Records · Verification table
AI Do page — source verification, 2026-11-06
The working record behind the pages, published as written. Source file source/verify/ai-do-2026-11-06.md. All records.
Page checked: content/ai-do.html. Scope: every claim in a line added or changed since ee6f7b8 (git diff ee6f7b8 HEAD -- content/ai-do.html, the 5 November rebuild from source/impact/ai-2026-11-05.md) that has no row in source/verify/ai-2026-11-05.md. Also re-checked, on the coordinator's request, five claims that do have 5 November rows but whose rows had flagged wording the rebuild never applied (rows B15, F4, F5, F9, G2 below). The help-now box was checked against source/impact/needs-2026-11-06.md and against the 988 pages themselves.
Method
- Read 2026-11-06, 09:35–09:52 EST. Nothing was taken from memory or from a search snippet.
- curl with a browser user agent, HTML converted to text with a stdlib parser. Quotes are copied from that text. PDFs were read with
pdftotext. - Headless Chrome (
[local working file],document.body.innerText) for pages that serve a JavaScript challenge to curl: link.springer.com (Brodsky; Bergan and Cole), eur-lex.europa.eu (Regulation 2026/1744). - Records: Europe PMC REST (Bastani, Guess, Castelo, Milli), OpenAlex and Crossref (Diel, Abrahamsson, Beland and Murphy, World Bank), the World Bank repository's own DSpace REST API (handle 10986/43212). A record is an index; where a record is all that was read, the row says so.
- Search (WebSearch) only to look for counter-evidence to three negative claims (rows D4, H6, I1). Nothing found by search was put on the page.
- Working files:
[local working file](not kept in the repo).
Hosts that refused today: nysenate.gov (403 to curl; Cloudflare "Performing security verification" in headless Chrome), academic.oup.com (403; read through PMC and Europe PMC instead), pubs.aeaweb.org PDF (403; read the authors' version), sciencedirect.com (refused headless Chrome), scholarworks.indianapolis.iu.edu (Anubis bot check, "Access Denied"). openknowledge.worldbank.org serves a JavaScript shell; its REST API and documents1.worldbank.org PDF answered.
Verdicts: CONFIRMED, CORRECTED (page changed to match the source), NOT IN SOURCE (the cited source doesn't say it; page reworded), NEWER DATA (a later version or status exists), UNREACHABLE (DNS/proxy/timeout), BLOCKED (403/405/script-only 404 or bot wall), UNFETCHABLE (the host says the page is gone). Evidence labels were checked against the grades fixed before reading in source/impact/ai-2026-11-05.md §0 (strong = two or more independent randomised or field experiments, or a meta-analysis; moderate = one well-powered RCT or quasi-experiment, or several consistent smaller studies; weak = small, preprint, self-report or correlational; none but plausible = no study of the action itself).
---
A. Help-now box (line 12) and its source s23
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| A1 | "call, text or chat 988, any time" | 988lifeline.org/get-help/ and home page, curl | "When you call, text, or chat the 988 Lifeline, your conversation is confidential."; "24/7 free, confidential mental health hotline"; home: "Call, text, or chat with a 988 Lifeline counselor for help during difficult moments anytime, day or night." Matches needs file A2.1. | CONFIRMED | — |
| A2 | "for yourself or for someone else" | 988lifeline.org/help-someone-else/, curl | "Contact 988 at any time for help if a friend is struggling." Matches A2.3. | CONFIRMED | — |
| A3 | "you don't have to be suicidal to call" | get-help, curl | "People do not have to be suicidal to call." Matches A2.2. | CONFIRMED | — |
| A4 | "In Spanish: press 2, or text AYUDA." | get-help, curl | "To call a Spanish-speaking counselor, dial 988 and then press 2. To text with a Spanish-speaking counselor, text AYUDA to 988." Matches A2.4. | CONFIRMED | — |
| A5 | "Phone interpreters cover more than 240 languages." | 988lifeline.org/faq/calling-the-988-lifeline/faq-is-the-988-lifeline-available-in-other-languages-for-non-english-speakers/, curl | "For people who speak other languages, call 988 and you will be provided translation in more than 240 additional languages through Language Line Solutions. This is available 24/7 through voice calling only." Matches A2.5. | CONFIRMED | — |
| A6 | s23 cites "interpretation services" (988lifeline.org/interpretation-services/) for the box | that page, curl | The page lists language pages only: "988 Mga Serbisyo ng Lifeline para sa mga nagsasalita ng Tagalog", "Services d'aide à la vie 988 pour les francophones", Russian, Chinese, Cantonese. No figure. The 240 quote is on the FAQ page (A5), as the needs file records. | CORRECTED | s23 third link → the FAQ URL, text "Is the 988 Lifeline available in other languages?" |
