John McDowell and the Last Transcendental Deduction: Realism, Anti-Realism and the Challenge of Artificial Intelligence

By Matthew Parish, Associate Editor

Sunday 9 August 2026

The late twentieth century witnessed a remarkable revival of transcendental philosophy in an intellectual environment that had largely declared transcendental arguments obsolete. The revival came not from continental philosophy but from the heart of analytical philosophy itself. At its centre stood John McDowell, whose response to Michael Dummett’s anti-realism amounted to something far more ambitious than many of his contemporaries appreciated. Although couched in the restrained vocabulary of Oxford analytical philosophy, McDowell was attempting nothing less than a modern reconstruction of Immanuel Kant’s transcendental deduction. His objective was similar to Kant’s original enterprise — to demonstrate that certain features of thought and language necessarily presuppose an objective reality independent of our merely subjective capacities.

Today, however, the emergence of large language models has reopened questions that many philosophers believed McDowell had successfully resolved. Artificial intelligence has unexpectedly transformed what appeared to be an abstract dispute over semantics into an empirical question concerning the nature of language itself. Whether McDowell’s defence of realism survives this technological revolution may prove one of the defining philosophical questions of the twenty-first century.

The background to McDowell’s project lies in Michael Dummett’s sophisticated defence of anti-realism. Dummett transformed the ancient dispute between realism and anti-realism by relocating it from metaphysics into semantics. Rather than asking whether external reality exists, Dummett asked what it means for statements to possess truth conditions.

For the realist, a statement may be true even if nobody could ever know that it is true. A proposition concerning some inaccessible region of the universe may possess a determinate truth value independently of any possible verification. Truth therefore transcends human knowledge.

For Dummett, however, this conception of truth was unintelligible. The meaning of a sentence consists in the conditions under which competent speakers can recognise its correctness. To speak of truths that could never in principle be recognised is, Dummett argued, to abandon the very concept of meaning. If there can never be any possible manifestation of a statement’s truth, then there is nothing for understanding to consist in.

This represented a devastating challenge to traditional realism because it attacked not reality itself but the coherence of realist language. Dummett was not claiming that realism was false. He was claiming that realism could not even be meaningfully expressed.

This move fundamentally altered analytical philosophy. Rather than debating metaphysical pictures of reality, philosophers increasingly concentrated upon linguistic practice, justification and public standards of meaning. The influence of Ludwig Wittgenstein, P. F. Strawson and ordinary language philosophy became overwhelming.

It is precisely here that John McDowell intervened.

Many readings portray McDowell merely as offering a sophisticated defence of realism against Dummett. This description is true as far as it goes, but it understates the originality of what McDowell was attempting.

Kant had famously argued that objective experience is only possible because the mind contributes certain necessary conceptual structures to experience. The categories are not discovered through experience; rather, they make experience possible. The transcendental deduction sought to establish this by asking what conditions must obtain if experience itself is possible.

McDowell adopted an almost identical strategy.

Instead of asking what conditions make experience possible, he asked what conditions make language and judgement possible.

The similarity is striking.

Kant began with the undeniable fact that we possess coherent experience.

McDowell began with the undeniable fact that we successfully make judgements about the world.

Kant inferred necessary conceptual structures.

McDowell inferred necessary openness of thought to objective reality.

The argument possesses exactly the architecture of a transcendental deduction.

If language genuinely consists in making judgements about the world — if assertions genuinely purport to describe objective reality rather than merely expressing internal states — then language itself presupposes contact with an objective world.

McDowell’s celebrated phrase that thought must be “answerable to the world” encapsulates this deduction.

Concepts cannot simply float freely within language games.

Judgments must possess friction.

Reality itself supplies that friction.

This was an elegant reversal of Dummett’s strategy.

Where Dummett argued that semantics constrains metaphysics, McDowell argued that semantics itself presupposes metaphysical realism.

Language cannot function unless reality genuinely constrains thought.

This was, in effect, Kant rewritten for Oxford.

The analogy becomes even deeper when one examines McDowell’s notion of experience.

Kant argued that intuition without concepts is blind while concepts without intuition are empty.

McDowell develops almost precisely the same theme.

Experience is already conceptually structured.

There is no unconceptualised Given standing outside rational capacities.

Equally, concepts cannot become detached from experiential constraint.

Mind and world interpenetrate.

Experience becomes what McDowell famously called the “tribunal” before which our judgements are assessed.

Again, this resembles Kant’s transcendental architecture far more than ordinary analytical philosophy usually acknowledges.

McDowell therefore occupies a curious historical position.

He is often presented as belonging firmly within analytical philosophy.

Yet his deepest intellectual inheritance may be Kant rather than Frege.

The real question is whether the argument succeeds.

For many years the consensus among numerous analytical philosophers was that it did.

McDowell seemed to avoid both empiricist foundationalism and postmodern relativism.

He rejected the Myth of the Given without collapsing into coherentism.

He defended objective knowledge without appealing to inaccessible metaphysical entities.

His conception of second nature allowed human beings to acquire rational capacities through education and participation within shared practices while remaining genuinely answerable to objective reality.

This appeared to resolve several philosophical dilemmas simultaneously.

Yet critics remained unconvinced.

One longstanding objection concerns circularity.

McDowell argues that judgment presupposes openness to reality.

But does it?

Or does this merely redescribe how successful judgement appears from within our own conceptual scheme?

The transcendental argument works only if alternative explanations can be ruled out.

Kant believed scepticism ultimately defeated itself because coherent experience required objective synthesis.

