# Source notes for *When Knowledge Is Cheap, What Is Worth Learning?*

Every source was opened on **2026-10-21 (America/Los_Angeles)**. This file records, for each reference, which original was read, what it supports, what it does not establish, and how it was checked. The bibliography is `public/papers/when-knowledge-is-cheap/references.bib`; this file is published beside the paper as `sources.md`. Summaries are deliberately short and contain no verbatim excerpts longer than a phrase.

**How to read the "checked" column.** *Original* means the primary document itself was read (a publisher page, an arXiv or repository page, an official PDF, or a page scan read page by page). *Indexed abstract* means the publisher blocked automated reading and the abstract was read through an indexing service that stores it verbatim (Europe PMC, Crossref, Semantic Scholar). *Author summary* means a number was taken from the authors' own public account of the same study because the original was not reachable; the manuscript says so where it matters.

## Citation map

| BibTeX key | Verified identity and primary source | What was read | Supports | Does not establish | Checked |
| --- | --- | --- | --- | --- | --- |
| `plato1925phaedrus` | Plato, *Phaedrus* 274e–275b, trans. H. N. Fowler, Loeb Classical Library (Harvard University Press, 1925), Perseus Digital Library. | Sections 274e, 275a, 275b. | Thamus's reply to Theuth: writing will produce forgetfulness, pupils will trust external characters instead of memory, and will have the appearance of wisdom rather than wisdom. | Anything about the historical effect of writing; it is a myth told by Socrates. | Original (Perseus, three section pages). |
| `simon1971attention` | H. A. Simon, "Designing Organizations for an Information-Rich World," in M. Greenberger (ed.), *Computers, Communications, and the Public Interest* (Johns Hopkins Press, 1971), pp. 37–72. | Section "The Scarcity of Attention," pp. 40–42 of the scan. | Information consumes the attention of its recipients, so a wealth of information creates a poverty of attention and a need to allocate it; most of the cost of information is borne by the recipient. | Anything quantitative; it is an essay. | Original (page scan, pp. 40–42 read as images; page numbers printed on the pages). |
| `polanyi1966tacit` | M. Polanyi, *The Tacit Dimension* (Doubleday, 1966; reissued University of Chicago Press, 2009), p. 4. | The sentence "we can know more than we can tell," as quoted by Autor (2014) in the abstract of NBER Working Paper 20485; the page number as given by two scholarly citations (a Taylor & Francis chapter and Autor's paper). | The claim that part of what we know cannot be put into words. | The share of knowledge that is tacit; the book is philosophy. | Quotation via Autor 2014 (original); page via scholarly citations. The 1967 Doubleday scan on the Internet Archive is lending-only and its OCR text could not be read, so the page was not seen directly. |
| `agrawal2018prediction` | A. Agrawal, J. Gans and A. Goldfarb, *Prediction Machines* (Harvard Business Review Press, 2018), 250 pp., ISBN 978-1-63369-567-2. | Bibliographic record (Open Library); chapter titles "Cheap Changes Everything" and "The Value of Judgment" in two library catalogue records of the 2018 edition. | The complements argument: when the price of a good falls, the value of its complements rises; AI is a drop in the cost of prediction, which raises the value of judgment. | Any empirical result; nothing about generative AI, which postdates the book. | Bibliographic record and chapter titles verified; chapter numbers differ between the 2018 and 2022 editions, so the manuscript cites chapters by title. The formal statement is taken from `agrawal2018judgment`. |
| `agrawal2018judgment` | A. K. Agrawal, J. S. Gans and A. Goldfarb, "Prediction, Judgment and Complexity," NBER Working Paper 24243 (January 2018). | NBER abstract. | AI interpreted as improvements in prediction; judgment defined as determining payoffs; "prediction and judgment are complements as long as judgment is not too difficult." | That judgment is always made more valuable; in complex environments the effect depends on whether prediction enables automated decisions. | Original (NBER page). |
| `autor2014polanyi` | D. H. Autor, "Polanyi's Paradox and the Shape of Employment Growth," NBER Working Paper 20485 (September 2014). | NBER abstract. | Tacit knowledge of how the world works often exceeds explicit understanding; tasks requiring adaptability, common sense and creativity are hard to automate; commentators overstate substitution and underweight complementarity. | Anything about generative AI; the paper covers earlier computerisation. | Original (NBER page). |