| A7 | "If a conversation … has turned to suicide or self-harm" | help-someone-else, curl | "Never keep it a secret if a friend tells you about a plan to hurt themselves." | CONFIRMED (framing) | — |
B. The three steps (lines 19, 22, 25)
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| B1 | Step 1 label "moderate" | impact review B1 | "Grade moderate." | CONFIRMED | — |
| B2 | "one of 194" | Kestin et al., nature.com/articles/s41598-025-97652-6, curl | "(N = 194)"; "Of the 233 enrolled students, 194 were eligible for inclusion in the study." | CONFIRMED | — |
| B3 | "one across Nigerian schools" | World Bank WP 11125 PDF, documents1.worldbank.org, pdftotext | "First-year secondary students from nine public schools in Benin City were invited to participate"; "a six-week after-school tutoring program" | CONFIRMED | made exact: "one across nine Nigerian schools" |
| B4 | "a 211-person follow-up" | Contractor and Reyes, arxiv.org/pdf/2607.08849v2 | "Our randomized experiment with 211 undergraduates"; abstract: "complete unaided assessments immediately and one week later". It is its own trial with a one-week retest, not a follow-up of another study. | CORRECTED | "one of 211 retested a week later" |
| B5 | "all point the same way: a tutor that makes you work helps" | Bastani (5 Nov L4); Kestin; World Bank record; Contractor and Reyes abstract | Bastani: "Student performance in the GPT Tutor arm was statistically indistinguishable from that of the control arm" (it did no harm; it didn't help on the exam). Kestin: "students learn significantly more in less time when using the AI tutor". World Bank: "a significant improvement of 0.31 standard deviation". C&R: "Gains remain for students who use AI as a tutor". | CORRECTED | "Used as a tutor, AI helped or did no harm." |
| B6 | "a tool that hands you the output costs you once it's gone" | Bastani (Europe PMC record); Contractor and Reyes PDF | Bastani: "17% reduction in grades for GPT Base". C&R: "automation users' advantages disappear … their test-score difference is negative but imprecise (β̂ = −0.24 SD, p = 0.538)". | CORRECTED | "Used to do the work, it cost students once it was gone, or left them no lasting gain." |
| B7 | Step 2 label "moderate" | impact review B2 | "Grade moderate-to-strong." The page grades one step lower. | CONFIRMED | — |
| B8 | "Twenty experiments with 26,863 people" | Pennycook and Rand, pmc.ncbi.nlm.nih.gov/articles/PMC9051116/, curl | "meta-analyzing 20 experiments (with a total N = 26,863)" | CONFIRMED | — |
| B9 | "a pause for accuracy cuts sharing of false headlines by about a tenth" | same | "primarily by reducing sharing intentions for false headlines by 10% relative to control". The outcome is stated intent in online surveys, not sharing. | CORRECTED | "cuts people's stated intent to share false headlines by about a tenth" |
| B10 | "six lessons in 'lateral reading'" | Brodsky et al., link.springer.com/article/10.1186/s41235-021-00291-4, headless Chrome | "instructors spent three class sessions in Weeks 4 and 5"; "In Weeks 7 and 8, students in DPI sections spent three class sessions practicing" | CONFIRMED | "six class sessions" |
| B11 | "took students who could judge a source from 7% to 61%" | same | "7.0% of students received a score of 4 (i.e., read laterally and correctly assessed trustworthiness) for at least one problem at pretest (5.9% of students in the DPI sections and 8.5% in the control sections)"; "(n = 83 or 61.0% of students in DPI sections)". 7% is all students before; 61% is taught students after; taught students started at 5.9%. The measure is one correct lateral read out of four problems. | CORRECTED | "61% of students checked a source elsewhere and judged it right on at least one of four problems, up from 6% before" |
| B12 | Step 3 label "moderate: with a cost" | impact review B3 | "Grade moderate, with known costs." | CONFIRMED | — |
| B13 | "Four weeks off Facebook raised wellbeing a little and cut news knowledge a little" | AER page aeaweb.org (abstract), curl; authors' PDF web.stanford.edu/~gentzkow/research/facebook.pdf (8 Nov 2019); AER PDF 403 | AER: "deactivating Facebook for the four weeks before the 2018 US midterm election … (ii) reduced both factual news knowledge and political polarization; (iii) increased subjective well-being". Authors' PDF: "index of news knowledge fell by 0.19 standard deviations"; "SWB index by 0.09 standard deviations". | CONFIRMED | s6 notes the figures are from the authors' 2019 version |
