The corpus from 1983–1989 reveals something stranger than a failed prediction: it reveals a field that was right about the wrong problem. Dreyfus and Athanasiou, Roszak, and Penrose all converged on common sense as the insurmountable wall — the place where rule-based AI would inevitably break down. Dreyfus put it precisely: "natural-language understanding, speech recognition, story understanding, and learning" would all fail because their structure "mirrors the structure of our everyday physical and social world," which no formal representation system could capture. [Athanasiou & Dreyfus, Mind over Machine, 1986]. They were describing a real problem. And then, roughly thirty years later, a system trained not on rules but on the statistical residue of human language at scale walked straight through that wall — not by solving the commonsense knowledge problem, but by routing around it entirely.

The 1980s thinkers couldn't imagine this because they were arguing about architecture. The debate was symbolic manipulation vs. human intuition — logic machines vs. embodied know-how. What they couldn't see was a third option that wasn't really an option yet: systems that don't represent knowledge as propositions at all, but absorb patterns from such vast corpora that something functionally resembling commonsense understanding emerges as a byproduct. GPT-4 doesn't know that fire is hot the way a child does, but it can navigate that fact in conversation with eerie competence. Dreyfus's critique was philosophically correct and practically obsolete simultaneously.

Roszak adds a dimension the technical critics miss: the political economy of AI hype. His 1986 account of DARPA funding, IBM arrangements, and researchers "associating their professional interests with sky's-the-limit commercial hype" [Roszak, The Cult of Information, 1986] maps almost perfectly onto the 2020s landscape — substitute OpenAI and Microsoft for the fifth-generation boosterism, and the structure is identical. What he got right for the right reasons: that the gap between capability and claim would be consistently weaponized for investment. What he missed: that eventually the claims, however inflated, would start producing something real enough to matter.

Penrose stands apart. Where Dreyfus and Roszak were sociological critics of AI's limitations, Penrose was making a deeper wager: that no algorithm, however complex, could capture genuine understanding, because human consciousness depends on non-computable physical processes [Penrose, The Emperor's New Mind, 1989]. By 2020, this argument had become unfalsifiable in a productive way — large language models perform so many cognitive tasks so fluently that the question of whether they genuinely understand anything has become philosophically live again, and Penrose's challenge looks less like a failed prediction than an unresolved provocation.

---

[Athanasiou & Dreyfus, Mind over Machine, 1986] does the heaviest lifting. The five-stage model of skill acquisition — from novice rule-following to expert intuition — frames the central argument: that human expertise is irreducibly embodied and situational, not extractable into formal rules. Their catalog of AI's failed promises (SHRDLU's brittleness, the commonsense knowledge problem, learning as an unsolved horizon) is historically precise and intellectually serious. The miss: they couldn't anticipate that "learning" would be solved not by cracking the epistemological problem but by scaling statistical inference beyond anything their era could conceive.

[Roszak, The Cult of Information, 1986] contributes the political economy layer. His documentation of AI's funding cycles, promotional culture, and the gap between researcher claims and actual capability is the most durable analysis in the corpus. Minsky's 1970 prediction that computers would "decide to keep us as pets" within fifteen years reads differently now — not as absurdity, but as the template for a recurring genre of AI prophecy that each generation reinvents.

[Penrose, The Emperor's New Mind, 1989] introduces the philosophical stakes that the sociological critics bracket. His engagement with "strong AI" — the claim that mental activity is simply algorithm execution — is the period's most rigorous attempt to articulate what would have to be true for AI to actually think. His invocation of Gödel's incompleteness theorems as evidence against algorithmic consciousness remains contested but not dismissed.

[Turkle, The Second Self, 1984] (appearing in footnote material here) shadows the corpus with the dimension everyone else underweights: the human relationship to AI systems. Her documentation of researchers developing emotional bonds with robots, of "technical missing" becoming "just missing," anticipates the 2020s discourse around parasocial relationships with AI chatbots in ways none of the technical critics touched.

---

The arc across 1983–1989 moves from technical optimism's collapse toward philosophical entrenchment. Early in the decade, the critique of AI was primarily empirical — here are the specific things it has failed to do, here is why those failures are structural. By the late 1980s, the critique had become ontological — Penrose arguing that the problem wasn't bad implementation but wrong assumptions about what computation could ever be. This shift from "AI hasn't solved X yet" to "AI cannot in principle solve X" is the decade's intellectual signature.

What the pattern reveals, seen from 2020: both moves were partially wrong. The empirical critique underestimated scale; the ontological critique may have misidentified where the hard problem actually lives. The genuine surprise of 2020-era AI isn't that it achieved general intelligence — it didn't, by most rigorous definitions — but that it achieved functional adequacy across so many domains that the question of whether it's "really" intelligent became practically secondary. The 1980s corpus was arguing about a threshold that the 2020s simply stepped around.

---

This corpus is almost entirely white, male, and Anglo-American — Dreyfus, Penrose, Roszak, Turkle are all working within a Western analytic tradition that takes for granted which cognitive tasks count as "intelligence" and which languages and cultural contexts constitute "commonsense." The commonsense knowledge problem as defined in this corpus is implicitly English-language commonsense rooted in a specific cultural world. The 2020 reality of AI includes systems trained on multilingual corpora, deployed in contexts these authors never imagined, producing failures that their frameworks don't have vocabulary for — not brittleness in the old sense, but bias encoded at scale, hallucination as a structural feature, and value alignment as a problem that isn't about common sense at all.

The corpus also has no voice from anyone who would be subject to AI systems rather than building or theorizing them — no labor economists worried about automation, no civil rights thinkers concerned about surveillance, no Global South perspectives on who benefits from and who is harmed by these technologies. The 1980s debate was almost entirely conducted among the people designing the systems and the philosophers questioning their assumptions. The people who would live with the consequences were outside the frame entirely.