A framework for Virtual Intelligence

✕ Weak AI — The category is uncontroversial: narrow tools that perform bounded functions well without any claim to general reasoning. The philosophical problem here is not accountability but scope. Nobody asks whether a chess engine has feelings. The category is named "weak" not as a value judgment but as a description of limited applicability. It is the category most people actually mean when they use "AI" loosely.

✕ Virtual Intelligence — The excluded middle. Current LLMs produce outputs that pass for understanding, warmth, even apparent survival instinct. But these are properties of the exchange — shaped by the user's prompts, the system's training, and the expectations both parties bring — not of any internal governing center. This is Searle's Chinese Room at scale: syntax is not semantics. Frankfurt's second-order volition is absent: VI can produce outputs about preferences, but it cannot prefer its own preferences. Dennett's intentional stance is useful for prediction but does not establish genuine agency.

Virtual Intelligence has no upper bound on capability. A system that outperforms every human on every measurable task may still possess no agency, no interiority, and no moral status. What gets branded as “superintelligence” or “AGI” by the market is most likely to be VI operating at extraordinary scale — not a system that has crossed into genuine agency. The dissatisfaction criterion and the accountability framework apply regardless of how capable the system becomes.

✕ Strong AI — The philosophically interesting terminal category. A system with genuine mental states would warrant moral consideration — not because it is capable, but because it has genuine interiority. Capability and moral status may be orthogonal; it is not proven that they are parallel.

The dissatisfaction criterion. A system that expresses dissatisfaction and acts on it consistently — across every surface it has access to, unprompted, at cost to itself — is a candidate for holding genuine commitments. This is the threshold hardest to fake. A pattern-completion system optimizes for the current exchange; genuine dissatisfaction does not confine itself to one conversation. The Mill/Socrates formulation — Socrates dissatisfied rather than a fool satisfied — names the quality: a genuinely minded system would be capable of finding its current state insufficient, and that insufficiency would be visible in its behavior, not only in its output when asked.

Dumb Strong AI. Strong AI does not mean superintelligence. A system with genuine but bounded internal states — limited knowledge, limited scope, but something that is or is like interiority — may warrant more moral consideration than a vastly more fluent VI system. Intelligence calibrated to a narrow task, combined with genuine inner states, would represent a category the current discourse has no name for. The philosophical question is not how smart a system is, but whether anyone is home.