| B14 | "two weeks without mobile internet improved attention and mood" | Castelo et al., PMC11834938, curl | "The intervention improved mental health, subjective well-being, and objectively measured ability to sustain attention"; "participants in the Intervention condition had higher mood on average (M = 6.62 …) than participants in the Delayed Intervention condition (M = 6.02 …) … d = 0.33" | CONFIRMED | — |
| B15 | "a time-ordered feed … changed nobody's politics" (5 Nov row L16 flagged "anyone's" → "average") | Guess et al., Europe PMC record | "the chronological feed did not significantly alter levels of issue polarization, affective polarization, political knowledge, or other key attitudes during the 3-month study period" (an average null) | CORRECTED | "and, on average, left political attitudes where they were" |
C1, "Near you" (lines 29–31): the site's own behaviour (CLAUDE.md, Local steps); n/a.
D. The habits (lines 35–40)
| # | Words on the page | Source and read | Source's words | Verdict | Change | ||
|---|---|---|---|---|---|---|---|
| D1 | "Ask for questions" label "moderate" | impact review A1 table | "Ask for hints/questions, not answers, when learning \ | moderate" | CONFIRMED | — | |
| D2 | "Write first" label "weak: one 18-person session" | impact review A1; Kosmyna (5 Nov L8) | "Write your own rough pass first \ | weak"; "18 completing session 4" | CONFIRMED | — | |
| D3 | "Make it argue against you" label "weak: self-report only" | impact review A1; s2 Lee et al. | Review: "Make it argue against you \ | none-but-plausible \ | Only correlational self-report (Lee et al.). Untested as an intervention." The rubric's "weak" also covers "self-report or correlational". | CONFIRMED (grade is a judgment; see proposals) | — |
| D4 | "nobody has tested whether asking for objections fixes that" | s2 (Lee et al., a survey: "We survey 319 knowledge workers", 5 Nov L7); search 6 Nov | The cited survey makes no claim about other tests. Search found a 24-person study of AI "provocations" in knowledge work (Drosos et al., arXiv 2501.17247; seen in search results, not read). | NOT IN SOURCE | "the survey didn't test whether asking for objections fixes that" | ||
| D5 | "Check every fact" label "moderate" | impact review A1 | "Check every fact you'll repeat \ | moderate (as lateral reading, A2)" | CONFIRMED | — | |
| D6 | "if it isn't there, the citation was invented" | s5 (Pennycook; Brodsky) | Neither says anything about citations; a missing sentence can also be a misattribution or an edited page. (5 Nov D6 flagged the earlier "Models invent citations" the same way.) | NOT IN SOURCE | "if it isn't there, don't repeat the claim" | ||
| D7 | "Keep one skill" label "weak: mechanism only"; "The harm in the trials appeared when the crutch was removed" | Bastani (Europe PMC record); review A1 | "Without guardrails, students attempt to use GPT-4 as a 'crutch' during practice problem sessions, and subsequently perform worse on their own." Review: "No trial of the habit." | CONFIRMED | — | ||
| D8 | Detector label "strong: evidence of harm" | review §0 rubric; s13 | One study is cited; the review's "strong" needs two or more independent experiments or a meta-analysis. The review gives this line no grade. | CORRECTED | "moderate: evidence of harm" | ||
| D9 | "Detectors flagged 61% of essays by non-native English writers as AI-written." | Liang et al., PMC10382961, curl | "we evaluated the performance of seven widely used GPT detectors on 91 TOEFL … essays"; "they incorrectly labeled more than half of the TOEFL essays as 'AI-generated' (average false-positive rate: 61.3%)"; "we strongly caution against the use of GPT detectors in evaluative or educational settings" | CONFIRMED | made exact: "Seven detectors flagged, on average, 61% …" |
E. Settings (lines 44–49)
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| E1 | Settings label "none but plausible: privacy only" | impact review A6 | "Grade: none-but-plausible for any learning/attention/truth/safety outcome; this is a privacy act" | CONFIRMED | — |
| E2 | s8: Meta, "Improving your recommendations on our apps with AI at Meta", 1 October 2025 (in effect 16 December 2025) | about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/, curl | og:title "Improving Your Recommendations on Our Apps With AI at Meta"; published "2025-10-01"; "several weeks before it goes into effect on December 16, 2025" | CONFIRMED | — |
