Shuichiro Ogawa
日本語

Notes · updated 2026-08-08

Scope and Method

The starting point is the distinction drawn in the generative-AI note: where search delegated the location of information, generative AI delegates the generation of thought itself. To exclude harm patterns from a learning service, the prior step is to lay out the full set of claimed technological harms to education and judge which harms survive peer review and which have been refuted. Over-fitting a design to refuted harms is as much a design error as missing robust ones.

Collection ran two parallel academic streams: 12 original harm claims and 13 peer-reviewed refutations, replications, and boundary conditions, for 25 works (source/review/technology-education-harm-patterns/papers.md). The originals for the Google effect (Sparrow et al. 2011) and cognitive offloading theory (Risko & Gilbert 2016), and the generative-AI evidence, are canonical in the generative-AI note’s corpus. Effect claims are reported as the papers’ claims.

Six Mechanism-Based Harm Patterns

The originals typologize into six mechanisms.

  • P1 Delegated memory encoding: expecting external storage lowers cognitive processing of content. Google effect (Sparrow et al. 2011), photo-taking impairment (Henkel 2014), GPS and spatial memory (Dahmani & Bohbot 2020).
  • P2 Skipped generative activity: omitting the generative encoding of putting things in one’s own words. Verbatim laptop note-taking (Mueller & Oppenheimer 2014).
  • P3 Attention distraction: irrelevant processing consumes cognitive resources. Classroom multitasking harming users and nearby peers (Sana et al. 2013), the West Point computer-ban RCT (Carter et al. 2017), habitual media multitasking (Ophir et al. 2009), smartphone mere presence (Ward et al. 2017).
  • P4 Medium characteristics: screen-versus-paper reading differences (meta-analyses by Delgado et al. 2018 and Clinton 2019).
  • P5 Early exposure: passive video during developmental sensitive periods (Zimmerman et al. 2007; DeLoache et al. 2010).
  • P6 Delegated thought generation: generative AI supplying answers and reasoning itself; evidence organized in the generative-AI note (the +48% practice / −17% exam divergence, and so on).

Calculators (Hembree & Dessart’s 1986 meta-analysis of 79 studies) sit among the originals as the classic feared harm that was never supported: only continued use in grade 4 impaired basic-skill development, while calculator use alongside instruction improved performance at every other grade.

The Refutation Lineage: How Well Did Famous Harms Replicate?

The peer-reviewed refutations the request centers on split sharply by pattern.

The Google effect (P1) failed replication at its flagship experiment. Sparrow et al.’s Experiment 1 (difficult questions prime internet-related words in a Stroop task) was judged a failed replication in the pre-registered Social Sciences Replication Project (Camerer et al. 2018, Nature Human Behaviour) and failed again in two independent replications (Hesselmann 2020). A caution is needed, though: no direct replication was found of Experiments 3–4 (remembering where rather than what). What was refuted is the priming experiment; the location-memory shift itself, the most-cited part, remains untested.

Laptop note-taking harm (P2) lost its performance difference under direct replication. Morehead et al. (2019) found no consistent significant differences in test performance across conditions, and Urry et al. (2021, pre-registered) replicated the verbatim-transcription tendency but not the test-performance difference, with a mini meta-analysis of eight similar studies agreeing. The design prescription “force longhand and learning improves” is not supported by current evidence.

Media-multitasking harm and the smartphone mere-presence effect (parts of P3) shrank drastically. Ophir et al.’s strong effects shrank to d = 0.17 with low consistency in two direct replications plus a meta-analysis (Wiradhany & Nieuwenstein 2017). Ward et al.’s Brain Drain failed a pre-registered direct replication (Ruiz Pardo & Minda 2022), and a 2023 meta-analysis (22 studies) put the overall effect at g = −0.14, significant only for memory and effectively zero in North American samples.

Screen-time harm and infant-media harm (P5) were hollowed out by reanalysis. Orben & Przybylski (2019) applied specification-curve analysis to 355,000+ adolescents and showed the negative association explains at most 0.4% of variance (on par with eating potatoes); Przybylski & Weinstein (2017) found an inverted-U with moderate use not harmful. Zimmerman et al.’s baby-video harm was reanalyzed by Ferguson & Donnellan (2014), who reported the effect turns positive, null, or negative depending on analytic choices, with the dispute unresolved.

Meanwhile, some harms have no refutation or survive conditionally. Within attention distraction, no failed replication was found for the West Point classroom RCT (−0.18 SD) or Sana et al.’s spillover to nearby peers (−17%). Photo-taking impairment was supported by a further test (Soares & Storm 2018), with no refutation found. GPS harm has no replication either way (its longitudinal arm is small, n = 13). Screen inferiority in reading is supported conditionally by two independent meta-analyses: paper’s advantage (g = −0.21 to −0.25) appears only for informational texts under time pressure and vanishes for narrative texts. And cognitive offloading itself has evidence in the opposite direction: Storm & Stone (2015) showed that saving a list improves memory for the next new material (saving frees cognitive resources), with Grinschgl et al. (2021) identifying the boundary condition that memory for the offloaded information itself declines.

