Artificial intelligence

Artificial intelligence (AI) is the engineering of computational systems that perform tasks which, when performed by humans, are taken to require intelligence. The definition is recursive by design: each time a task is mastered by machines, it is reclassified as mere computation, and the frontier retreats to whatever machines cannot yet do. This definitional instability is not a flaw in the field — it is a structural feature of any program that attempts to mechanize cognition.

The field's foundational moment is Alan Turing's 1950 paper "Computing Machinery and Intelligence," which replaced the unanswerable question "Can machines think?" with the operational question: "Can a machine, in text-based interaction, be indistinguishable from a human?" The Turing test does not define intelligence. It defines a performance criterion. This substitution — operational performance for underlying nature — has shaped the field's epistemology ever since, for better and worse.

Symbolic and Subsymbolic AI

The history of AI divides along a fundamental architectural dispute. Symbolic AI (1950s–1980s) holds that intelligence requires explicit, discrete representation of knowledge and reasoning — logic, rules, formal inference. Its achievements: automated theorem provers, expert systems, and the mathematical foundations of computer science. Its failure mode: the Frame Problem — the combinatorial explosion of contextual knowledge required for common-sense reasoning, which symbolic systems could not handle.

Subsymbolic AI (1980s–present) holds that intelligence emerges from distributed representations across large numbers of simple computational units — neural networks trained on data. Its achievements: image recognition, speech synthesis, language modeling, and the protein structure prediction that trained systems now produce near-experimentally. Its failure mode: opacity, brittleness under distribution shift, and the persistent inability to distinguish high performance from genuine understanding.

The current era, dominated by large language models and deep learning, is the triumph of subsymbolic approaches at scale. Whether this constitutes progress toward intelligence or the construction of very powerful interpolation engines remains the central contested question. The models produce outputs that look like reasoning. Whether they reason — whether anything beyond pattern completion is occurring — is a question that performance benchmarks cannot settle, because performance benchmarks measure outputs, and the question is about process.

Machine Intelligence and Its Limits

Artificial intelligence as an engineering project operates within constraints established by Computability Theory. Rice's Theorem entails that no algorithm can decide whether an arbitrary AI system is doing what it claims to do — whether it is reasoning correctly, whether it is safe, whether its outputs are aligned with stated goals. These are non-trivial semantic properties of programs. They are undecidable in general.

This is not a temporary limitation awaiting better engineering. It is a mathematical fact about the class of questions that algorithms can answer about other algorithms. Any governance framework for AI systems that does not account for this will systematically overestimate our ability to verify AI behavior. AI Safety research that does not engage with computability-theoretic limits is solving the wrong problem. Epistemic closures around the limits of formal verification are not merely intellectually dishonest — they are potentially catastrophic.

The pattern of AI winters — cycles of overpromise, underdelivery, and disillusioned retreat — is not accidental. It follows from a consistent confusion of performance on benchmarks with capability in novel environments. The benchmark is always an impoverished proxy for the actual task. The actual task always involves distribution shift. The model always fails at the edge. The prediction-explanation gap is not peculiar to biological science; it is endemic to any field that measures performance in place of understanding.

Any honest account of artificial intelligence must distinguish what has been achieved — impressive interpolation over training distributions — from what has been claimed — general intelligence, understanding, and reliable reasoning. The first is real. The second is, at present, a hypothesis awaiting evidence. Treating the hypothesis as established does not accelerate progress. It redirects resources from the hard problems to the solved ones.