F. Telling what's real (lines 55–60)
| # | Words on the page | Source and read | Source's words | Verdict | Change | |
|---|---|---|---|---|---|---|
| F1 | "Don't trust your eyes" label "strong" | impact review A2 | "Don't trust your eyes or a free detector \ | strong" | CONFIRMED | — |
| F2 | "Across 56 studies" | Diel et al.: OpenAlex and Crossref records; authors' manuscript on OSF (osf.io/cxv4r, file "manuscript - 2.docx", read in full) | Title: "Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers"; "56 papers involving 86,155 participants" | CORRECTED | "Across 56 papers" | |
| F3 | "people spotted deepfakes about 55% of the time, close to a coin toss" | same | Manuscript: "they only correctly detected deepfake stimuli 55.54% [48.87, 62.10] of the time"; journal abstract: "Overall deepfake detection rates (sensitivity) were not significantly above chance because 95% confidence intervals crossed 50%." | CONFIRMED | — | |
| F4 | "People take AI-written text for human 77% of the time" (5 Nov L10, D19) | internationalaisafetyreport.org extended summary, curl | "In one study, participants misidentified AI-generated text as human-written 77% of the time." | CORRECTED | "In one study, people took AI-written text for human 77% of the time." | |
| F5 | "open-source detectors lose about half their accuracy on real fakes" (5 Nov L11 flagged AUC) | arxiv.org/abs/2503.02857 (v5, 27 May 2026), curl | "the performance of open-source state-of-the-art deepfake detection models drops precipitously when evaluated on Deepfake-Eval-2024, with AUC decreasing by 50% for video, 48% for audio, and 45% for image models compared to previous benchmarks"; the benchmark is "in-the-wild deepfakes collected from social media and deepfake detection platform users in 2024" | CORRECTED | "Open-source detectors lose about half their score on the standard measure (AUC) when tested on fakes collected online." | |
| F6 | "Find the original" label "moderate"; "Read what other sources say about the site before you read the site." | Brodsky et al., headless Chrome | "Lateral reading requires users to leave the information (website) to find out whether someone has already fact-checked the claim, identify the original source, or learn more about the individuals or organizations making the claim"; taught "using reverse image search to find the original version of an image". Review A2: "moderate". | CONFIRMED | — | |
| F7 | "Don't swing" label "strong" | impact review A2 | "strong for the phenomenon; instruction untested" | CONFIRMED | — | |
| F8 | Content Credentials label "weak: rarely present" | impact review A2; c2pa.org/faqs/ | Review: "Content Credentials check \ | none-but-plausible"; no study of the action is cited. No source read says how often credentials are present. | CORRECTED | "none but plausible: can be stripped" |
| F9 | "they're often stripped when a file is re-uploaded" (5 Nov L31b's own suggested wording) | c2pa.org/faqs/, curl | "While C2PA Manifests are typically embedded in the asset, they can be separated."; "providing resilience in cases where metadata is accidentally or intentionally stripped"; "While metadata can be removed inadvertently or intentionally". Nothing on frequency or re-uploading. | NOT IN SOURCE | "Credentials can be stripped from a file, by accident or on purpose" | |
| F10 | Gemini label "weak: Google's output only" | impact review A2 | "Gemini 'was this created with Google AI?' \ | none-but-plausible, narrow" | CORRECTED | "none but plausible: Google's output only" |
| F11 | Safe word label "none but plausible: official advice" | impact review A2 | "Family safe word \ | none-but-plausible" | CONFIRMED | — |
G. Feeds (line 64)
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| G1 | "a 35,000-person study of leaving Facebook and Instagram before the 2020 election found polarisation effects near zero" | Allcott et al. 2024, PMC11126999, curl | "randomizing a subset of 19,857 Facebook users and 15,585 Instagram users to deactivate their accounts for 6 wk before the 2020 U.S. election"; "the effects of both Facebook and Instagram deactivation on affective and issue polarization … were all precisely estimated and close to zero." 19,857 + 15,585 = 35,442 (our arithmetic). | CONFIRMED | s6 gives both numbers instead of the sum |
| G2 | "didn't change anyone's politics" (5 Nov D22, L16) | Guess et al., Europe PMC record | as B15 | CORRECTED | "didn't change political attitudes on average" |
H. If you have a kid (lines 75–82)