Robustness Verdict Table

PatternOriginal claimRefutation/replicationVerdict
Google effect (Stroop)Hard questions auto-activate internet conceptsFailed in SSRP + 2 independent replicationsNot replicated
Google effect (location-memory shift)Remember where, not whatNo direct replication foundUntested (not proof of robustness)
Laptop note-taking harmLonghand superior on conceptual testsPerformance difference not replicated in 2 lines of direct replicationNot replicated (verbatim tendency replicates)
Media-multitasking harmBroad cognitive-control deficitsShrunk to d = 0.17 in meta-analysis, low consistencyDrastically shrunk
Smartphone mere presencePresence alone reduces capacityDirect replication failed; meta g = −0.14, unstableDrastically shrunk, unstable
Screen-time harmUse damages well-being0.4% explained variance; inverted-UPractically negligible
Infant media harmBaby videos impair languageReanalysis: analysis-dependent (dispute ongoing)Contested (the RCT showing babies don’t learn from DVDs stands separately)
Calculator harmErodes basic computation79-study meta-analysis unsupportive overallUnsupported (grade-4 boundary condition only)
Classroom attention distractionUsers and nearby peers score lowerNo refutation found (incl. RCT)Comparatively robust
Photo-taking impairmentPhotographed objects remembered worseSupporting replication onlyTentatively robust
GPS and spatial memoryHabitual use erodes spatial memoryNo replication found (small longitudinal sample)Tentative (thin evidence, no refutation)
Screen reading inferiorityScreens always worseTwo meta-analyses specify conditionsConditional (informational texts + time pressure; vanishes for narrative)
Delegated thought generation (generative AI)Answer-providing AI harms retention and transferSupported by multiple RCTs; mitigated by guardrail designConditionally robust (the generative-AI note)

Design Implications: What to Exclude, What Not to Over-Fit To

Three design judgments follow from the verdict table.

First, the robust harms to exclude are two: attention distraction and delegated thought generation. Attention distraction has no refutation, RCTs included, justifying designs that separate irrelevant stimuli and notifications from the learning context. Delegated thought generation is supported by multiple RCTs (looks good during practice, nothing remains without the AI), justifying designs that withhold answers in favor of prompting and scaffolding. This is the same conclusion as the failure-design lineage’s principle of not delegating the cognitive processing that is the learning objective.

Second, do not over-fit the design to refuted harms. Forcing longhand (laptop harm not replicated), banning smartphones outright (the mere-presence effect is unstable, though notification-driven distraction is a separate matter), avoiding screen materials (inferiority vanishes for narrative texts and unpressured reading), and prohibiting calculator-like tools (unsupported by meta-analysis) lack grounding in current evidence. Over-fitting to harm panics costs the learning benefits of technology, including the cognitive resources freed by offloading.

Third, the axis of judgment is not whether something is offloaded but what is offloaded. As Storm & Stone’s saving effect shows, offloading can free resources for new learning. The single harm the calculator meta-analysis found (delegating basic skills in grade 4, while they were still being acquired) and the generative-AI harm (delegating novices’ thought generation) share one structure: harm appears when the delegated processing is itself the current learning objective. Delegating processing outside the learning objective, like the location of information or subordinate procedural computation, shows weak evidence of harm or even benefit. The opening distinction (from delegating location to delegating thought generation) matters for design precisely because what generative AI offloads by default is, for novices, exactly the learning objective: assembling thought.

Gaps

  • No direct replication exists of Sparrow et al.’s location-memory experiments (Exp. 3–4); the most-cited part of the Google effect circulates untested.
  • No failed replications were found for photo-taking impairment or GPS harm, but replications are few in absolute number; absence of refutation is not proof of robustness.
  • No direct replications were found for classroom attention distraction (Sana et al.; Carter et al.) either; its robustness rests on the absence of refutation and the strength of the RCT design.
  • Replication work on generative-AI harms has not yet accumulated (the originals date from 2023 onward); the refutation history in this note forecasts that the same scrutiny will come for those claims.
  • No Japanese-language replication studies were confirmed.

Unverified Items

  • Specific effect sizes for H01, H02, H04, H10 (paywalled full texts)
  • In-text confirmation of d = 0.97 in Mueller & Oppenheimer (2014); sample size of DeLoache (2010)
  • Direct confirmation of Hembree & Dessart’s (1986) DOI
  • First author of the Brain Drain meta-analysis (Behavioral Sciences 2023)
  • Replications of Sparrow 2011 Exp. 3–4: none found (noted as a gap)
  • Failed replications of photo-taking impairment and GPS harm: none found (same)

References

All accessed 2026-08-08. For ledger details, see source/review/technology-education-harm-patterns/papers.md. For Sparrow 2011, Risko & Gilbert 2016, and the generative-AI works, see the generative-AI note.

Original harm claims

Refutations, replications, and boundary conditions

← All Notes · Home