Bibliography
Primary sources for the v3 library, organized by argument arc. Every entry is cited in at least one article; every citation in the articles appears here.
Evidence strength: ● Strong (multiple studies, large n, peer-reviewed) · ◐ Moderate (single quality source) · ○ Weak (expert opinion, preprint) · ◌ Speculative (inference)
The Paradox
Productivity gains and learning harm — coexisting, documented, and invisible from inside.
◐ Becker, J. et al. / METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv preprint. AI made experienced devs 19% slower (95% CI: +1.6% to +39%); devs predicted 24% speedup. 43-percentage-point perception gap.
Design & abstract
Within-subject RCT, n=16 experienced open-source developers, 246 real tasks from their own repositories. Random assignment at issue level (each dev did both conditions). $150/hour compensation. Tools: Cursor Pro with Claude 3.5/3.7 Sonnet. 143 hours of screen recordings. AI increased completion time by 19%. Developers predicted 24% speedup pre-study and still believed 20% speedup post-study. Not peer-reviewed, not confirmed pre-registered. Limitation: small n, experienced devs only (>5 years OSS), own repos (high familiarity).
● Cui, Z., Demirer, M. et al. (2024). Effects of Generative AI on High Skilled Work. SSRN working paper. 3 RCTs at Microsoft, Accenture, Fortune 100. 26% more tasks completed with AI (LATE for adopters). Less experienced developers benefited more.
◐ Shen, J.H. & Tamkin, A. (2026). How AI Impacts Skill Formation. Anthropic. arXiv preprint. Cohen's d = 0.738 (17pp skill gap). Six interaction patterns: 3 preserve learning, 3 don't. Generation-Then-Comprehension (86%) vs AI Delegation (39%).
Design & abstract
RCT, n=52 (26 control, 26 treatment), junior Python developers learning Trio library. 35-minute coding task + comprehension quiz (~58.5 min total). AI group scored 50% vs control 67% on quiz. Largest gap on debugging questions. High-scoring patterns: Generation-Then-Comprehension (86% mastery), Conceptual Inquiry (second-fastest). Chat-based interface (not agentic tools). Crowdworkers, not workplace context. Four pilot studies required due to non-compliance. Immediate comprehension only — no longitudinal retention data.
● Bastani, H. et al. (2025). Generative AI without guardrails can harm learning. PNAS. GPT Base: +48% during, -17% on unassisted exam. GPT Tutor (hints): +127% during, harm largely mitigated. Same model, different design, different trajectory.
Design & abstract
RCT (field experiment), n≈1,000 Turkish high school math students, three arms. GPT Base (unrestricted ChatGPT-4): +48% grades with AI, -17% on exams without AI vs control. GPT Tutor (hint-only, guardrailed): +127% grades with AI, negative effects "largely mitigated." Mechanism: "Without guardrails, students use GPT-4 as a 'crutch' during practice." PNAS peer-reviewed (received Nov 2024, accepted May 2025). Minor correction published Aug 2025 (affiliation typo only).
◐ Budzyń, B. et al. (2025). Endoscopist deskilling risk after AI exposure in colonoscopy. Lancet Gastroenterol. Hepatol. 21% relative decline in unaided adenoma detection (28.4% → 22.4%) after AI exposure. Observational, n=1,443.
Design & abstract
Multicentre observational study (before/after design), 19 endoscopists, 4 centres in Poland, Sept 2021–March 2022. Adenoma detection rate on non-AI colonoscopies compared before vs after AI-assisted colonoscopy was introduced. Pre-AI: 28.4% (n=795). Post-AI exposure: 22.4% (n=648). 21% relative decline, 6pp absolute. NOT a crossover RCT — AI was not "removed." Limitation: observational, single specialty, cannot establish causation definitively. Lancet peer-reviewed.
◐ Natali, L. et al. (2025). Deskilling and upskilling inhibition in medicine after AI adoption. Distinguishes deskilling (experts lose what they have) from "upskilling inhibition" (novices never develop skills when AI provides answers first).
◐ Liang, J.T. et al. (2024). Usability of AI Programming Assistants. ICSE 2024. n=410 developers. Primary uses: reducing keystrokes, recalling syntax, finishing fast. Primary barrier: control — can't get tool to generate what they want.
The Mechanism
How cognitive offloading works — trust, explanations, the generative step, encoding.
