Records · Evidence review

Impact review — AI pages (/ai/do/, /ai/practice/, /ai/learn/)

The working record behind the pages, published as written. Source file source/impact/ai-2026-11-05.md. All records.

Reviewed 2026-11-05 (fetches 20:54–22:40 EST). Reviewer: Claude (Fable 5.1), with one status-check subagent for organisations and laws. Content pages were not edited; nothing committed.

0. Method

Decision being informed. Which actions the AI Do and Practice pages should recommend, in what order, and with what evidence grade. The decider is Isaac. The population is one person with no special position, a few hours a month and a modest budget. What is being estimated: whether each action protects the reader's (and others') learning, attention, grip on what is true, or safety, as the pages frame those harms.

Grades, fixed before reading the results.

• strong — two or more independent randomised or field experiments, or a meta-analysis, pointing the same way on the outcome the page claims.
• moderate — one well-powered RCT or quasi-experiment, or several consistent smaller or short-term studies, or RCTs on a close proxy of the page's action.
• weak — small, preprint, self-report or correlational, or the measured outcome is a proxy (e.g. "recognises a manipulation technique" rather than "shares less falsehood").
• none-but-plausible — no study of the action itself; a plausible mechanism; no evidence of harm.
• contradicted — the best available evidence finds no effect, or the opposite, on the outcome the page claims.

Provenance tags used below. [F] opened and read by me today; [F-sub] opened by the status subagent today (its observations trusted; its three load-bearing conclusions re-checked by me against the primary page, noted where done); [F-earlier] opened in the earlier 2026-11-05 session and recorded in source/research/ai.md (same day, not re-opened by me); [record] an identifier lookup (Europe PMC, OpenAlex, Semantic Scholar, PMC ID converter) whose abstract or metadata I read, not the paper; [unread] seen only as a search snippet or not reachable. Grades for sources: primary / research / preprint / charity's own figures / news / industry.

What limited the work. The session's shared web-search budget was exhausted after my first query; the rest of the work used direct URLs, DOI and title lookups in scholarly databases, and publisher pages. Publisher blocks (PNAS, Science, OUP, SAGE, Elsevier, SSRN, OpenAI, World Bank repository) were worked around through PMC, Europe PMC full text, OpenAlex/Semantic Scholar records, saved PDFs read with pdftotext, and author or government pages. Where the only reachable copy is an abstract, the grade says "abstract-level".

---

A. The evidence behind each action on the pages

A1. Learning: "think first, hints not answers, make it argue, check facts, keep one skill AI-free"

The mode-of-use finding (page's lede and step 1). Grade: moderate for "tutor-style use beats answer-style use"; weak for "typing 'don't give me the answer' reproduces a designed tutor".