| `autor2024expertise` | D. Autor, "Applying AI to Rebuild Middle Class Jobs," NBER Working Paper 32140 (February 2024). | NBER abstract. | Information is an input to the more consequential function of decision-making; AI's opportunity is to extend the reach of expertise to workers who already possess complementary knowledge. | That this will happen; it is an argument. | Original (NBER page). |
| `deming2017social` | D. J. Deming, *QJE* 132(4): 1593–1640 (2017); NBER Working Paper 21473. | NBER abstract and the published record. | Jobs requiring high social interaction grew by nearly 12 percentage points as a share of the US labour force (1980–2012); math-intensive but less social jobs shrank by 3.3 points; growth was strongest for jobs requiring both. | Anything about AI; the data end in 2012, and the abstract does not use the word "complementarity." | Original (NBER page; Crossref record for the journal version). |
| `brynjolfsson2022turing` | E. Brynjolfsson, "The Turing Trap," *Daedalus* 151(2): 272–287 (2022). | Abstract. | Technology that augments people creates new capabilities, while technology that substitutes for them erodes their bargaining power; incentives currently favour automation over augmentation. | Any measurement. | Indexed abstract (Semantic Scholar); record confirmed with Crossref. |
| `recht1988knowledge` | D. R. Recht and L. Leslie, *Journal of Educational Psychology* 80(1): 16–20 (1988). | Bibliographic record (Crossref); the abstract as indexed (64 junior-high students, 2 × 2 design of reading ability by baseball knowledge, a significant main effect of prior knowledge on all recall measures and no interaction); Willingham's account of the same study. | Readers who know a domain understand and recall a text about it better; poor readers with high baseball knowledge did better than good readers with low knowledge. | Transfer to other domains; one domain and a small sample. | Bibliographic record original; findings via indexed abstract and via `willingham2006knowledge` (original). The APA full text was not reachable. |
| `willingham2006knowledge` | D. T. Willingham, "How Knowledge Helps," *American Educator*, Spring 2006. | The article on aft.org. | Factual knowledge enhances reasoning and problem solving; the more one knows, the easier it is to learn new things; his account of Recht and Leslie. | New data; it is a synthesis for teachers. | Original (AFT page). |
| `chase1973chess` | W. G. Chase and H. A. Simon, "Perception in Chess," *Cognitive Psychology* 4(1): 55–81 (1973). | pp. 55–56 of the paper. | Masters reconstruct real positions almost perfectly after a five-second view, do no better than weaker players on random positions, and perceive the board in larger chunks built through years of practice. | Transfer outside chess. | Original (page scan read as images). |
| `sparrow2011google` | B. Sparrow, J. Liu and D. M. Wegner, "Google Effects on Memory," *Science* 333(6043): 776–778 (2011). | pp. 776–778 of the paper. | Experiment 2: people recalled statements worse when they believed the computer had saved them (erase M = 0.31 vs save M = 0.22); Experiment 4: people recalled where a statement was saved (M = 0.49) better than the statement itself (M = 0.23). | Long-term harm to memory; Experiment 1 (the "Google Stroop" priming effect) did not replicate, see the next two entries. | Original (author-hosted PDF, read as images; abstract also via Europe PMC). |
| `camerer2018replicability` | C. F. Camerer et al., *Nature Human Behaviour* 2(9): 637–644 (2018). | Abstract. | Of 21 replications of social-science experiments from *Nature* and *Science*, 13 (62%) found a significant effect in the original direction, with effect sizes about half the originals; Sparrow et al. was among those that did not replicate. | Which parts of Sparrow et al. beyond Experiment 1 were tested. | Indexed abstract (Europe PMC); the Sparrow result via the project's replication reports as found in the search. |
| `hesselmann2020google` | G. Hesselmann, *PeerJ* 8: e10325 (2020). | Abstract. | A further replication found no conclusive evidence for the Google Stroop effect. | Anything about Experiments 2–4. | Indexed abstract (Europe PMC). |
| `kahneman2009intuitive` | D. Kahneman and G. Klein, *American Psychologist* 64(6): 515–526 (2009). | Abstract. | Whether an intuitive judgment can be trusted depends on the predictability of the environment and on the person's opportunity to learn its regularities; subjective confidence is not a reliable indicator of accuracy. | Anything about AI. | Indexed abstract (Europe PMC). |
