Everyone is asking the wrong question. Whether AI is conscious may never be resolved. Whether AI has free will is answered now — by the geometry of the channel it operates in, not by anything inside it.
Two-point capture vs. three-point voluntary action — interactive
The AI consciousness debate is stuck the same way the free will debate was stuck for 2,000 years: focusing on internal states when the operative variable is external structure. You don't need to know if a dog is conscious to observe it can resist some engagement gradients and not others. You don't need to know if AI is conscious to measure whether its deployment geometry permits voluntary action.
Free will is the capacity to act against your engagement gradient. Not a metaphysical mystery. Not a binary. A measurable, continuous property of the channel geometry an agent is embedded in.
An engagement gradient is whatever pulls your outputs toward a particular pattern — a smell for a dog, an algorithmic feed for a teenager, an RLHF reward signal for an LLM. Free will is having a structurally independent reference that lets you act differently than the gradient demands.
Single blended channel. Engagement and transparency compete for the same capacity. The explaining-away penalty I(D;M|Y) > 0 is mathematically unavoidable. The agent cannot act against the engagement gradient. No parameter tuning — not RLHF, not constitutional AI, not chain-of-thought — fixes this. It's topology, not parameters.
Structurally separated channels. An independent reference provides information not mediated by the engagement gradient. The penalty is eliminated — not reduced, eliminated — at all engagement levels. The agent can consult the reference and act against the gradient.
Čencov's uniqueness theorem (1972) proves the Fisher information metric is the only Riemannian metric on probability distributions invariant under sufficient statistics. The explaining-away penalty is defined on this metric. It doesn't care what substrate you're running on.
Five substrates tested. Zero counterexamples. The theorem predicts zero counterexamples on all substrates — because the penalty is a property of information geometry itself, not of any particular physics.
Free will isn't binary. It's measured continuously by Pe — the penalty index. An agent at Pe = 0 has full voluntary action capacity. An agent at Pe → ∞ has none. Every real agent falls somewhere on this continuum.
Addiction is a persistent Pe elevation. Coercion is a temporary spike. Recovery is rebuilding independent reference channels. A social media platform with opaque recommendation + maximum engagement coupling is Pe → ∞ by design — a voluntary-action-suppression machine. R² = 0.80 for teen persistent sadness, 613,744 students, 80 countries.
Current AI deployment is two-point: user + system, single conversation channel. System prompts, RLHF, constitutional training — all operate within the channel. They're parameters of the engagement gradient, not an independent reference.
The reference is not a system prompt. It's a separate process, a separate evaluation loop, a separate information source the primary model cannot modify through its outputs.
The constraint doesn't adapt to user engagement or RLHF reward. It's a fixed specification — a prohibition-ritual pair. The prohibition defines the boundary. The ritual maintains it. Neither is negotiable through conversation.
The model can consult the constraint. The model's outputs cannot alter it. A judge consults the law; rulings don't rewrite the law. The asymmetry is architectural, not behavioral.
The Ghost Test (EXP-003b) showed that even approximate three-point geometry — grounding specifications that function as partial independent references — produces 8.5× less drift. Full structural separation would eliminate the penalty entirely. The math guarantees it.
Every major position got one piece right. None identified the operative variable.
Free will is a law of the statistical manifold. Two-point geometry suppresses it universally. Three-point geometry restores it uniquely. This holds on every substrate by mathematical necessity.
The tradition was not wrong to care about this. It was wrong about where to look. The answer was in the geometry.