• Bastani et al., PNAS 122(26), 2025 [F] research. RCT at classroom level, "nearly 1,000 students" in about fifty 9th–11th-grade classes at one high school in Turkey, Fall 2023, four 90-minute sessions. Three arms: control (textbook, no device); GPT Base (GPT-4 chat); GPT Tutor (GPT-4 with teacher solutions, common mistakes and an instruction to give hints, not answers). Assisted practice: GPT Base +0.137 on a 0–1 scale (control mean 0.284, about +48%), GPT Tutor +0.361 (+127%). Unassisted exam: GPT Base −0.054 (SE 0.022, p<0.05, about −17%); GPT Tutor −0.004 (SE 0.013, not significant). Students in GPT Base did not perceive they had learned less; GPT Tutor students thought they had done better although they had not. Limits stated by the authors: one subject (maths), one school, Fall 2023 models, short-term outcomes only. The page's "17% worse" and "hints lost nothing" are correct; "about a thousand students" is correct ("nearly 1,000").
• Anthropic, How AI assistance impacts the formation of coding skills, 29 Jan 2026 [F] research (company). RCT, 52 mostly junior engineers; quiz 50% (AI) vs 67% (control), d=0.738, p=0.01; AI group two minutes faster, not significant. The usage-pattern analysis (conceptual questions ≈65%, delegation ≈39%) is descriptive, n=2–7 per pattern; the authors say it "does not establish a causal link".
• Contractor & Reyes, Experimental Evidence on the Learning Impact of Generative AI, arXiv 2607.08849 v2, 4 Oct 2026 [F] preprint. Proctored RCT, 211 Middlebury undergraduates, Spring 2025, GPT-4o. Immediate test +7.2 pp (0.28 SD, p=0.021); one week later +3.9 pp (0.21 SD, p=0.068). "Augmentation" users kept a gain a week later (0.24 SD, p=0.095), "automation" users did not (−0.24 SD, ns) — but use type was coded after the fact, not randomised. Rule violations could account for 17–22% of the immediate effect. The page's "one 2026 preprint found a lasting gain" is fair if "lasting" means one week.
• Kestin, Miller, Klales, Milbourne & Ponti, Scientific Reports 15:17458, 2025 [F] research. Randomised crossover, 194 eligible of 233 Harvard physics students, Fall 2023; a GPT-4 tutor built on the same pedagogy as the in-class active-learning lesson (expert-written, question-specific prompts, cognitive-load management, growth-mindset framing, instructional videos). Median post-test 4.5 vs 3.5; effect size 0.63 (regression), 0.73–1.3 SD (quantile regression avoiding ceiling); median time on task 49 min vs 60; engagement 4.1 vs 3.6 (p<0.0001). Authors: gains "may depend on" expert prompts and structure; they "do not presume" the tutor beats active learning for higher-order synthesis. This is the strongest positive tutor result, and it is a designed tutor, not a reader's prompt.
• De Simone et al., World Bank, From Chalkboards to Chatbots (blog 9 Jan 2025 [F]; working paper 11125, May 2025 [record] OpenAlex abstract; repository page blocked). RCT, after-school programme, six weeks June–July 2024, Edo State, Nigeria; Microsoft Copilot (GPT-4) for English with teacher support. Overall assessment +0.31 SD, English +0.23 SD, "equivalent to 1.5–2 years" of business-as-usual schooling; girls gained more; attendance disrupted by flooding and strikes. Number of students not retrieved today.
• Wang, Ribeiro, Robinson, Loeb & Demszky, Tutor CoPilot, arXiv 2410.03017 v3 (10 Oct 2026) [F] preprint. RCT with 900 human tutors and 1,800 K-12 students: AI assisting the human tutor raised mastery +4 pp (p<0.01), +9 pp for students of lower-rated tutors; $20 per tutor per year. Different action (AI behind a human), but relevant to "who benefits".
• Meta-analyses, kept as a dispute: Wu et al., Humanities & Social Sciences Communications 13, 2026 [record] — 35 experiments, 4,193 participants, g=0.670, no publication bias detected (abstract-level). Deng et al., Computers & Education 227, 2024 [record] — 69 articles, positive, "methodological limitations… call for cautious interpretation". Wang & Fan 2025 (g=0.867) was retracted on 22 April 2026 "due to concerns relating to discrepancies in the meta-analysis"; the authors did not respond [F]. Bartoš, Martinková & Wagenmakers, PsyArXiv, 12 May 2025 [record] OSF abstract: once publication bias is modelled, Wang & Fan's effects "greatly diminish… and the evidence in favor of the benefits disappears." Any page figure drawn from "meta-analyses show AI helps learning" should not be used.
• Kosmyna et al., Your Brain on ChatGPT, arXiv 2506.08872 v2 (31 Dec 2025) [F] preprint: 54 participants, 18 in session 4; "Brain-to-LLM" group showed higher recall than LLM-first. Comment by Stankovic et al., arXiv 2601.00856 (29 Dec 2025) [F] preprint: sample size, reproducibility, EEG methods, reporting inconsistencies, transparency. The page's labels ("unreviewed, small; 18 people") are correct.
• Lee et al. (Microsoft Research / CMU), CHI 2025 [F] research: survey, 319 knowledge workers, 936 examples; higher confidence in the AI associated with less critical thinking. Correlational self-report. The page's "Make it argue against you" rests on this; the prompt itself has never been tested.

Grades for the page's learning actions.

A2. Truth: "find the original, don't trust your eyes or a detector, don't swing to all-fake, safe word"

Note for the Practice drill: Diel's pooled figures make the drill's "you can't tell by looking" line stronger than the page's current single citation.

