world sense. It only knows how the word usually appears in medical writin