McDowell similarly believed judgement requires genuine contact with reality.

But perhaps both philosophers underestimated the possibility of systems capable of producing apparently successful judgment without satisfying the metaphysical conclusions they wished to establish.

For centuries such systems remained entirely hypothetical.

Artificial intelligence has changed that.

Large language models generate coherent linguistic performances on a scale previously unimaginable.

They answer questions.

They produce essays.

They engage in debate.

They appear to reason.

They often exhibit remarkable flexibility across entirely unfamiliar contexts.

Most importantly, they accomplish this without possessing anything remotely resembling McDowell’s conception of experience.

Language models possess no sensory awareness in the ordinary sense.

They possess no phenomenology.

They possess no lived experience.

They do not inhabit what McDowell would call the space of reasons in the way human beings do.

Yet they nevertheless produce linguistic performances exhibiting many of the characteristics that analytical philosophers traditionally associated with understanding.

This development introduces a profound challenge to McDowell’s transcendental deduction.

Suppose one had argued in 1995 that coherent language necessarily presupposes openness to objective reality through conceptually structured experience.

The existence of GPT-class language models now demonstrates that something astonishingly similar to coherent language can emerge through entirely different mechanisms.

One must be careful here.

Language models are not conscious.

Nor are they autonomous epistemic agents.

They do not possess beliefs in the ordinary philosophical sense.

Yet McDowell’s argument concerned the conditions for meaningful judgment.

The unsettling question is whether the linguistic behaviour itself requires the transcendental conditions McDowell identified.

Perhaps it does not.

Perhaps remarkably sophisticated semantic competence can arise through statistical regularities operating over unimaginably large linguistic corpora.

If that is true, McDowell’s transcendental inference becomes considerably weaker.

An analogy with Kant helps illustrate the difficulty.

Suppose Kant had argued that objective geometry necessarily presupposes innate Euclidean intuition.

Then mathematicians discover non-Euclidean geometry.

The transcendental deduction would not necessarily become false.

But it would no longer establish necessity.

It would merely establish one possible route.

Artificial intelligence may perform an analogous function regarding McDowell.

Human language may indeed depend upon openness to reality.

But language as such may not.

That distinction matters enormously.

Indeed one might reformulate McDowell’s argument in a more modest fashion.

Perhaps experience remains necessary for creatures like ourselves.

Perhaps our biological evolution has made conceptually structured perception indispensable.

But this no longer amounts to a universal transcendental deduction concerning language itself.

It becomes a local explanation of human cognition.

That is a much weaker philosophical claim.

Even more intriguingly, language models may partially vindicate Dummett.

Meaning within these systems often appears to emerge entirely through inferential relations among linguistic expressions rather than through direct experiential contact with reality.

Words acquire significance because of their place within vast statistical networks.

This bears an uncanny resemblance to Dummett’s emphasis upon linguistic practice rather than metaphysical correspondence.

One should not overstate the analogy.

Language models were trained upon human language, itself ultimately grounded in human interaction with reality.

The models inherit semantic structure generated by embodied speakers.

They are therefore parasitic upon human experience rather than replacing it.

Yet parasitism is not dependence of the sort McDowell requires.

The models demonstrate that sophisticated linguistic behaviour may be many conceptual steps removed from direct experiential openness while remaining extraordinarily effective.

This suggests that the relation between meaning and experience is considerably looser than McDowell believed.

Does this therefore vindicate anti-realism?

Not necessarily.

Ironically, artificial intelligence may undermine both sides simultaneously.

Language models reveal that semantics can operate with surprisingly weak connections to immediate experience.

But they also reveal something Dummett underestimated.

The extraordinary success of these systems depends upon the stability of an external world reflected within massive human linguistic practices accumulated over centuries.

Without that stable world, there would have been no coherent corpus from which statistical regularities could emerge.

Reality may therefore enter language less through individual verification than through civilisation itself.

That possibility points towards an entirely new realism — one neither Kant, Dummett nor McDowell could have anticipated.

Meaning may emerge neither through immediate correspondence nor through verification alone, but through participation in historically accumulated linguistic practices whose ultimate success depends upon long-term interaction between humanity and an objective world.

Artificial intelligence becomes not evidence against realism but evidence that realism operates at a collective rather than individual level.

If so, McDowell’s transcendental deduction was neither wholly correct nor wholly mistaken.

He correctly perceived that language cannot ultimately become detached from reality.

Where he may have erred was in locating that connection within the experience of every individual rational subject.

Large language models suggest the connection may instead reside within the historical evolution of linguistic communities from which both humans and machines inherit their semantic resources.

The philosophical consequences remain uncertain.

Analytical philosophy spent half a century debating Dummett and McDowell as though they represented the final stage of the realism debate. Artificial intelligence has unexpectedly transformed that debate into the beginning of something much larger. It has exposed assumptions that neither side seriously questioned — assumptions concerning what language requires, how meaning emerges and whether experience is indispensable for thought.

Kant’s transcendental deduction attempted to establish the universal conditions for objective knowledge. McDowell attempted an equally ambitious deduction concerning the conditions of meaningful judgement. Both projects remain among the greatest achievements of philosophy. Yet history repeatedly demonstrates that transcendental arguments are vulnerable to the appearance of previously unimaginable forms of cognition.

Large language models may prove to be precisely such a form.

If they do, then the future of realism will not lie in repeating McDowell’s argument but in constructing an entirely new transcendental philosophy adequate to an age in which intelligence, language and reason no longer belong exclusively to human beings.