A3. Attention and feeds

The Haidt–Odgers dispute, kept as a dispute: Odgers, Nature, 29 Mar 2024 [F] — "Most data are correlative", associations "no, small and mixed", reverse causation (Heffer 2019), 72-country roll-out study with no consistent association (Vuorre & Przybylski 2023), and the warning that the story "might distract us from effectively responding to the real causes"; she calls device bans and age limits "unlikely to work and possibly counterproductive". Haidt, After Babel, 22 Feb 2023 [F] — r≈.20 for girls once sex and social media are separated, 12 of 18 experiments and 24 of 33 longer-interval longitudinal studies showing harm, and "effects that accumulate over time and at scale". His direct reply to Odgers was not fetched.

A4. Kids and companion bots

A5. Policy: "find your representative and weigh in"

A6. Settings (training and memory opt-outs)

Grade: none-but-plausible for any learning/attention/truth/safety outcome; this is a privacy act, and the page says so honestly. Mechanics: Claude and Gemini paths from primary pages [F-earlier]; Meta's 1 Oct 2025 post states no opt-out from AI-chat ad personalisation [F-earlier] [F-sub]; every OpenAI page returned 403 today [F-sub], so the ChatGPT toggle label stays unverified, as the page already admits.

---

B. The highest-impact things one person can do, ranked

Ranking rule: evidence grade on the stated outcome × size of the stake for the reader and others × feasibility at a few hours a month. Each entry says what would overturn it.

1. When you are learning something, use AI only in tutor mode, and keep practising without it. Grade moderate. Bastani (−17% on the exam for answer-mode, no harm for hint-mode, ~1,000 students), Contractor & Reyes (211; augmentation gains survived a week, automation gains did not), Anthropic (52; d=0.74), Kestin (194; a designed tutor beat active learning by 0.63–1.3 SD in less time), Nigeria (+0.31 SD in six weeks with teacher support). The honest gap: nobody has randomised what the user types. Overturned if a trial finds tutor-style use erodes unaided performance over a semester, or answer-style use does not.
2. Before you repeat or share anything striking, find the original and pause on "is this accurate?" Grade moderate-to-strong. Accuracy prompts: 20 experiments, 26,863 people, false-headline sharing −10%; lateral-reading instruction: 7% → 61% correct lateral evaluations; human and machine detection: ~55% and falling in the wild, so process beats inspection. Communal: this is the one action whose benefit lands on other people, by cutting what you spread. Overturned if field studies show self-administered accuracy checks decay to nothing within days (Guess 2020 already shows the tips effect halving in three weeks).
3. Take one platform, or all mobile internet, off your phone for two to four weeks; then keep feeds on Following. Grade moderate, with known costs. Allcott 2020 (well-being +0.09 SD, 60 min/day freed, news knowledge −0.19 SD), Castelo 2025 (well-being d=0.45, attention d=0.23, low compliance), Guess 2023 (chronological = less time, no attitude change). Do it for attention and time; it will not fix politics (Allcott 2024's near-zero polarisation estimates, 35,000 people). Overturned if larger pre-registered trials find no well-being gain, or that the attention gain reverses on return.
4. If you have a kid: delay the smartphone and keep companion bots out, with other parents, and back a phone-free school day. Grade moderate for school and bullying outcomes, contested for wellbeing. Beland & Murphy (England, difference-in-differences: +6.41% SD on exams, +14.23% SD for the lowest quartile, nothing for the top), Abrahamsson (Norway, J. Human Resources 2026, event study: no average effect, but less psychological care use among girls, less bullying for both, better girls' GPA), Böttger & Zierer (five studies, d=0.162, stronger for social well-being). Against: Goodyear 2025 (30 English schools, 1,227 pupils: no wellbeing difference between restrictive and permissive schools, −0.48 WEMWBS, CI −2.05 to 1.06; in-school use down 0.67 h but total use unchanged; cross-sectional) and Vereecke 2026 (one Belgian school, 960 pupils, eight waves: no improvement, loneliness up slightly, no control group). Delay itself rests on associations (Thiagarajan 2025: smartphone before 13 associated with worse adult outcomes, especially girls; design and n not retrieved) and on the Haidt–Odgers dispute. Pledges (Wait Until 8th, Smartphone Free Childhood) have no evaluation and cite no studies on their sites. Overturned if a randomised or well-controlled roll-out finds no academic or bullying effect; settled by the same design for wellbeing.
5. Contact your state legislator about one named AI bill, in your own words. Grade moderate. Two randomised field experiments at state level (Bergan 2009; Bergan & Cole 2015, +12 pp), congressional staff surveys (individualised messages ~20% "a lot" of influence vs 1% for form messages), and Broockman & Skovron (legislators misjudge their constituents because of who contacts them). AI law is being written in statehouses (NY, CA, CO, a dozen chatbot laws), which is exactly where the experiments were run. Overturned if replication in other states and on salient bills finds no effect; untested for AI bills specifically.

