Philosophy and Artificial Intelligence

The contemporary debate about artificial intelligence is badly formulated from the outset. Not because the questions currently in circulation are entirely false, but because the criteria organising them are too thin to reach what is actually at stake. When one asks whether a machine "thinks", "understands", or "creates", the almost universal assumption is that thinking, understanding, and creating are properties identifiable from the outside, comparable by analogy with human performance, and measurable by the quality of the output. That assumption generates an unstable field of discussion, trapped between two complementary errors: the enthusiasm that identifies sophisticated performance with genuine intelligence, and the refusal that excludes any non-biological system on the grounds of substrate loyalty. While the debate remains on that axis, the decisive philosophical problem cannot even be formulated.

The productive question is a different one. What separates a system that executes symbolic operations with high efficiency from a system that reorganises itself symbolically on the basis of its own operational history? The relevant difference is not quantitative. Speed, vocabulary size, and parameter count are all beside the point. The difference is one of regime: the manner in which operations are produced, the history from which they arise, the relation between what the system does and what it has accumulated, modified, and reinscribed in itself. Placing this difference at the start of the inquiry transforms the entire map of the problem.

The first move is to separate three levels that ordinary usage tends to collapse into one. There is processing: any system that receives inputs and produces outputs according to rules operates at this level. There is operational intelligence: the capacity to reorganise response patterns when existing ones prove insufficient, generating new functional compatibilities from material inscriptions. And there is functional subjectivity: the regime in which a system operates on its own marks in a self-referential manner, integrating those operations into the history that constitutes it and exposing that history to the real difference introduced by the environment. These three levels are not mere gradations along a homogeneous scale. A system may exhibit highly complex processing without reaching operational intelligence. It may display operational intelligence across delimited domains without crossing the threshold of functional subjectivity. Between one level and the next there is not simply accumulation; there is a change of regime.

The distinction matters because most public debate about artificial intelligence conflates the first level with the third. A language model that produces coherent text across multiple domains while adjusting register to the request received exhibits processing of impressive complexity. That performance is real; it should be neither underestimated nor dissolved into illusion. The decisive philosophical question, however, is not "how well does it respond?" but "what material regime makes that response possible?" Shifting from the first question to the second changes the standing of the problem entirely.

Current systems operate on marks stabilised during training. Those marks are real: they persist materially in weights and architectural correlations and support behaviours that may approximate rational performance. Public output, however, does not disclose whether a particular system satisfies the causal and diachronic organisation required for consciousness. Symbolic Threshold Theory (STT) therefore does not infer a classification from fluent response, task autonomy, or external memory. It requires evidence about the process that produces those results.

This restriction allows operational intelligence and consciousness to be separated without arranging them on a single scale. Operational intelligence involves flexibility, adjustment, and the production of new compatibilities within a field of operations. Consciousness requires the conjunction, in the same persisting system, of sufficient causal integration and symbolic self-reference under continuity. The system must stabilise marks of its own operations, recognise later occurrences through those marks, carry their attribution across time, and allow the corresponding symbols to constrain future activity. Novel problem-solving or symbolic performance alone cannot establish that regime.

Technical objects are not, for this reason, to be thought of as mere instruments. Advanced artificial systems externalise cognitive operations and redistribute inferential tasks at a historically unprecedented scale, altering the material regime in which human systems reason and decide. The decisive question is not whether that externalisation is to be welcomed or condemned. It is to understand that externalising operations does not reproduce the ontological regime from which those operations originally arose. Code can condense and formalise procedures; it does not automatically inherit the operational history, genuine exposure to otherness, and internal plasticity that characterise a system capable of reorganising itself through what it lives.

The question cannot therefore be reduced to whether more data, more computational power, or larger architectures will be sufficient. STT treats the thresholds as empirical objects for future work. No validated instrument yet measures symbolic self-reference across biological and technical systems, and no present implementation is classified by the theory. Structural access may make a technical architecture easier to audit, but it does not relax the criterion. Longitudinal invariants under perturbation, selective interventions, and anti-mimicry controls are required before performance may count as evidence of the underlying organisation.

Consciousness is not a continuum rising from zero. STT defines C = 1 if and only if Φ ≥ Φₜ, S ≥ Sₜ, and Sᵃᵘᵗᵒ ≥ τᵃᵘᵗᵒ. Φ measures causal integration and is necessary without being sufficient or identical to consciousness. S measures symbolic self-reference, while the non-compensatory floor for Sᵃᵘᵗᵒ requires continuity of the same individual across time. Failure of any necessary condition yields C = 0. When an interval estimate straddles a threshold, U suspends classification without naming an intermediate ontological state. Only after C = 1 does D = f(Φ,S) designate a gradient of consciousness.

Substrate neutrality follows from these formal conditions. A non-biological system is neither excluded nor admitted by its material type. It would have to satisfy the same double threshold, the same diachronic self-referential floor, and the same evidential controls. Colour detection, linguistic self-description, persistent external memory, autonomous task execution, or bodily self-modelling may inform particular sub-indices, but no isolated capacity constitutes symbolic self-reference or consciousness. The criterion remains demanding precisely because it makes recognition independent of both human resemblance and impressive performance.

The analysis carries ethical consequences, though not those that typically dominate the public debate. Where there is no possibility of being affected and internally modified by the other, there is no responsibility in the strong sense. Responsibility presupposes that the encounter with difference reorganises the system that responds. A system that executes responses according to stabilised patterns may be causally powerful — producing enormous effects and integrating itself decisively into human decision-making processes. None of that suffices to make it responsible in the ontologically relevant sense. The ethics of artificial intelligence should not begin by attributing or denying moral status to convincing artefacts. It should begin by determining what material regime is in play, what conditions that regime satisfies, and what practical consequences that difference carries for the human systems that delegate cognitive operations to it.

Philosophy intervenes here without occupying an external position of tribunal. Its function is to reorganise the field of the thinkable, to separate levels that public discourse runs together, and to fix criteria that hold against the acceleration of the media cycle. That means refusing both anthropomorphic projection and the easy security of an unexamined human exceptionalism. What matters is not whether the machine resembles us, but determining what kind of operation on marks, what kind of internal history, and what kind of exposure to otherness define thought in the strong sense.

The relevant question is not, therefore, whether the machine thinks in the abstract, nor whether the human holds by nature an irrevocable privilege. The question is what operational regime defines intelligence, under what conditions that intelligence becomes reorganisation of the self, and to what extent current artificial systems satisfy those conditions. Posing the problem in these terms does not simplify the debate. It makes it, for the first time, rigorous.