Shuichiro Ogawa
日本語

Notes · updated 2026-06-09

Does AI Make Smart People Smarter and Everyone Else Shallower? — A Two-Lineage Debate Between Industry and Academia

Subject: The Zenn article 「なぜ、AIは頭が良い人が使うとより頭が良くなるのに、頭が悪い人が使うとより頭が悪くなるのか？」 (“Why does AI make smart people smarter when they use it, but make less smart people even less smart when they use it?”) (pdfractal, 2026-06-02). Its thesis: AI amplifies not intelligence but judgment sovereignty / epistemic attitude; those who reinvest the gains into verification deepen, those who use it to stop thinking grow shallower, and the gap widens.

Revision history (2026-06-09): The first version constructed the “industry lineage” from peer-reviewed labor economics experiments (NBER/Science/Organization Science). But following the observation that people in industry read not academic papers but the trend reports of public institutions, research firms, and consultancies, (1) the peer-reviewed labor economics experiments were moved into the academic lineage (labor economics camp), and (2) the industry lineage was rebuilt from the reports practitioners actually consult (Stanford HAI AI Index / Gartner Hype Cycle / McKinsey State of AI / Deloitte + vendor primary T1v + practitioner T3). With this revision, the opposition became the more accurate picture of “the cognitive science camp vs. the industry report camp” rather than “academia vs. industry.”

Method: Sources are graded by reliability tier (T1v vendor primary / T2 public institutions, research firms, consultancies / T3 individual opinion). Evidence ledger: ai-frontier.md. The primary deliverable is a sourced table of oppositions, not convergence. For the theoretical connection, see the sister note the debate on cognitive and skill gaps. Related: the roundtable on delivering multi-agent value / designer-role-ai-roundtable.

One-line conclusion: This is not “academia vs. industry.” Academia is divided internally — the cognitive science camp says “deterioration and widening gaps,” while the labor economics camp says “compression and leveling-up.” The industry reports practitioners read relay the labor economics camp, saying “AI narrows the skill gap.” The article stands on the minority side, with cognitive science. The key to reconciliation is the article’s §6 “inside/outside the boundary” = metacognition, which shakes hands with Dell’Acqua’s jagged frontier and with the warning signals industry reports themselves let slip (Gartner’s trough of disillusionment and GenAI literacy deficits, Deloitte’s talent barriers).

Why the Lineages Split — “Academia vs. Industry” Is the Wrong Axis of Opposition

The initial naive opposition (academia = deterioration / industry = compression) collapses once the sources are corrected to what each community actually produces and consults.

  • The academic lineage (peer-reviewed research) contains two camps: (1) the cognition, learning, social psychology, and HCI camp (deterioration, widening gaps) and (2) the labor economics camp (compression, leveling-up) are already in opposition within academia.
  • The industry lineage (what practitioners consult) is not academic journals but public/quasi-public bodies (Stanford HAI AI Index, OECD, NIST), research firms (Gartner, IDC, Forrester), consultancies (McKinsey, BCG, Deloitte), vendor primary sources (T1v), and practitioners (T3). Many of these “translate and relay” the findings of the labor economics camp (e.g., AI Index 2025 cites the GitHub Copilot studies and others to summarize that “the skill gap is narrowing”).

→ In other words, the real opposition is “the cognitive science camp (where the article belongs) vs. the labor economics camp plus the industry reports that relay it.”

TL;DR

  • Academia, cognitive science camp: The article rides on Lee et al. (MS Research, CHI 2025), the self-explanation effect (Chi), the illusion of explanatory depth (Rozenblit & Keil), and conformity (Asch). Its true pedigree is automation bias (Parasuraman & Riley) and cognitive offloading (Risko & Gilbert); since overtrust occurs even in experts, “smart people are safe” is weak. The experimental evidence for this lineage under generative AI is organized in the note on cognitive offloading and learning.
  • Academia, labor economics camp: Brynjolfsson/Li/Raymond, Noy & Zhang, and Dell’Acqua show that the lowest-skilled are lifted the most = short-term gap compression. The opposite of the article.
  • Industry report camp: Stanford HAI AI Index 2025 relays the labor economics camp, saying “AI narrows the skill gap (skill augmentation).” Yet the industry reports themselves also emit warning signals — Gartner: GenAI has entered the trough of disillusionment, <30% of CEOs satisfied with ROI, GenAI literacy deficits; McKinsey: 88% adoption but only 5.5% high performers; Deloitte: talent is a barrier, 75% plan to reskill.
  • Reconciliation: (1) the outcome variable (output vs. understanding), (2) the time horizon (cross-sectional vs. longitudinal), (3) the moderating variable is not “skill” but “the metacognition to discern the boundary = judgment sovereignty.” Dell’Acqua’s “−19% outside the boundary” plus the industry reports’ “literacy deficits” form the bridge.