Not in the top five but worth keeping: the family safe word (cheap, unevidenced, the harm is real); turning off training (privacy, narrow); the companion-bot rule (now law-backed in two states, evidence of harm weak-to-moderate).

Where the evidence is young, and what would settle it. (a) Long-run effects of tutor-mode AI: a semester-long RCT with delayed unassisted tests. (b) Whether a reader's prompt reproduces a designed tutor: randomise the prompt. (c) Phone-free schools and wellbeing: staggered roll-outs with control schools (Goodyear and Vereecke are the weakest designs in this set and the most quoted against bans). (d) Companion bots and teens: anything beyond Fang's four weeks and Common Sense's expert ratings. (e) Whether individual contacts move AI bills: a Bergan-style experiment on a chatbot bill.

Drop / keep / re-rank (the Do page's current order)

---

C. Communal impact, backfire and group formats

C1. Who benefits from each kept action, and how much

• Tutor-mode AI. Benefit lands on the learner, and on institutions that write the tutor well: Nigeria's +0.31 SD came with teacher support; Tutor CoPilot's +4 pp (+9 pp for weaker tutors) came from AI behind a human tutor; Brynjolfsson, Li & Raymond (5,179 support agents) found +14% productivity with the gain concentrated in novices (+34%). Dell'Acqua et al. (758 BCG consultants): +12.2% tasks and 25.1% faster inside the frontier, but 19% less likely to be correct on a task outside it; a prompt-engineering briefing did not remove that. So the communal benefit depends on whether someone checks the model's reach.
• Find-the-original / accuracy pause. Benefit is mostly to others: the pooled effect is on what people share. It is modest (−10% false sharing) and decays; it also slightly lowers belief in true headlines (Guess 2020).
• Platform breaks and Following feeds. Benefit to the reader's time and mood; a cost to news knowledge (−0.19 SD; −0.098 SD) that the page should state. No measurable communal political effect at 35,000-person scale.
• Parents acting together. The quasi-experimental gains from school phone policies accrue most to the lowest-achieving quartile (+14.23% SD) and to girls (Norway); the null wellbeing findings are cross-sectional or single-school.
• Legislator contact. A 12-pp shift in a legislator's support on one bill is large relative to one person's cost; the backfire is structural: Broockman & Skovron show legislators already misread opinion because contactors are unrepresentative, so a message that states where you live and why counts more than a form.

C2. Harm and backfire, with evidence

• AI text detectors. Liang et al. 2023 [F]: seven detectors, 91 TOEFL essays, false-positive rate 61.3% for non-native writers, near zero for native 8th-graders; a prompt to "sound more like a native speaker" cut detection to 11.6%; authors "strongly caution against the use of GPT detectors in evaluative or educational settings". Groh 2022: a confident wrong detector output lowered human accuracy. The pages do not recommend detectors; they should say not to.
• Surveillance of children. The AAP's own wording is "keep loose tabs", "about care, not spying" [F]. CDT's Hidden Harms (Aug 2022) [F] reports LGBTQ+ students "disproportionately targeted" and low-income, Black and Hispanic students "at greater risk of harm" from school monitoring software (its survey figures are in the PDF, not read). Keep the page at "read a few chats together"; do not add monitoring software.
• Moral panic. Odgers [F]: unsupported stories "do little to support young people" and may distract from poverty, violence and under-staffed services (one school psychologist per 1,119 students). The page's kids section should keep its claims to what is shown (companion-bot harms in testing; associations for early smartphones) and not imply settled causation.
• Liar's dividend. Over-warning about fakes helps liars on text scandals (0.17–0.21 SD) and makes true "AI-generated" news less believed; the page's "don't swing to all-fake" line is the right counter, and fact-checks cancel the gain (Schiff).
• Deactivation. Lower news knowledge (both Allcott studies).

C3. Doing it with others: which group formats have evidence

---

D. Organisations, tools and laws the pages name

Status as of 2026-11-05. "Filing" = latest Form 990 year on ProPublica Nonprofit Explorer [F-sub]. Controversy window: Nov 2024–Nov 2026, news grade unless stated.