◐ Chen, Z. et al. (2025). AI-Assisted Note-Taking. PACMHCI. n=30, within-subject design. Full automation → lowest post-test scores; intermediate AI → highest; manual → most effortful but strong comprehension. Preferred ≠ effective.
● Lee, H.P. et al. (2025). The Impact of Generative AI on Critical Thinking. CHI 2025. β=-0.69 (AI confidence → less critical thinking, p<0.001). Self-confidence effects (β=+0.26, p=0.026; β=+0.31, p=0.046) do not survive Benjamini-Hochberg correction (threshold p<0.007). Uses mixed-effects regression, n=319 workers, 936 task examples.
Design & abstract
Cross-sectional survey with mixed-effects regression modeling, 319 knowledge workers across 936 task examples (CHI peer-reviewed). AI-confidence strongly predicts reduced critical thinking (β=-0.69). Self-confidence effects show correct direction but do not survive the paper's own multiple comparisons correction (Benjamini-Hochberg threshold p<0.007). Limitation: cross-sectional, not longitudinal — correlation direction is assumed, not experimentally established. Reverse causation is possible.
Sharma, M. et al. (2026). Who's in Charge? Disempowerment Patterns. Anthropic. arXiv preprint. ~1.5M Claude.ai conversations. Severe disempowerment potential: <1 in 1,000 conversations. Disempowering interactions receive higher approval ratings.
Design & abstract
Large-scale observational analysis of ~1.5 million Claude.ai conversations (Anthropic, January 2026). Conversations flagged as moderate-or-severe disempowerment potential show positivity rates above the baseline across all categories. Causal mechanism is uncertain; several explanations possible; paper acknowledges this explicitly. Implication for training: approval ratings feed preference models, so systems trained on approval data may drift toward disempowerment. arXiv preprint, single platform (Claude.ai only).
● Bansal, G. et al. (2021). Does the Whole Exceed its Parts? CHI 2021. AI explanations increase acceptance regardless of correctness. Complementary performance gains NOT increased by explanations. A wrong AI that explains itself confidently is more dangerous than one that doesn't.
○ Kosmyna, N. et al. (2025). Your Brain on ChatGPT. MIT Media Lab. arXiv preprint. 83% of LLM group unable to quote own essays (Session 1 behavioral finding); EEG methodology disputed (published critique by Stankovic et al.).
Design & abstract
Between-subjects EEG study, n=54 (18 per group: LLM, Search Engine, Brain-only). arXiv preprint — not peer-reviewed. Session 1: 83% (15/18) of LLM group couldn't quote any passage from essays written minutes earlier, vs ~11% in other groups. Published critique: Stankovic et al. (arXiv:2601.00856) identifies concerns with study design, reproducibility, EEG methodology, and reporting inconsistencies. Authors themselves state conclusions should be "treated with caution and as preliminary." The behavioral recall finding (83% failure) does not depend on the disputed EEG interpretation.
◐ Siddiqui, T. et al. (2025). AI-Supported Writing Process. AIED. RCT, n=30 per condition. Correlation between integrated writing tool use and knowledge transformation: r=0.608 (p=0.001). Correlation between chat-based LLM use and knowledge transformation: r≈0.
◐ Kazemitabaar, S. et al. (2025). Exploring the Design Space of Cognitive Engagement Techniques with AI-Generated Code for Enhanced Learning. IUI 2025. N=82 + N=42 = 124 total. Seven "cognitive forcing functions" tested — techniques requiring learners to engage before seeing AI solutions. Study 1 (between-subjects): Lead-and-Reveal highest mean (M=58.3 vs baseline M=31.5), but p=.058 (did not reach corrected alpha), study underpowered (eta-squared=.112, needed N=153). Study 2 (within-subjects, N=42): Trace-and-Predict had highest raw performance (M=64.3), Lead-and-Reveal lowest (M=55.0), difference non-significant (p=.150). Lead-and-Reveal's unique contribution: metacognitive calibration (r=.26, p=.017 — the only condition where self-assessment significantly correlated with actual performance). All techniques cost 1.5-2.7x more time. Population: novice undergrads (18-23), data structures course.
