AI Learns That You're the Same Person Across All Your Data

Machine learning is usually formalized through samples—but a new paper argues that misses the persistent referent those samples come from. The paper proposes making the “unit” an explicit primitive: a learning task declares a population of persistent referents and a sameness criterion. In the supervised setting, what gets learned is a tokenizer that produces a contextual unit token, plus a shared response-law form that reads it. A key formal result: unlinked single-row observations can fail to distinguish a heterogeneous unit world from a homogeneous pooled world, while trusted same-unit pairs can separate a restricted witness. It’s a rigorously grounded case for putting the individual—defined as a unit, not a person—back at the center of learning.