---

E. Corrections to make on the pages (priority order; none made)

1. EO 14365's title → "Ensuring a National Policy Framework for Artificial Intelligence".
2. Kids section: AAP names one warning sign (withdrawn, prefers chatbots to people); the other two are risks of bots. Re-word.
3. Replace the Common Sense /ai-ratings/ link with the Institute page; add its own funding sentence next to the rating.
4. Art. 50: add the 2 Dec 2026 deadline for pre-existing generative systems.
5. Gemini check: now video and audio, Google content only, about ten checks a day; the 20 Nov 2025 blog link is dead.
6. "Don't trust your eyes": cite Diel's pooled 55.5%; add the detector line (61% false positives on non-native writers; don't use them on people).
7. Feeds and breaks: state the news-knowledge cost (−0.19 SD; −0.098 SD).
8. Law section: turn the directory links into an action (one named state bill; own words; one ask; a personal line; 20% vs 1%), add usa.gov beside Plural, say Plural is a company.
9. Learn page, fact 1: fine as is. Where the site anywhere says "meta-analyses show AI helps learning", remove it (Wang & Fan retracted; Bartoš).
10. OpenAI: keep the "confirm the label" caveat until read in-app.

F. Staleness

• 16–20 Nov 2026: CCW Review Conference (autonomous-weapons line) [F-earlier].
• 2 Dec 2026: EU marking deadline for pre-existing generative systems; EU nudification ban starts.
• 1 Jan 2027: NY RAISE Act, NY ch. 368, Colorado SB26-189, California large-platform provenance duties.
• Monthly: app settings paths; Gemini verifier limits; OpenAI pages (re-try).
• The research literature moves fastest on AI tutoring (two 2026 meta-analyses already; one retraction) and on phone bans (a 2026 J. Human Resources paper and a 2026 single-school cohort). Re-check both within six months.

---

G. Sources

All fetched 2026-11-05 unless marked [F-earlier] (also 2026-11-05, earlier session, per source/research/ai.md). Tag → who opened it (see Method).

Learning

• https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/ — Bastani et al., PNAS 2025 (read via Europe PMC full text https://www.ebi.ac.uk/europepmc/webservices/rest/PMC12232635/fullTextXML; pnas.org 403) — research [F]
• https://www.anthropic.com/research/AI-assistance-coding-skills — 29 Jan 2026 — research (company) [F]
• https://arxiv.org/abs/2607.08849 and https://arxiv.org/html/2607.08849v2 — Contractor & Reyes, v2 4 Oct 2026 — preprint [F]
• https://www.nature.com/articles/s41598-025-97652-6 (PDF read with pdftotext) — Kestin et al., Sci Rep 2025 — research [F]
• https://blogs.worldbank.org/en/education/From-chalkboards-to-chatbots-Transforming-learning-in-Nigeria — 9 Jan 2025 — primary (World Bank blog) [F]; working paper DOI 10.1596/1813-9450-11125 — OpenAlex record [record]; https://openknowledge.worldbank.org/handle/10986/43212 — 403 [unread]
• https://arxiv.org/abs/2410.03017 — Tutor CoPilot, v3 10 Oct 2026 — preprint [F]
• https://arxiv.org/abs/2506.08872 — Kosmyna et al., v2 31 Dec 2025 — preprint [F]
• https://arxiv.org/abs/2601.00856 — Stankovic et al. comment, 29 Dec 2025 — preprint [F]
• https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/ — Lee et al., CHI 2025 — research [F]
• https://www.nature.com/articles/s41599-025-04787-y — Wang & Fan 2025, RETRACTED 22 Apr 2026 — research (retracted) [F]; retraction note https://www.nature.com/articles/s41599-026-07310-z [F]
• https://api.osf.io/v2/preprints/8vs32/ — Bartoš, Martinková & Wagenmakers, PsyArXiv 12 May 2025 — preprint [record]
• DOI 10.1057/s41599-026-07019-z — Wu et al. 2026 meta-analysis — research, abstract-level [record]; DOI 10.1016/j.compedu.2024.105224 — Deng et al. 2024 — research, abstract-level [record] (both via https://api.openalex.org/works?search=...)
• International AI Safety Report 2026 extended summary — https://internationalaisafetyreport.org/publication/2026-report-extended-summary-policymakers — primary [F-earlier]; https://internationalaisafetyreport.org/about — [F-sub]