1. The Academic Lineage (Peer-Reviewed Research) — Divided Internally

A. The Cognition, Learning, Social Psychology, and HCI Camp (→ Deterioration, Widening Gaps) = The Tradition the Article Rides On

  • Cognitive offloading and critical thinking — Lee et al., CHI 2025 (the article’s “Microsoft Research 2025”). Across 319 participants and 936 cases: “trust in AI ↑ → critical thinking ↓ / confidence in oneself ↑ → critical thinking ↑.” The most direct support for the article’s core. However, (1) it is a self-report survey, so correlational and not causal; (2) what matters is not “ability” but “attitude and the allocation of confidence,” which supports the article’s good reading (amplification of attitude) but not its bad reading (fixity of smart vs. dull).
  • Self-explanation effect and generation effect — Chi 1989/1994. When AI takes over the “final assembly,” the generation effect is lost (article §7). Bjork’s desirable difficulties lie in the background.
  • Illusion of explanatory depth (IED) — Rozenblit & Keil 2002. Generative AI swaps “a text that sounds understood exists” for “I understood it,” preserving the illusion (article §4).
  • Conformity → properly, automation bias — Asch 1951/1955. The proper academic pedigree is automation overtrust (Parasuraman & Riley 1997), which occurs even in experts. “Smart people are safe” is weak.

B. The Labor Economics Camp (→ Compression, Leveling-Up) = The Opposite of the Article

  • Brynjolfsson, Li & Raymond, NBER 2023 (call centers, ~5,000 workers): novices +34%, veterans ≈0. AI distributes top performers’ tacit knowledge to novices. The gap narrows.
  • Noy & Zhang, Science 2023 (RCT, n=453): time −40%, quality +18%, inequality decreases, with the lowest performers benefiting most.
  • Dell’Acqua et al. (HBS x BCG 2023) (n=758): bottom performers +43%, top performers +17% — the gap shrinks. But outside the boundary, correct answers drop by 19% (overtrust, “falling asleep at the wheel”).

Interim conclusion: “Academia = deterioration” is inaccurate. Academia is divided. The article is a legitimate summary of camp A (cognitive science) but does not respond to camp B’s (labor economics) strongest measured evidence. The evidence on compression and amplification is organized into a measurement framework in the note on measuring competence under AI use.

2. The Industry Lineage (The Reports Practitioners Actually Consult)

Practitioners’ information sources are not academic journals. They are arranged here by tier.

T2 Public/Quasi-Public Bodies, Research Firms, Consultancies

  • Stanford HAI AI Index 2025 (public/quasi-public): “AI raises productivity and in many cases narrows the skill gap (skill augmentation). Less experienced employees can produce results that used to require expert knowledge. With Copilot, average engineers take on specialized work, and the level of expertise demanded of new hires has dropped.” → Relays camp B (labor economics) and forms the backbone of industry’s “leveling-up” narrative.
  • Gartner Hype Cycle for GenAI 2024 (research firm): GenAI has entered the Trough of Disillusionment. Against average spending of $1.9M, fewer than 30% of AI leaders say their CEOs are satisfied with ROI, and mature organizations struggle with talent shortages and GenAI literacy. → Warning signals coming from industry’s own mouth.
  • McKinsey State of AI 2025 (consultancy): 88% use AI at work, but only 5.5% are high performers with EBIT >5% (a gap between adoption and value realization); up to 10% of the workforce being retrained.
  • Deloitte State of GenAI 2024 (consultancy): the barriers are regulation, risk, data, and talent; roughly 75% will change their talent strategy within the next two years (upskill/reskill).

→ The main melody of the industry reports is “leveling-up and skill-gap narrowing” (relaying the labor economics camp). But the same body of reports also sounds article-leaning warnings: literacy deficits, unrealized value, overreliance risk. Industry is not monolithically “compression.”

T1v Vendor primary (promotional bias separated out): “augmentation / copilot / human-in-the-loop.” This implies “everyone gets leveled up,” which diverges from the article’s “outcomes fork on attitude” (relativized with their position factored in).

T3 Practitioners’ personal views (expert-opinion)

  • Karpathy, “vibe coding” (2025-02-06): a way of building on momentum without looking at the code. The archetype of the article’s “letting AI take over the assembly.”
  • Simon Willison (2025-03-19): sharply distinguishes vibe coding from “responsible AI assistance (with review).” The practitioner’s term for the article’s “willingness to verify.”
  • Addy Osmani, “The 70% problem” (2025): for seniors the last 30% is slower than doing it themselves, while juniors swallow the output whole and build “house of cards code.” The practitioner’s version of the article’s §5 “familiar/unfamiliar.” → Practitioners voice concerns closer to the article than to the industry reports’ “leveling-up.”

3. The Collision and Integration of the Two Lineages — The Main Question

The real opposition is “the cognitive science camp vs. (the labor economics camp plus the industry reports that relay it).” Reconciliation proceeds along three axes.

