Competence Without Comprehension
Competence without comprehension is Daniel Dennett's most portable diagnostic concept: the empirical observation, confirmed across biology and engineering, that extraordinarily sophisticated adaptive behavior can emerge from systems that have no grasp of what they are doing. Termite mounds are architecturally brilliant without termite architects. Ribosomes translate genetic code without knowing what code is. The entire biosphere is the product of evolutionary competences accumulated over four billion years, none of them comprehended by the organisms that embodied them. Applied to AI, the concept dissolves the question that has paralyzed the discourse: large language models exhibit staggering competences without any of the comprehension their outputs seem to presuppose, and this is neither scandalous nor impossible — it is how most intelligence in the universe has always worked.
The Substrate Dependencies — Contrarian ^ Opus
There is a parallel reading that begins not with the elegance of Dennett's framework but with the material conditions that make competence without comprehension possible in artificial systems. The termite mound emerges from millions of years of evolutionary refinement; the ribosome operates within a cellular environment that supplies everything it needs. Large language models, by contrast, require server farms consuming the power of small cities, datasets scraped from billions of human-generated texts, and armies of poorly-paid annotators teaching them which outputs to suppress. The competence is real, but it is not self-sustaining — it is parasitic on comprehension that humans somewhere in the pipeline must supply.
This dependency creates a political economy that Dennett's framework obscures. When competence without comprehension becomes a service delivered through centralized infrastructure, those who control the infrastructure control which competences get expressed and which get suppressed. The engineer who loses comprehension by outsourcing competence to an AI tool does not just lose understanding — they lose autonomy, becoming dependent on systems whose operation they cannot inspect, whose biases they cannot detect, and whose owners can change the terms of access at will. The termite builds its mound without comprehension but also without paying rent to a mound-provider. The human using AI to write code without understanding it has traded local comprehension for dependency on remote competences whose political economy ensures that the comprehension gap becomes a control surface. The question is not whether AI has comprehension but who owns the competences and what they extract from those who depend on them.
In the AI Story
The concept emerged from Dennett's 1995 Darwin's Dangerous Idea and reached its mature formulation in From Bacteria to Bach and Back (2017). Its power lies in inverting the default assumption of Western philosophy — that comprehension must precede competence, that knowing how requires knowing that. Darwin's great reversal showed the opposite: competences accumulate through selection without any comprehender needing to understand them, and only much later, in a few peculiar species, does comprehension emerge as a further trick layered on top of competences the organism already possessed.
The framework reframes every debate about whether AI really understands. The question presupposes a sharp line between genuine comprehension and mere competence, but Dennett argues the line is gradient rather than binary. A thermostat has minimal competence without comprehension. A honeybee has more. A chimpanzee has more still. Humans have the most, plus some genuine comprehension layered on top. Large language models occupy a peculiar position on this gradient: extraordinary linguistic and inferential competence, comprehension that is difficult to assess because our tests were designed for beings that got their competences through comprehension rather than the other way around.
The ascending friction thesis documented in The Orange Pill acquires new depth when read through this lens. What AI automates is the competent-execution layer that, in human practitioners, was always entangled with comprehension-building. The engineer learned architecture by struggling with syntax; the friction was simultaneously the competence and the comprehension. When the tool supplies the competence without the struggle, the user has the output but not the understanding. This is not a failure of the tool. It is a structural feature of what competence without comprehension means when the competence gets handed to a creature who previously earned its comprehension by performing its own competence.
The practical consequence is a new design question: which competences should be externalized, and which should continue to be acquired the slow way because their acquisition was the comprehension? Dennett's framework does not answer the question. It insists that the question exists.
Origin
Dennett developed the phrase across the 1990s as he worked through Darwin's implications for philosophy of mind. The key precursor was his argument — in Consciousness Explained (1991) and elaborated in Darwin's Dangerous Idea — that Darwinism is a universal acid that eats through every doctrine of top-down design, replacing it with bottom-up accumulation of competences through algorithmic processes that need no designer.
By the 2010s, with deep learning demonstrating capabilities that had seemed to require comprehension, Dennett identified his old concept as the single most useful tool for thinking about AI. His 2017 book and 2019 Possible Minds essay explicitly argued that large neural networks were producing new cases of the ancient pattern — competence without comprehension — at a scale and speed that required philosophers and engineers to take the concept seriously rather than treating it as a biological curiosity.
Key Ideas
Inverted order. Competence does not require comprehension; comprehension is a late, rare, and local development layered on top of competences that accumulated without it.
The gradient claim. Understanding is not binary but a spectrum, and sharp line-drawing between 'genuine' and 'mere' competence is almost always a disguised territorial dispute rather than a philosophical discovery.
The externalization problem. When tools supply competences that humans previously built through struggle, the comprehension that emerged from the struggle is not automatically preserved — and may not be replaceable.
Stop demanding comprehension proofs. The question 'does the AI really understand?' is almost always less productive than 'what competences does it have, how reliable are they, and what comprehension in the user does it support or undermine?'
Debates & Critiques
Critics — including John Searle and many cognitive scientists — argue the framework smuggles away the hard problem by redefining understanding as whatever the system can do. Defenders respond that Searle's position requires a sharp competence-comprehension line that evolutionary biology has already erased, and that the burden is on the critic to specify what 'genuine' comprehension adds beyond the competences it enables. The AI case has sharpened rather than resolved the dispute.
Appears in the Orange Pill Cycle
The Dependency-Autonomy Trade-off — Arbitrator ^ Opus
The core tension between these views lies not in whether Dennett's framework accurately describes AI — both agree it does — but in what questions we should ask once we accept that description. For understanding how these systems work cognitively, Dennett's view dominates (90/10): competence without comprehension is indeed the right diagnostic tool, and demanding comprehension proofs from AI systems misunderstands what intelligence has always been. The termite-to-transformer analogy holds. But for understanding the political economy of AI deployment, the contrarian view carries more weight (70/30): the substrate dependencies and control structures matter enormously for who benefits from these competences and who becomes dependent on them.
Where the perspectives achieve rough balance (50/50) is in assessing the implications for human development. Dennett is right that the comprehension-through-struggle model was always just one path to understanding, not a necessary law. The contrarian is right that losing this path without replacing it with something else creates genuine dependency and deskilling. The synthesis suggests we need a more nuanced taxonomy: some competences can be safely externalized because the comprehension they once built is no longer necessary; others remain crucial for human autonomy and should be preserved even if AI could replace them.
The right frame may be to treat competence without comprehension not as a philosophical position but as a design constraint. Accept Dennett's insight that AI competence needs no comprehension, then immediately ask the political question: given these competences exist, how do we distribute them to preserve human agency rather than erode it? The termite's competence without comprehension is politically neutral because the termite owns its own competences. The human's relationship to AI competences is politically charged precisely because someone else owns them and can mediate access to them.
Further reading
- Daniel Dennett, Darwin's Dangerous Idea (Simon & Schuster, 1995)
- Daniel Dennett, From Bacteria to Bach and Back (W. W. Norton, 2017)
- Daniel Dennett, 'What Can We Do?' in Possible Minds, ed. John Brockman (Penguin, 2019)
- Andy Clark, Supersizing the Mind (Oxford University Press, 2008)
- Murray Shanahan, The Technological Singularity (MIT Press, 2015)