How Close Can We Get — Summary

Not in theory. Not in philosophy. Right now, with current approaches, what can AI actually model about human behavior?

The honest answer has tiers. Explicit, stated, routine preferences in structured domains: 85 to 95 percent accuracy. If Margaret says she prefers morning appointments and always responds faster to texts than emails, those patterns are learnable. Implicit preferences inferred from behavioral signals: 65 to 80 percent. When the stated preference diverges from the revealed preference, AI is getting better at noticing the gap but is still imperfect. Contextual adaptation — knowing which version of Margaret is present right now — lands around 55 to 70 percent. Anticipatory reasoning, predicting needs before they are expressed, runs 45 to 65 percent. And deep understanding: the meaning behind a request, the significance of a particular Tuesday, the layered social dynamics of someone who values independence and fears isolation simultaneously — 20 to 40 percent, and probably no higher for the foreseeable future.

These numbers are estimates, not benchmarks. But they frame realistic expectations. The 70 to 80 percent of human behavior that is pattern-based and context-dependent is increasingly within reach. The remaining 20 to 30 percent — contradictions, transformations, meanings — will likely remain irreducibly human. Designing AI systems that are honest about this boundary, that communicate uncertainty, defer when confidence is low, and fail gracefully when approximation breaks down, matters more than pretending the boundary does not exist.

We are past obvious incompetence, not yet at obvious competence. Careful assessment serves better than either hype or despair.