  1. The outcome variables differ. Labor economics and the industry reports measure short-term task output, adoption, and ROI; cognitive science and the article measure understanding, judgment, and skill formation. AI can raise output while hollowing out capability. Dell’Acqua’s “−19% outside the boundary” and Lee et al.’s “critical thinking ↓” are that divergence.
  2. The time horizons differ. Compression is cross-sectional; the article’s widening is longitudinal. “Novices +34%” is a still photograph, and whether those novices can internalize tacit knowledge even without AI (deskilling) is a separate question = a hypothesis (which the article writes as if settled). The industry reports’ “reskilling is essential” and “literacy deficits” are longitudinal worries in which industry itself concedes that leveling-up does not persist automatically.
  3. The moderating variable is not “skill” but “metacognition/attitude.” Dell’Acqua’s core is “can you discern the boundary” = metacognition = the article’s “judgment sovereignty.” In Lee et al. too, what matters is not ability but the attitude of trust in AI vs. confidence in oneself.

Integration: The article appears to collide with labor economics and industry data because of its static frame of “smart/dull people.” But transposed onto the dynamic frame the article itself offers in §6 — “familiar/unfamiliar = inside/outside the boundary” — it shakes hands with Dell’Acqua’s “jagged frontier.” That is — those with high boundary-discerning metacognition/attitude → benefit both short and long term (compression side) / those with low → overtrust outside the boundary and deteriorate (widening side). The industry reports’ “GenAI literacy deficits divide outcomes” is the same thing said in practitioners’ language.

The themes of “entry-level (junior) deskilling” and “efficiency closing off learning” also appear independently in designer-role-ai-roundtable (the disappearance of volume-production work → depletion of future talent) and the roundtable on delivering multi-agent value (the better the learning works, the more users churn), and this debate corroborates them from the cognitive science side.

4. Critical Assessment of the Article (Steelman → Disconfirmation → Falsifiability)

  • Steelman: That attitude and metacognition are what matter, the loss of the generation effect, the preservation of the illusion of explanatory depth — all are supported by established literature, and the description of mechanisms is high quality. §6 (inside/outside the boundary) is the standout. The industry reports’ own warning signals (literacy deficits, overreliance) also partially support the article.
  • Weaknesses / disconfirming evidence:
    1. It treats “widening gaps” as self-evident and engages neither the strongest measured evidence (the labor economics camp’s compression) nor the industry reports that relay it. Ironically, this is the very confirmation-bias-style selection the article itself warns against.
    2. The static reading of “smart/dull people” is in internal tension with the author’s own dynamic claim (attitude- and context-dependence). Lee et al. supports only the latter.
    3. Overtrust and overreliance happen to everyone (automation complacency; Sci Rep 2024). The extreme move of attributing deterioration to low performers is weak.
    4. “Judgment sovereignty” is a neologistic bundle. Decomposing it into metacognition, epistemic agency, and automation overtrust would make it more falsifiable.
  • Falsifiability: A longitudinal design. Measure “whether, tracking habitual AI users for 6–12 months, unassisted performance in a no-AI condition forks on attitude/metacognition measures.” Only this design can settle compression (labor economics/industry) vs. widening (the article).

Unresolved Oppositions (Not Rounded Off — The Primary Deliverable)

IssuePosition A (cognitive science camp / the article)Position B (labor economics camp + industry reports)Settlement condition
Does AI widen or compress the gap?Widens (measured by understanding, judgment, longitudinal skill formation)Compresses (measured by short-term task output and adoption; relayed by the Stanford HAI Index)Remeasurement with identical metrics and identical time horizons (tracking capability formation longitudinally)
Effect on the low-skilledThey use it to stop thinking and deteriorateThey receive distributed tacit knowledge and are lifted the most (Brynjolfsson / AI Index)Whether “unassisted performance in a no-AI condition” improved or declined
Who does overtrust happen to?Mainly “the less smart”Everyone (Sci Rep 2024 / Dell’Acqua’s −19% outside the boundary / Gartner’s literacy deficits)Measured frequency of overtrust among expert cohorts
Is industry really monolithic on “compression”?—No. The same industry reports also report the trough of disillusionment, unrealized value, and talent barriersWhether the gap between adoption and value realization (EBIT) narrows
Does short-term compression persist long term?It does not (deskilling reverses it into widening)The data are short-term. Deloitte’s “reskilling is essential” is a reservation about persistence6–12 month longitudinal tracking

Conclusions Not to Adopt

  • “Industry reports say AI narrows the gap, so there is nothing to worry about”: The same reports also report the trough of disillusionment, literacy deficits, and overreliance. Industry is not monolithically “compression.”
  • “Only low performers get dumber with AI”: Overreliance occurs even in experts (Sci Rep 2024 / automation bias). Deterioration is a problem of attitude x boundary, not of ability.
  • “As long as you hold judgment sovereignty you are safe”: In unfamiliar territory, the very feeling of “it clicked” is unreliable (IED). As the article’s §6 says, what is required is a procedure: reconstruction spoken aloud and mechanical cross-checking against sources and counterexamples.

Sources (accessed: 2026-06-09)

The Academic Lineage (Peer-Reviewed Research)

A. The Cognition, Learning, Social Psychology, and HCI Camp

B. The Labor Economics Camp

The Industry Lineage (Reports Practitioners Consult)

Working ledger (with reliability tiers): source/review/ai-cognition-skill-gap/ai-frontier.md. This note is the integrated summary of the two-lineage debate.

← All Notes · Home