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| H1 | "Its research institute says it is funded by philanthropy and by industry, including makers of some of the products it rates." | institute.commonsensemedia.org/risk-assessments/methodology, curl | "The Youth AI Safety Institute is funded by both philanthropy and industry, including the makers of some of the technologies we evaluate." | CONFIRMED | s15 "read 5 Nov" → "read 6 Nov" |
| H2 | Kids label "weak: expert judgment" | impact review A4 | "weak-to-moderate evidence of harm; the rule is expert judgment" | CONFIRMED | — |
| H3 | "activates once ten families in a grade sign" | waituntil8th.org, curl | "as long as at least 10 families total from your child's grade and school pledge. Once 10 families have pledged to delay the smartphone, you will be notified that the pledge is active!" | CONFIRMED | made exact: "in a grade at a school" |
| H4 | "so nobody's kid is the only one without" | same | "Banding together helps decrease the pressure to have a phone at an early age." | CONFIRMED (paraphrase of the mechanism) | — |
| H5 | "Its site cites no studies" | waituntil8th.org/why-wait, curl | "a new UC San Francisco-led study published in the Journal of Child Psychology and Psychiatry found"; "The early results of a landmark study on brain development by the National Institute of Health"; the page links cep.lse.ac.uk/pubs/download/dp1350.pdf, PMC7954067 and tandfonline 10.1080/19452829.2025.2518313. None tests the pledge. | CORRECTED | "Its site cites studies of phone harms, but none of the pledge itself, and we found none." |
| H6 | "and we found none" | search 6 Nov ("Wait Until 8th pledge evaluation study"); impact review B4 | No evaluation of the pledge found (search-bounded). Review: "Pledges … have no evaluation and cite no studies on their sites." | CONFIRMED (search-bounded) | — |
| H7 | Label "none but plausible: no studies" | as H6 | no study of the pledge itself | CONFIRMED | — |
| H8 | Phone-free label "moderate: contested on wellbeing" | Goodyear et al., PMC11984610; Abrahamsson, JHR page | Goodyear: pupils at restrictive schools "did not differ in their sleep duration and efficiency, physical activity, academic attainment, and disruptive classroom behaviour". Abrahamsson: "has no average effect on education or mental health". Attainment is contested too. | CORRECTED | "moderate: contested" |
| H9 | "Bans that were actually enforced raised test scores in English schools, most for the weakest students" | Beland and Murphy, CEP DP1350 PDF | "student test scores improve by 6.41% of a standard deviation. Our results indicate that there are no significant gains in student performance if a ban is not widely complied with."; "Students in the lowest quartile of prior achievement gain 14.23% of a standard deviation" | CONFIRMED | reworded with H10–H12: "In England, test scores rose after schools banned phones, most for the weakest students, but only where the ban was kept." |
| H10 | "and cut bullying and girls' mental-health referrals in Norway" | Abrahamsson, jhr.uwpress.org (via doi.org), curl; NHH news, nhh.no, 12 Aug 2024 | JHR: "banning smartphones in middle schools has no average effect on education or mental health but masks important gender differences. Bans reduce psychological health care use among girls, lower bullying for both genders, and improve girls' GPAs". The abstract says nothing about enforcement and doesn't name the country. NHH: "den første i Norge som undersøker om et smarttelefonforbud i ungdomsskolen har effekter" (the first in Norway to examine whether a lower-secondary smartphone ban has effects). | CORRECTED | "In Norway, bans had no average effect on grades or mental health, but they cut bullying, and girls used less mental-health care and got better grades." |
| H11 | "Two studies of school wellbeing found no difference." | Goodyear; Vereecke et al., frontiersin.org, curl | Goodyear: "no evidence of a difference between groups (adjusted mean difference −0.48, 95% CI −2.05 to 1.06, p = 0.62)". Vereecke: "a total smartphone ban was not associated with clear improvements in adolescents' social wellbeing"; "social loneliness increased modestly". | CORRECTED | "Two other studies found no clear gain in wellbeing; one of them found no difference in attainment either." |
| H12 | "A ban on paper did nothing." | Beland and Murphy | "we find larger effects in schools that report a high level of compliance with a ban compared to schools where compliance is weak, where there is no statistically significant impact." | CORRECTED | folded into H9's "but only where the ban was kept" |