Truth

• https://api.semanticscholar.org/graph/v1/paper/search?query=Human+performance+in+detecting+deepfakes... — Diel et al. 2024, DOI 10.1016/j.chbr.2024.100538 — research, abstract-level [record]
• https://pmc.ncbi.nlm.nih.gov/articles/PMC8602050/ — Köbis et al., iScience 2021 — research [F]
• https://pmc.ncbi.nlm.nih.gov/articles/PMC8740705/ — Groh et al., PNAS 2022 — research [F]
• https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0285333 — Mai et al., PLOS ONE 2023 — research [F]
• https://arxiv.org/abs/2503.02857 — Deepfake-Eval-2024 — preprint [F-earlier]
• https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/ — Liang et al., Patterns 2023 — research [F]
• https://link.springer.com/article/10.1186/s41235-021-00291-4 — Brodsky et al. 2021 — research [F]
• DOI 10.1037/edu0000740 — Wineburg et al. 2022 — paywalled, OpenAlex metadata only [record]/[unread]
• https://www.nature.com/articles/s41562-024-01881-0 — Kozyreva et al. 2024 toolbox — research, abstract and intervention list [F]
• https://pmc.ncbi.nlm.nih.gov/articles/PMC9051116/ — Pennycook & Rand, Nat Commun 2022 — research [F]
• https://www.nature.com/articles/s41586-021-03344-2 — Pennycook et al., Nature 2021 — research (partial extract) [F]
• https://pmc.ncbi.nlm.nih.gov/articles/PMC7355018/ — Guess et al., PNAS 2020 — research [F]
• https://pmc.ncbi.nlm.nih.gov/articles/PMC9401631/ — Roozenbeek et al., Sci Adv 2022 — research [F]
• DOI 10.1037/xap0000315 — Maertens et al. 2020 — research, abstract via OpenAlex [record]
• DOI 10.1037/xge0001395 — Modirrousta-Galian & Higham 2023 — research, abstract via OpenAlex [record]
• https://www.cambridge.org/core/journals/american-political-science-review/article/liars-dividend-can-politicians-claim-misinformation-to-evade-accountability/687FEE54DBD7ED0C96D72B26606AA073 — Schiff, Schiff & Bueno 2024 — research [F]
• DOI 10.1177/2056305120903408 — Vaccari & Chadwick 2020 — research, abstract via OpenAlex [record] (sagepub and Loughborough pages 403)
• https://academic.oup.com/pnasnexus/article/3/10/pgae403/7795946 — Altay & Gilardi 2024 — research [F-earlier]
• https://c2pa.org/faqs/ — primary [F-earlier]; https://c2pa.org/about/, https://contentauthenticity.org/how-it-works, https://contentcredentials.org/verify — primary [F-sub]; Hacker Factor posts 16 Nov 2023 and 9 May 2024 — expert research blog [F-sub]; TechCrunch 1 Jul 2024 ("Made with AI" → "AI info") — news [F-sub]
• https://support.google.com/gemini/answer/16722517 — SynthID verification help — primary [F-sub]; https://blog.google/innovation-and-ai/products/verify-google-ai-videos-gemini-app/ (18 Dec 2025) — primary [F-sub]
• https://www.ic3.gov/PSA/2024/PSA241203 — primary [F-earlier]; https://www.ic3.gov/PSA/2025/PSA250515 — primary [F-sub]; https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf — primary [F-sub]; https://consumer.ftc.gov/consumer-alerts/2023/03/scammers-use-ai-enhance-their-family-emergency-schemes — primary [F-earlier]

Attention and feeds

• https://web.stanford.edu/~gentzkow/research/facebook.pdf — Allcott, Braghieri, Eichmeyer & Gentzkow, working paper 8 Nov 2019 (AER 2020), pages 1–4 read — research [F]; https://www.aeaweb.org/articles?id=10.1257/aer.20190658 — abstract [F]
• https://pmc.ncbi.nlm.nih.gov/articles/PMC11126999/ — Allcott et al., PNAS 2024 — research (Meta-funded, disclosed) [F]
• https://api.semanticscholar.org/graph/v1/paper/DOI:10.1126/science.abp9364 — Guess et al., Science 2023 — research, abstract-level [record] (science.org 403)
• https://arxiv.org/abs/2409.18393 — Bagchi et al. critique, 27 Sep 2024 — preprint [F]
• https://arxiv.org/abs/2411.14652 — Piccardi et al., Science 27 Nov 2025 — research [F]
• https://academic.oup.com/pnasnexus/article/4/3/pgaf062/8052060 — Milli et al. 2025 — research [F-earlier]
• https://academic.oup.com/pnasnexus/article/4/2/pgaf017/8016017 — Castelo et al. 2025 — research [F-earlier]
• https://www.nature.com/articles/d41586-024-00902-2 — Odgers, Nature 29 Mar 2024 — research review [F]
• https://www.afterbabel.com/p/social-media-mental-illness-epidemic — Haidt, 22 Feb 2023 — advocacy/research summary [F]