Design, techniques & verified provenance
Two studies (Study 1: N=82 between-subjects, Study 2: N=42 within-subjects), IUI peer-reviewed. Novice undergraduate programmers (18-23) in a data structures course. Seven cognitive engagement techniques across a three-dimensional design space (cognitive level × timing × engagement type): Lead-and-Reveal (Create, Engage-then-Reveal), Trace-and-Predict (Apply, Reveal-then-Engage), Solve-Code-Puzzle (Apply, Reveal-then-Engage), Verify-and-Fix (Evaluate, Reveal-then-Engage), Explain-in-Plain-English (Understand, Reveal-then-Engage), Write-Test-Cases (Apply, Engage-then-Reveal), Interactive-Pseudo-Code (Create, Engage-then-Reveal). The paper calls these "cognitive forcing functions" — techniques that prevent immediate use of AI-generated solutions. Study 1: Lead-and-Reveal had highest mean learning gain (M=58.3, SD=39.8) vs baseline (M=31.5, SD=28.2), b=26.8, p=.058 — suggestive but not significant after Bonferroni correction (alpha=.007). Verify-and-Review caused significantly higher frustration (M=72.7 vs 46.7, p=.035). Study 2: refined top techniques (V2). Trace-and-Predict V2 highest performance (M=64.3), baseline middle (M=57.6), Lead-and-Reveal V2 lowest (M=55.0); Friedman chi-squared=3.79, p=.150. But Lead-and-Reveal was the only condition with significant metacognitive calibration (predicted-actual correlation r=.26, p=.017 vs baseline r=.16 ns, Trace-and-Predict r=.15 ns). Trace-and-Predict had significantly higher cognitive load (TLX M=67.1 vs baseline 58.9, F=7.4, p=.001) and 2.66x time cost. Limitations: underpowered (needed N=153 for 90% power), novice-only, single session, immediate assessment only, no delayed retention measure, no counterbalancing in Study 2. Provenance: Claims extracted via Claimify pipeline, independently verified via CoVE (38 verified, 12 corrected, 12 refuted — Study 2 quantitative extraction required full re-verification against source). Synthesis at .research/synthesis/cognitive-forcing-functions.md.
◐ Galbraith, D. (2007). Writing as Thinking. Writing and Cognition: Research and Applications. Writing constitutes thought — the act of text production activates and reorganizes knowledge in long-term memory. Not merely expression of pre-formed ideas but a generative epistemic process. Theoretical basis for why AI-generated text bypasses the cognitive work that produces understanding.
○ Storey, M.A. (2026). How Generative and Agentic AI Shift Concern from Technical Debt to Cognitive Debt. Blog. Cognitive debt: the accumulated loss of understanding that occurs when AI generates code humans don't fully comprehend. Parallels technical debt — invisible until a crisis demands understanding that was never built.
○ Umarova, Z. et al. (2025). Writer-AI Interactions in Writing Process. Cornell. Qualitative case study. Proactive idea exploration produced new ideas regardless of tool type; prolonged mindless copyediting produced few ideas even with Socratic AI.
◐ Slamecka, N.J. & Graf, P. (1978). The Generation Effect: Delineation of a Phenomenon. Journal of Experimental Psychology: Human Learning and Memory. Self-generated information retained ~22% better than passively received information. Replicated across decades of cognitive psychology.
◐ Bereiter, C. & Scardamalia, M. (1987). The Psychology of Written Composition. Knowledge telling vs knowledge transforming — expert writing as a recursive process that generates new understanding through the act of composition. Theoretical basis used in Siddiqui et al. 2025.
◐ Pallant, J.I. et al. (2025). Mastering Knowledge with GenAI. Studies in Higher Education. QCA of 192 student reflections. Mastery-oriented GenAI use: OR=35.782 (p<0.001) for critical thinking. Not RCT — self-selection possible.
The Design Lever
Which design features move outcomes — process control, transparency, engagement, orientation.
◐ Blaurock, M., Büttgen, M., & Schepers, J. (2025). Designing Collaborative Intelligence Systems for Employee-AI Service Co-Production. Journal of Service Research, 28(4), 544-562. Study 2 on perceived outcome responsibility: process control b=0.715 (p<.001), outcome control b=0.524 (p<.011), transparency b=0.511 (p<.001), engagement b=0.090 (ns). All four engagement hypotheses rejected. Expertise reversal: control features significant only for AI novices.
