Last updated: 2026-10-08

U
Undergraduate level

What Is the Philosophy of Artificial Intelligence?

A thermostat holds a target temperature, measures the room against it, and switches the heating on or off to close the gap. Ordinary language barely resists describing this as the thermostat "wanting" the room warmer and "believing" it's currently too cold. Almost everyone immediately feels that description is too generous, and almost everyone has more trouble than they expect explaining exactly why. The thermostat has a goal-state, a measurement of its current state against it, and a corrective action triggered by the mismatch — the same skeleton a much more sophisticated control system has. Whatever it is that a thermostat lacks and a mind has, naming it precisely turns out to be most of the philosophical content of this entire series.

Three Ways AI Asks the Same Question FoundationalKnowledge that endures for decades — core principles

Building, or trying to build, a system that behaves as though it perceives, decides, or understands puts three different projects into the same sentence, and this series keeps them visibly separate throughout.

  • AI as construction asks how an intelligent-seeming capacity can be built at all — what architecture, what training regime, what representation makes a capacity possible.
  • AI as explanation asks what succeeding (or failing) at construction reveals about cognition in general, including human cognition — a computational model is also a testable hypothesis about how a capacity might work.
  • AI as provocation asks which assumptions about minds, knowledge, and persons start to look unstable once a machine exhibits behaviour that used to be taken as sufficient evidence of those assumptions holding.

Margaret Boden's case for treating artificial intelligence as part of cognitive science runs through exactly this overlap: a working program is simultaneously an engineering artefact and a precise, checkable claim about how some aspect of thought — language, logic, creativity, perception — might actually be organised1. Her edited collection on the philosophy of AI gathers the foundational debates from several directions at once specifically because no single one of the three projects above, pursued alone, poses the questions this series is built around.a working system is also a theory under test

The Thesis This Series Develops FoundationalKnowledge that endures for decades — core principles

Later pages read against it rather than discover it piecemeal:

Guiding thesis. A sophisticated artificial mind may not be a single, unified reasoning engine with a complete internal description of itself. It may instead be a negotiated and continually revised achievement among specialised, partially opaque subsystems, situated in a world that constrains its actions and coordinated through selective forms of shared availability.

Two further propositions travel alongside it. First, that AI should be evaluated not only by what it can induce people to believe, but by whether it helps people form warranted beliefs — about the system, its capabilities, its limitations, its outputs, and the world it acts in. The Deception Criterion takes up exactly this distinction, against the test most people still reach for first. Second, that an artificial agent may not need a biological body, but it may need a world that can resist it — a claim this series defers to its page on embodiment, once the vocabulary for discussing it carefully is in place.

None of these three claims is settled here. Later pages test each one against its strongest objections.

What Is Intelligence? The Question That Won't Sit Still FoundationalKnowledge that endures for decades — core principles

Is intelligence a behavioural property (does the system act appropriately), an organisational one (is it structured the right way), a relational one (does it stand in the right relation to a world and other agents), or an experiential one (is there something it's like to be the system)? Different traditions in AI have quietly assumed different answers while using the same word. A chess engine that outperforms every human player is behaviourally remarkable and, on most readings, organisationally and experientially nothing like a mind at all — which is precisely why calling it "intelligent" without qualification settles nothing and begs several questions at once.

A further assumption deserves to be named rather than inherited silently: that human cognition is the correct template for every possible kind of intelligence. Octopus cognition, distributed across a nervous system with most of its neurons in the arms rather than a central brain, already shows that biological intelligence doesn't have to look humanlike to be real. An artificial system built from genuinely different materials, on genuinely different principles, has no obligation to resemble human cognition either — which means judging it by how closely it imitates human behaviour (the subject of the next page) may be measuring the wrong thing even when it's measuring something real.

Vocabulary This Series Keeps Fixed FoundationalKnowledge that endures for decades — core principles

A handful of terms recur across all fifteen pages, used consistently rather than loosely. A self-model is a system's representation of its own state, capabilities, history, boundaries, goals, or likely behaviour — not automatically a self, a conscious subject, or a truthful introspective report. Metacognition is processes that monitor, evaluate, or regulate other cognitive processes, not a synonym for consciousness. Global availability means information reachable by many relevant subsystems, not literal access by every component. Embodiment means participation in a structured relationship among perception, action, constraint, persistence, and consequence — biological embodiment is one form of this, not the only one Part IV considers. Later pages introduce homeostasis, feeling, emotion, agency, and selfhood with the same discipline: each is a cluster of related but separable ideas, not a single switch a system either has or lacks.

The Strongest Objection to This Whole Approach FoundationalKnowledge that endures for decades — core principles

A serious objection deserves stating before the series leans on the framework above: perhaps philosophy is simply the wrong tool here, and the right answer to most of this series' questions is "wait for better science." On this view, arguing about whether a transformer "understands" anything is a dispute about words, not about the system, and will dissolve once neuroscience and machine learning mature enough to replace the vocabulary entirely. This objection is taken seriously throughout rather than dismissed: several later pages (especially on self-reference and on consciousness) explicitly mark where a philosophical distinction is doing real work that no amount of additional engineering detail would settle on its own, and where it might genuinely just be a placeholder for evidence not yet available. The distinction between those two cases is itself one of this series' running jobs.

Provisional Conclusion FoundationalKnowledge that endures for decades — core principles

Nothing above establishes that any current system is intelligent, conscious, or a subject of experience, and nothing above establishes the opposite either. What's established is narrower and more useful: that "is AI intelligent" is not one question but several, that construction, explanation, and provocation pull in different directions when conflated, and that the series which follows owes the reader a specific account of which question is on the table at every point, rather than a single verdict delivered once and treated as covering all of them.

Questions for Further Thought

  • If a system's internal organisation were fully transparent to you, would that settle whether it's intelligent, or only whether it's organised a particular way?
  • Is there a behaviour a thermostat could exhibit that would make "it wants the room warmer" the right description rather than a convenient shorthand?
  • Which of behavioural, organisational, relational, and experiential intelligence matters most for deciding how to treat a system — and does the answer change depending on what "treat" means?
  • Is human cognition a reasonable default template for judging non-human intelligence, biological or artificial?

Further Reading

  • Boden, M. A. (Ed.). (1990). The Philosophy of Artificial Intelligence. Oxford Readings in Philosophy. Oxford University Press.

References

  1. Boden, M. A. (Ed.). (1990). The Philosophy of Artificial Intelligence. Oxford Readings in Philosophy. Oxford University Press. ↩