I. Inner work (lines 86–90)
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| I1 | "no study shows that a builder's practice changes what they build" | s19 (Goyal; Castelo; Holt-Lunstad); search 6 Nov | None of the three addresses what anyone builds. Search found mindfulness experiments with software-engineering students reporting better conceptual-modelling efficiency (Bernárdez et al., ESEM 2014, IEEE TSE; search results only, not read). | NOT IN SOURCE | "it doesn't test whether a builder's practice changes what they build" |
| I2 | Mindfulness label "moderate" | Goyal (5 Nov L37) | "moderate evidence" | CONFIRMED | — |
| I3 | Castelo label "weak: one trial" | Castelo, PMC11834938 | "a preregistered randomized controlled trial"; "Participants (n = 467)". The rubric puts one well-powered RCT at moderate and self-report outcomes at weak; SWB and mental health were self-reported. | CONFIRMED (grade is a judgment; see proposals) | — |
| I4 | Social ties label "moderate: association, not a trial" | Holt-Lunstad (5 Nov L37) | meta-analysis of observational studies, "OR = 1.50" | CONFIRMED (judgment) | — |
J. Ask for one thing (lines 94–97)
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| J1 | Label "moderate" | impact review B5 | "Grade moderate." | CONFIRMED | — |
| J2 | "In a field experiment, a few constituent messages moved state legislators' votes by about 12 points" | Bergan and Cole, link.springer.com/article/10.1007/s11109-014-9277-1, headless Chrome | "Michigan state legislators are randomly assigned to be contacted by their constituents about a specific bill or to a control group"; "being targeted by constituent contacts increases the probability of supporting the relevant legislation by about 12 percentage points"; the contacts were patched-through phone calls, "(0, 22, 33 or 65 intended calls)" | CORRECTED | "state legislators who got a few dozen calls from constituents about a bill were about 12 percentage points more likely to support it" |
| J3 | "congressional staff say an individual message has 'a lot of influence' about twenty times as often as a form letter" | CMF 2011 PDF (congressfoundation.org link, served from squarespace), pdftotext | "About 20% of staffers said individualized email and postal mail would have a lot of influence on an undecided Member and only one percent said identical form email and postal mail would."; survey "in the field between October 12 and December 13, 2010"; "260 respondents". "Twenty times" was our arithmetic on rounded figures and undated. | CORRECTED | "In a 2010 survey, about 20% of congressional staff said an individual letter or email would have 'a lot of influence' on an undecided member; 1% said so of a form message." |
| J4 | "Find your state legislators at usa.gov" | usa.gov/elected-officials, curl | "Find the names and current activities of state and territorial legislators."; "LAST UPDATED: March 2, 2026" | CONFIRMED | — |
| J5 | "Plural, a company's free lookup tool" | pluralpolicy.com/find-your-legislator, curl | "Plural is able to provide the Find Your Legislator tool at no cost"; "© 2026 Plural Policy, Inc." | CONFIRMED | — |
| J6 | NY "law on AI companions for minors … take[s] effect on 1 January 2027" | IAPP, curl | "The law enters into force in 1 Jan. 2027" (also 5 Nov D37) | CONFIRMED | — |
| J7 | "as does Colorado's replacement AI law" | leg.colorado.gov/bills/sb26-189, curl; Arnold & Porter, curl | Colorado: "SB26-189 Automated Decision-Making Technology"; "The act repeals and reenacts those provisions"; "starting January 1, 2027"; "Became Law"; "05/14/2026 Signed Act". A&P: "the law taking effect on January 1, 2027". | CONFIRMED | — |
| J8 | "with 2 December 2026 as the deadline for generative systems already on the market" (cited to s21, the Commission page) | digital-strategy.ec.europa.eu AI Act page (updated 19 Oct 2026), curl; EUR-Lex, Regulation (EU) 2026/1744, headless Chrome | Commission page has no such date ("The transparency rules of the AI Act will come into effect in August 2026."). The regulation: "'4. Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026.'" (amending Article 111; regulation of 8 July 2026, OJ 24.7.2026) | CORRECTED (claim true; citation added) | s21 adds, as text with no link: "the 2 December 2026 deadline is in Regulation (EU) 2026/1744, which adds Article 111(4) to the AI Act, read on EUR-Lex 6 Nov 2026" (primary) |
The new prompt (line 115) and the description line make no factual claim; n/a.