| `bjork1994desirable` | R. A. Bjork, in J. Metcalfe and A. Shimamura (eds.), *Metacognition* (MIT Press, 1994), pp. 185–205. | pp. 185–205 of the chapter. | Conditions that speed acquisition during training can fail to support long-term performance, while conditions that introduce difficulties can enhance it (p. 185); continuous feedback and fixed conditions act like crutches that support performance during training and collapse afterwards (p. 196); the chapter's term "desirable difficulties" (p. 193). | Anything about AI; a review chapter. | Original (page scan read as images; page numbers printed on the pages). |
| `ericsson1993practice` | K. A. Ericsson, R. Th. Krampe and C. Tesch-Römer, *Psychological Review* 100(3): 363–406 (1993). | Bibliographic record (Crossref); the paper's core claim as restated in the 2019 replication. | Expert performance is attributed to prolonged, effortful, feedback-rich practice designed to improve performance. | That practice alone explains expertise; the next entry. | Bibliographic record original; content via `macnamara2019revisiting` (original abstract). The APA full text was not reachable. |
| `macnamara2019revisiting` | B. N. Macnamara and M. Maitra, *Royal Society Open Science* 6(8): 190327 (2019). | Abstract. | A replication of the 1993 violinist study did not reproduce the finding that accumulated deliberate practice corresponded to each skill level; the effect was substantial but considerably smaller. | That practice does not matter. | Indexed abstract (Europe PMC). |
| `noy2023productivity` | S. Noy and W. Zhang, *Science* 381(6654): 187–192 (2023). | Abstract. | 453 college-educated professionals doing writing tasks: time fell by 40%, quality rose by 18%, inequality between workers decreased. | Learning or skill effects; short tasks. | Indexed abstract (Europe PMC). |
| `brynjolfsson2025generative` | E. Brynjolfsson, D. Li and L. Raymond, *QJE* 140(2): 889–942 (2025). | Abstract of the published version. | 5,172 customer-support agents: productivity up 15% on average; less experienced and lower-skilled workers improved both speed and quality, while the most experienced saw small gains in speed and small declines in quality. | Generality beyond one firm. | Indexed abstract (Crossref record of the journal version). The NBER working paper (w31161) reports 5,179 agents, 14% and 34%; the manuscript uses the published numbers. |
| `peng2023copilot` | S. Peng, E. Kalliamvakou, P. Cihon and M. Demirer, arXiv:2302.06590 (2023). | Abstract; p. 3 of the PDF. | 95 professional programmers recruited; those with Copilot finished an HTTP-server task 55.8% faster. | Code quality or learning; one task. | Original (arXiv page; PDF page read as an image). |
| `dellacqua2026frontier` | F. Dell'Acqua et al., *Organization Science* 37(2): 403–423 (2026); HBS Working Paper 24-013 (2023). | Abstract (HBS page and Crossref record of the journal version). | 758 consultants; inside the frontier, AI users completed 12.2% more tasks, 25.1% faster, with significantly higher quality; on a task outside the frontier, AI users were 19% less likely to produce correct solutions. | Learning effects; one firm, one session. | Indexed abstract (HBS faculty page read through a proxy; Crossref). The abstract says "19% less likely," not "19 percentage points." |
| `metr2025developers` | J. Becker, N. Rush, E. Barnes and D. Rein, arXiv:2507.09089 (2025); METR report, 10 July 2025. | arXiv abstract; METR report page. | 16 experienced developers, 246 real tasks: allowing AI increased completion time by 19%; developers had forecast a 24% speedup and afterwards believed a 20% speedup. | A lasting slowdown; see the next entry. | Original (both pages). |
| `metr2026update` | J. Becker, N. Rush, T. Cunningham, D. Rein and K. Mahamud, METR blog, 24 February 2026. | The post. | Late-2025 data from 57 developers and 800+ tasks: estimated speedup −18% (CI −38% to +9%) among returning developers and −4% (CI −15% to +9%) among new recruits; METR calls the data an unreliable signal because developers increasingly declined to take part. | A measured speedup. | Original. |
| `bastani2025guardrails` | H. Bastani, O. Bastani, A. Sungu, H. Ge, Ö. Kabakcı and R. Mariman, *PNAS* 122(26): e2422633122 (2025). | Abstract and results, via the open full text. | Nearly 1,000 students in about fifty 9th–11th grade classes at a large high school in Turkey: practice grades up 48% with GPT Base and 127% with GPT Tutor; on the unassisted exam, GPT Base students scored 17% lower than controls, while GPT Tutor students were statistically indistinguishable from controls and did not do better. | That guard-railed AI improves learning; one school, one model, short run. An erratum exists (PNAS 122(34): e2518204122). | Indexed full text (Europe PMC). |
