The Philosophy of Artificial Intelligence

Beneath every technical advance in this field sits a question the engineering never quite answers: what would it actually mean for a machine to be intelligent, and how should you go about building one? For seventy years, two very different answers have been competing for the honor of being right.

Symbolic AI versus machine learning

Symbolic AI — the tradition of John McCarthy and his generation — starts from a clean idea: intelligence is the manipulation of symbols according to explicit rules. You write knowledge down in a formal language, and reasoning is what happens when logical rules are applied to it. Its great virtues are transparency and precision. You can inspect the reasoning, and you can trace any conclusion back to the facts and rules that produced it.

Machine learning, and the deep learning that now dominates, begins from the opposite instinct. Instead of encoding knowledge by hand, it lets a system discover statistical patterns in enormous piles of examples. Its strength is exactly where symbolic methods were weakest: the messy, high-dimensional stuff — images, speech, natural language — that no one could ever reduce to tidy rules. Its weakness is the mirror image of symbolic AI’s strength. What the system has “learned” is smeared across millions of numerical parameters, and reading meaning back out of them is genuinely hard.

Rule-based intelligence versus data-driven intelligence

Put the two side by side and you get a clean trade-off. Rule-based systems are explainable and predictable, but brittle — they know only what they were explicitly told, and they fail the moment the world hands them a case their authors never anticipated. Data-driven systems are flexible and robust to variation, but opaque, hungry for data, and prone to failing in ways that are hard to predict and harder to explain.

Neither, on its own, looks sufficient for general intelligence. That is why much of the interesting research now tries to marry them — pairing learned components with explicit reasoning or structured knowledge — on the bet that the future belongs to whoever can combine the transparency of rules with the reach of learning.

Limitations of current systems

Whatever they can do on specific tasks, today’s systems share a set of limitations that no amount of scaling has yet erased:

• They lack robust commonsense understanding of the physical and social world.
• They do not reliably tell correlation from causation.
• They can produce fluent, confident output that is simply wrong, with no internal signal that anything is off.
• They have no goals, understanding, or awareness in the human sense. What they have is a model of the statistical relationships in their training data.

Do “thinking machines” exist today?

On the honest reading of the evidence: no. Modern systems can perform narrow tasks brilliantly, sometimes above human level — and that is a real achievement — but it is not the same thing as thinking. They do not grasp meaning, hold beliefs, or reason about the world the way a person does; they compute outputs from inputs based on patterns in data.

Whether genuine machine understanding is possible even in principle remains open, a question as much philosophical as technical. It is worth remembering that McCarthy’s founding claim — that any feature of intelligence could, in principle, be described precisely enough for a machine to simulate it — has never been proven or disproven. Seventy years on, we are still arguing about it. That may be the most honest measure of how hard the problem really is.