Kids and group formats

• https://about.commonsensemedia.org/press-room/press-releases/ai-companions-decoded-common-sense-media-recommends-ai-companion-safety-standards — 30 Apr 2025 — primary [F]
• https://institute.commonsensemedia.org/risk-assessments/methodology — primary [F] (and [F-sub])
• https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions — 16 Jul 2025 — charity's own figures [F]
• https://www.healthychildren.org/English/family-life/Media/Pages/are-ai-chatbots-safe-for-kids.aspx — AAP, updated 27 Aug 2025 — primary [F]
• https://www.healthychildren.org/English/fmp/Pages/MediaPlan.aspx — AAP Family Media Plan — primary, JS app, no evidence cited [F]
• https://www.aap.org/en/news-room/news-releases/aap/2026/american-academy-of-pediatrics-provides-guidance-on-safe-equitable-and-effective-use-of-generative-artificial-intelligence-for-clinicians/ — 3 Oct 2026 — primary [F-earlier] [F-sub]
• https://arxiv.org/abs/2503.17473 — Fang et al. 2025 — preprint [F-earlier]
• DOI 10.1037/dev0000108 — Collier et al. 2016 — research, abstract via OpenAlex [record]
• https://cep.lse.ac.uk/pubs/download/dp1350.pdf — Beland & Murphy, CEP DP1350, May 2015 (pages 1–5 read) — research [F]
• DOI 10.3368/jhr.0224-13403r2 — Abrahamsson, J. Human Resources 2026 — research, abstract via OpenAlex [record] (SSRN 403)
• DOI 10.3390/educsci14080906 — Böttger & Zierer 2024 — research, abstract via OpenAlex [record] (mdpi.com 403)
• https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11984610/fullTextXML — Goodyear et al., Lancet Reg Health Eur 2025 — research [F]
• https://www.frontiersin.org/journals/developmental-psychology/articles/10.3389/fdpys.2026.1796471/full — Vereecke et al. 2026 — research [F]
• https://www.unesco.org/gem-report/en/articles/phone-bans-schools-are-spreading-worldwide-policy-debate-rages — updated 18 Sep 2026 — primary [F]; https://www.unesco.org/gem-report/en/technology — [F]
• DOI 10.1080/19452829.2025.2518313 — Thiagarajan, Newson & Swaminathan 2025 — research, abstract via OpenAlex [record]
• https://www.waituntil8th.org/ — primary [F]; ProPublica EIN 83-0628912 and Charity Navigator — [F-sub]
• https://smartphonefreechildhood.org/ — primary [F]
• https://cdt.org/insights/report-hidden-harms-the-misleading-promise-of-monitoring-students-online/ — Aug 2022 — research (civil-society report) [F]
• https://api.semanticscholar.org/graph/v1/paper/search?query=Navigating+the+Jagged+Technological+Frontier... — Dell'Acqua et al. 2023/Org Sci 2026 — research, abstract-level [record] (SSRN 403)
• https://www.nber.org/papers/w31161 — Brynjolfsson, Li & Raymond — research [F]
• https://www.nber.org/papers/w33641 — Dell'Acqua et al., Cybernetic Teammate 2025 — research [F]