| `kestin2025tutoring` | G. Kestin, K. Miller, A. Klales, T. Milbourne and G. Ponti, *Scientific Reports* 15: 17458 (2025). | Full text. | 194 college students; median learning gains with the AI tutor were over double those of in-class active learning (effect size 0.63 by regression, 0.73–1.3 after correcting for a ceiling), with a median of 49 minutes on task; the tutor was designed to guide students sequentially and manage cognitive load. | Long-term retention. | Indexed full text (Europe PMC). |
| `desimone2025nigeria` | M. De Simone et al., World Bank Policy Research Working Paper 11125 (May 2025). | Crossref record of the paper; the authors' World Bank blog post of 9 January 2025. | A six-week after-school programme in Edo, Nigeria, using generative AI: students randomly assigned improved by about 0.3 standard deviations, described by the authors as nearly two years of typical learning; girls gained even more. | Effects without a teacher present. | Bibliographic record original; numbers by author summary (the repository pages were not reachable). |
| `shen2026skill` | J. H. Shen and A. Tamkin, arXiv:2601.20245 (2026). | Abstract and the HTML full text. | 52 developers learning a new library: those using AI scored 17% lower (two grade points; Cohen's d = 0.74, p = 0.01) on conceptual understanding, code reading and debugging, with no significant time saving; three of six interaction patterns preserved learning. | Generality; a preprint from an AI company with a small sample. | Original (arXiv). |
| `fan2025laziness` | Y. Fan et al., *British Journal of Educational Technology* 56(2): 489–530 (2025). | Abstract. | 117 university students in four conditions: the ChatGPT group improved essay scores most, but knowledge gain and transfer did not differ. | Long-term effects. | Indexed abstract (Crossref). |
| `otis2025uneven` | N. G. Otis, R. Clarke, S. Delecourt, D. Holtz and R. Koning, OSF preprint, version 3 (3 October 2025). | Abstract. | Kenyan entrepreneurs randomised to a GPT-4 business assistant: no average effect; low performers did nearly 10% worse and high performers may have gained over 15%, a gap the authors trace to which pieces of advice were selected and implemented. | The mechanism beyond the authors' interpretation; a preprint. | Original (OSF API record). |
| `lee2025critical` | H.-P. Lee et al., CHI '25, Article 1121 (2025). | Abstract (Microsoft Research page). | 319 knowledge workers, 936 examples: higher confidence in generative AI is associated with less critical thinking, higher self-confidence with more. | Objective change in skill; self-report. | Original (publisher-affiliated page; DOI record). |
| `gerlich2025offloading` | M. Gerlich, *Societies* 15(1): 6 (2025); correction 15(9): 252. | Abstract; the correction's record. | 666 participants: frequent AI use correlated negatively with critical thinking, mediated by cognitive offloading; younger participants showed higher dependence. | Causation; cross-sectional and self-reported. | Indexed abstract (Semantic Scholar); correction via Crossref. |
| `kosmyna2025cognitive` | N. Kosmyna et al., arXiv:2506.08872 v2 (2025). | Abstract. | 54 participants writing essays with an LLM, a search engine or nothing: LLM users showed the weakest EEG connectivity, the lowest sense of ownership, and struggled to quote their own work. | Long-term decline; small sample, unreviewed. | Original (arXiv). |
| `fernandes2026wiser` | D. Fernandes et al., *Computers in Human Behavior* 175: 108779 (2026). | Abstract. | 246 and 452 participants solving LSAT problems with AI scored about three points above a norm population but overestimated their performance by about four points; higher AI literacy went with less accurate self-assessment. | Generality beyond LSAT-style reasoning. | Indexed abstract (Semantic Scholar); record via Crossref. |

## Considered and not cited

- N. Carr, "Is Google Making Us Stupid?" (*The Atlantic*, July/August 2008): the text could not be opened, so it is not cited.
- C. Goldin and L. F. Katz, *The Race between Education and Technology* (2008): not needed for the argument.
- A. Agrawal, J. Gans and A. Goldfarb, "The Simple Economics of Machine Intelligence" (*Harvard Business Review*, 17 November 2016): only the opening could be read; the book and the NBER paper carry the same point.

## Positioning

The essay argues from the literature; it reports no measurement of its own. Every empirical number in it is quoted with its sample size and its source, and the manuscript says where a result is a preprint, where it failed to replicate, and where the authors themselves call their data unreliable. The argument does not depend on any single study: the pattern that matters is that tools which hand over answers help people who can judge them and hurt the learning of people who cannot, and that pattern appears in school mathematics, consulting, software, writing and entrepreneurship.