AI researchers discover neurons aren't compressing concepts — they're just responding to the same word spelled different ways

Researchers tested whether artificial neural networks in AI models compress multiple unrelated concepts into single neurons (a trait called superposition). They found that most of the apparent overlap is just neurons responding to the same word in different contexts, not actually storing two different meanings in the same place. Filtering out this word-form confusion makes AI better at understanding which meaning of a word is intended, and makes AI edits more precise.

For years, AI researchers have assumed neurons in large language models pack multiple unrelated ideas into single processing units as a compression strategy. This work suggests that assumption has been wrong — what looked like clever compression was mostly just pattern-matching on spelling. This matters because it changes how we think about what's actually happening inside these models and where the real limits of their understanding sit. It also suggests that some of the weirdness in how AI behaves (why edits in one area break things elsewhere) might be easier to fix than previously thought.