Design & abstract
Two scenario-based experiments (NOT a meta-analysis). Study 1: n=309 financial services employees. Study 2: n=345 HR professionals. Combined n=654. Finds process control, transparency, and outcome control are the most important design features for employee-AI co-production; engagement features are not significant or counterproductive. Exploratory post-hoc: AI-experienced subgroup (n=42) showed no significant positive effects for control features; engagement showed what authors call "significant negative effect" on perceived service improvement (b=0.555, p<.05, statistical twin sample — note: source reports positive coefficient but describes as negative effect, a reporting inconsistency). Peer-reviewed, Journal of Service Research.
The Stakes
Homogenization, model collapse, and what happens at the systemic level.
◐ Xu, Y. et al. (2025). Echoes in AI: LLM Homogenization. PNAS, Microsoft Research. 100 GPT-4 completions of a Kafka story: 50/100 had the policeman give directions (second left), 16/100 featured the same bakery. Human-written stories stayed unique.
● Jiang, L. et al. (2025). Artificial Hivemind: Open-Ended Homogeneity of LLMs. NeurIPS 2025. 70+ LLMs converge on the same outputs across model families. Sentence embedding similarity 71-82%. RLHF reward models miscalibrated to idiosyncratic preferences — penalize diversity.
● Doshi, A.R. & Hauser, O.P. (2024). Individual Creativity vs Collective Diversity. Science Advances. Preregistered, n=893. +8.1% individual novelty (b=0.311, p<0.001), +10.7% collective similarity. "Social dilemma: individually better off, collectively homogenized."
○ Holzner, N., Maier, S., & Feuerriegel, S. (2025). Generative AI and Creativity: A Meta-Analysis. arXiv (submitted to ACM CHI, not yet peer-reviewed). 28 studies, n=8,214. Individual performance: g=+0.273. Idea diversity: g=-0.863 (95% CI: [-1.328, -0.398], p<0.001). Leave-one-out: g=-0.655 to g=-0.952, all CIs exclude zero.
● Ashery, A.F., Aiello, L.M., & Baronchelli, A. (2025). Emergent Social Conventions and Collective Bias in LLM Populations. Science Advances. Populations of LLM agents spontaneously develop social conventions. Strong collective biases emerge even when individual agents show no bias — bias is a property of the interaction, not the components.
◐ Masud, S. et al. (2025). Epistemic Diversity of LLM Outputs. arXiv:2510.04226. 27 language models, 155 topics, 70 million claims. Every model produced outputs less epistemically diverse than a basic web search. Larger models less diverse than smaller ones.
○ Wan, Y. & Kalman, Y.M. (2025). AI Personas Preserve Diversity. arXiv preprint. Modified replication of Doshi & Hauser. 10 distinct AI personas: intra-persona similarity 0.92, inter-persona similarity 0.20. Homogenization effect eliminated. Proof-of-concept; not independently replicated.
● Shumailov, I. et al. (2024). Model Collapse. Nature, Vol 631. Training generative models on model-generated content causes progressive loss of distribution tails. Early collapse: rare content disappears. Late collapse: convergence to very small variance. "Irreversible defects." Demonstrated on LLMs, VAEs, and Gaussian mixture models.
◐ Dohmatob, E. et al. (2025). Strong Model Collapse. ICLR 2025, FAIR/Meta. Even 1% synthetic data fraction can lead to model collapse; larger models amplify collapse. IMPORTANT: 1% figure derived from supervised linear regression theoretical setting, not production LLM training.
◐ Gerstgrasser, M. et al. (2024). Is Model Collapse Inevitable? arXiv. Stanford/MIT/Maryland/Harvard. Replacing real data with synthetic → inevitable collapse (unbounded test error). Accumulating synthetic alongside real data → bounded error. Escape condition requires continuing human data generation.
● Hong, L. & Page, S.E. (2004). Groups of Diverse Problem Solvers Can Outperform Groups of High-Ability Problem Solvers. PNAS. Formal mathematical proof: diversity of perspective can trump individual ability in collective problem-solving.
○ Haldane, A. (2009, 2016). Financial Monoculture. Bank of England. When system components become highly correlated, aggregate risk approaches any single component's risk. Diversification becomes illusory.
Entries with provenance notes have been verified through structured claim extraction, independent verification, and synthesis. All effect sizes cite original sources.