Artifictional Intelligence
Artifictional Intelligence: Against Humanity's Surrender to Computers (Polity, 2018) is Collins's most direct engagement with the AI debate, published at the moment when transformer-based language models were beginning to transform the technology's public perception. The book's central thesis is that current AI — however impressive its fluency — lacks the social embedding required for genuine understanding, and that the real danger is not that machines will exceed human capability but that humans will prematurely surrender evaluative vigilance to systems that look competent but do not understand. The book develops Collins's framework of mimeomorphic and polimorphic action, tacit knowledge taxonomy, and interactional vs. contributory expertise into a unified argument about what AI is and is not.
The Material Substrate Problem — Contrarian ^ Opus
There is a parallel reading that begins not with social embedding but with the political economy of AI deployment. Collins's framework, while sophisticated in its treatment of tacit knowledge, may underestimate how thoroughly capital's logic shapes what counts as 'understanding' in practice. When a language model can pass medical licensing exams, draft legal briefs that win cases, and generate code that runs production systems, the distinction between 'real' and 'artifictional' intelligence becomes less a philosophical matter than an economic one. The systems that allocate resources, determine employment, and shape daily life don't require Collins's deep social understanding — they only need to be effective enough to displace human judgment in specific profitable domains.
The lived experience of those whose work is being transformed suggests a different calibration of danger. A radiologist whose diagnostic accuracy is surpassed by pattern-matching systems, a translator whose nuanced cultural knowledge is approximated well enough by statistical models, or a junior analyst whose exploratory work is automated away — these professionals experience not a 'surrender' but a structural displacement. The question isn't whether AI truly understands in Collins's sociological sense, but whether markets care about that distinction. The book's careful taxonomy of knowledge types may accurately describe what machines lack, but history suggests that capitalism has always been remarkably creative at restructuring work around whatever machines can do, rather than preserving work that requires what they cannot. The real transformation isn't happening at the level of epistemology but in the reorganization of labor processes, professional authority, and the economic value of different forms of human expertise.
In the AI Story
The book was written before GPT-3 and without knowledge of the scale-driven capabilities that would emerge in subsequent years. Its core claims have nevertheless proved durable. Collins's argument that machines would achieve impressive interactional competence without crossing into contributory competence describes the actual capabilities of 2024-2026 frontier models with eerie precision. His warning about the Surrender has acquired empirical support as studies document declining evaluative vigilance in AI-assisted professional work.
The book's distinctive contribution is its refusal of the binary that dominates popular AI discourse. Collins is neither a dismisser of AI (he recognizes the genuine achievement of sophisticated mimeomorphic reproduction) nor an enthusiast (he insists on the structural limitations that textual training cannot overcome). His position is a precise middle: these machines are real, their capabilities are real, but the capabilities they possess are not the capabilities their fluency suggests, and the cultural response should be calibrated to what they actually do rather than what they appear to do.
Origin
Published by Polity in 2018, the book consolidated arguments Collins had been developing across papers and lectures throughout the 2010s. The timing was prescient: the book appeared at the cusp of the transformer revolution, and its framework has provided one of the most durable lenses for analyzing subsequent developments.
Key Ideas
The Surrender. The book's titular warning: humans will defer to machines that look competent but do not understand, and the deference is more dangerous than the machines.
Deep vs. surface AI. The book distinguishes genuine AI (which would require socialization into human communities) from surface AI (which reproduces the form of intelligence without its substance).
The six levels. Collins enumerates six levels of AI capability, arguing that current systems achieve roughly level 4 but are often mistaken for level 6 (genuine social intelligence).
The sociological critique. The book's argument is sociological rather than technical: the barrier to AI is the social constitution of human knowledge, not computational limitations.
Appears in the Orange Pill Cycle
The Pragmatic Knowledge Boundary — Arbitrator ^ Opus
The tension between Collins's sociological framework and the material critique resolves differently depending on which question we're asking. If we're asking 'what is genuine understanding?' Collins's analysis is essentially correct (95/5) — current AI systems do lack the social embedding that constitutes human knowledge. The distinction between interactional and contributory expertise remains sharp and empirically grounded. But if we're asking 'what determines which capabilities matter in practice?' the material critique dominates (20/80) — market forces, not philosophical distinctions, shape AI deployment.
The question of danger shows the most balanced weighting (50/50). Collins is right that premature surrender to non-understanding systems poses risks we're only beginning to document — medical AI making errors no human would make, legal AI missing context that matters, coding assistants introducing subtle bugs. Yet the contrarian view correctly identifies that displacement doesn't wait for philosophical resolution. Both dangers are real: we simultaneously face the epistemic risk of trusting machines that don't understand and the economic risk of being displaced by machines that understand enough for market purposes.
The synthetic frame that emerges is one of pragmatic knowledge boundaries. Rather than a binary between 'real' and 'artificial' intelligence, we might map a spectrum of pragmatic adequacy — domains where surface competence suffices (routine legal drafting), domains where it catastrophically fails (novel medical cases), and the vast middle where the boundary shifts based on economic pressure, regulatory frameworks, and collective choices about acceptable risk. Collins provides the analytical tools to identify these boundaries; the material critique reminds us that identifying them is different from defending them. The book's lasting value may be less as a prediction of what AI cannot do than as a framework for deciding what we should not let it do, even when it can.
Further reading
- Harry Collins, Artifictional Intelligence: Against Humanity's Surrender to Computers (Polity, 2018)
- Harry Collins, Artificial Experts: Social Knowledge and Intelligent Machines (MIT Press, 1990)
- Harry Collins, Tacit and Explicit Knowledge (University of Chicago Press, 2010)