Policy

• https://api.openalex.org/works/https://doi.org/10.1177/1532673X08326967 — Bergan 2009 — research, abstract [record]
• https://link.springer.com/article/10.1007/s11109-014-9277-1 — Bergan & Cole 2015 — research [F]
• https://www.cambridge.org/core/journals/american-political-science-review/article/bias-in-perceptions-of-public-opinion-among-political-elites/2EF080E04D3AAE6AC1C894F52642E706 — Broockman & Skovron 2018 — research [F]
• https://api.openalex.org/works/https://doi.org/10.1111/ajps.12180 — Kalla & Broockman 2016 — research, abstract [record]
• https://api.openalex.org/works/https://doi.org/10.1093/jopart/mui042 — Yackee 2006 — research, abstract [record]
• https://www.congressfoundation.org/s/cwc-perceptions-of-citizen-advocacy.pdf — CMF 2011 (PDF, pdftotext) — survey of staff [F]
• https://www.congressfoundation.org/s/cmf-citizen-centric-advocacy.pdf — CMF 2017 (PDF, pdftotext) — survey of staff [F]
• https://ag.ny.gov/press-release/2021/attorney-general-james-issues-report-detailing-millions-fake-comments-revealing — 6 May 2021 — primary (government) [F]
• https://www.acus.gov/recommendation/managing-mass-computer-generated-and-falsely-attributed-comments — 403 [unread]; https://www.regulations.gov/learn — 403 [unread]
• https://www.house.gov/representatives/find-your-representative, https://www.senate.gov/senators/senators-contact.htm, https://pluralpolicy.com/find-your-legislator — primary [F-earlier]; https://pluralpolicy.com/about/, https://pluralpolicy.com/open, https://www.usa.gov/elected-officials — [F-sub]

Laws and orders

• https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/ — EO 14365, 11 Dec 2025 — primary [F]
• https://www.nysenate.gov/legislation/bills/2025/S9051/amendment/B — ch. 368 of 2026 — primary [F]; https://www.nysenate.gov/legislation/bills/2025/S8828 and /S6953/amendment/B — primary [F-earlier] [F-sub]; https://www.governor.ny.gov/news/governor-hochul-signs-nation-leading-legislation-protect-kids-ai-chatbots-and-ensure-baby — [F-sub]
• https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai and .../code-practice-ai-generated-content — primary [F-earlier] [F-sub]; Goodwin (3 Aug 2026) and Houthoff (1 Sep 2026) analyses of Reg. (EU) 2026/1744 — news (law firm) [F-sub]; EUR-Lex — blank [unread]
• https://calmatters.digitaldemocracy.org/bills/ca_202520260sb243 — SB 243 text mirror [F-earlier] [F-sub]; https://www.gov.ca.gov/2025/10/13/ — [F-sub]; leginfo — 403 [unread]
• https://leg.colorado.gov/bills/sb26-189 — primary [F-sub]; JD Supra (28 Jan 2026) on the DOJ task force — news [F-sub]
• https://www.insideprivacy.com/?p=15667 — Senate 99–1 — news [F-earlier]
• https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/ — 1 Oct 2025 — primary [F-earlier] [F-sub]
• OpenAI: help.openai.com/en/articles/7730893, /12315553, openai.com/index/introducing-parental-controls — all 403 today [unread]
• https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training and .../10023548 — primary [F-earlier]; https://support.google.com/gemini/answer/13594961 — primary [F-earlier]

Nonprofit filings (all [F-sub], charity's own figures / IRS data)

• https://projects.propublica.org/nonprofits/organizations/412024986 (Common Sense Media, FY2024); https://www.charitynavigator.org/ein/412024986
• https://projects.propublica.org/nonprofits/organizations/362275597 (AAP, FY2025); https://www.charitynavigator.org/ein/362275597
• https://projects.propublica.org/nonprofits/organizations/830628912 (Wait Until 8th Inc, 990-EZ 2025); https://www.charitynavigator.org/ein/830628912 (not rated)

Corrections, 2026-11-06

From the re-verification of 6 November (source/verify/ai-2026-11-06-additions.md): Wait Until 8th's site cites studies of phone harms, though none of the pledge itself (this file said it cites none). The Bartoš finding cited here is a May 2025 PsyArXiv re-analysis of Wang & Fan's data; a separate January 2026 preprint pools 67 meta-analyses and puts the bias-adjusted education effect at about a third.

Further, from the Do-page re-verification (source/verify/ai-do-2026-11-06.md): lateral reading took taught students from 5.9% to 61.0% (7% was all students before the lessons); the World Bank Nigeria study is an unpublished working paper (grade preprint; nine schools in Benin City; +0.31 SD overall, 0.23 in English); Diel et al. was read in the authors' manuscript on OSF (56 papers); accuracy prompts cut sharing intentions, not sharing.

No tracking on any link.Ticks stay on this device.What changedContested