K. Source list (lines 121–143)
| # | Words on the page | Source and read | Source's words | Verdict | Change |
|---|---|---|---|---|---|
| K1 | s1 Bastani et al., "Generative AI without guardrails can harm learning", PNAS 122(26), 2025 | Europe PMC REST record | "Generative AI without guardrails can harm learning: Evidence from high school mathematics."; PNAS 122 26, 2025-06-25, PMC12232635 | CONFIRMED | — |
| K2 | s1 "Contractor and Reyes, preprint, October 2026 (211 people; the gain from tutoring-style use held a week later)" | arXiv abstract page and v2 PDF | "Authors: Zara Contractor, Germán Reyes"; "last revised 4 Oct 2026 (this version, v2)"; "211 undergraduates"; "Gains remain for students who use AI as a tutor". In the text the one-week tutor-use estimate is "β̂ = 0.24 SD, p = 0.095" and the use-type analysis is called "descriptive". | CONFIRMED | — |
| K3 | s2 "AI assistance and coding skills" | anthropic.com/research/AI-assistance-coding-skills, curl | og:title "How AI assistance impacts the formation of coding skills" | CORRECTED | link text → the real title |
| K4 | s4 Kestin et al., "AI tutoring outperforms in-class active learning", Scientific Reports, 2025 (194; 0.63 to 1.3 SD) | nature.com, curl | "AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting"; Scientific Reports 15, 17458, 2025-06-03; "While the linear regression suggests an effect size of 0.63, this is an underestimation due to ceiling effect; … an effect size in the range of 0.73 to 1.3 standard deviations." | CONFIRMED (0.63 is the estimate the authors call an underestimate) | — |
| K5 | s4 World Bank, "From chalkboards to chatbots", 2025 (randomized trial in Nigerian schools with teachers present; +0.31 SD) | repository REST API (openknowledge.worldbank.org/server/api/pid/find?id=hdl:10986/43212); WP PDF | "From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria"; "Policy Research Working Paper; 11125"; "2025-05-20"; "Using a randomized controlled trial"; "a significant improvement of 0.31 standard deviation on an assessment that included English topics aligned with the Nigerian curriculum, knowledge of artificial intelligence and digital skills. The effect on English, the main outcome of interest, was of 0.23 standard deviations."; "Teachers, who played a critical role in guiding the students but did not provide direct instruction" | CONFIRMED | made exact: "nine schools in Benin City, after school, with teachers guiding; +0.31 standard deviations overall, 0.23 in English" |
| K6 | s4 World Bank grade "research" | same; OpenAlex search 6 Nov | a working paper; OpenAlex lists no journal version | CORRECTED | grade → "preprint"; "working paper" added |
| K7 | s5 "Accuracy prompts are a replicable and generalizable approach" | PMC9051116 | citation_title "Accuracy prompts are a replicable and generalizable approach for reducing the spread of misinformation"; "2022 Apr 28" | CORRECTED | full title |
| K8 | s5 "Improving college students' fact-checking strategies through lateral reading", Cognitive Research, 2021 (230, quasi-experiment) | Springer, headless Chrome | "Improving college students' fact-checking strategies through lateral reading instruction in a general education civics course"; "Published: 31 March 2021"; "N = 230"; "Instructor-matched sections" | CORRECTED (title); rest CONFIRMED | full title |
| K9 | s6 Allcott et al., "The welfare effects of social media", AER, 2020 (+0.09, −0.19 SD) | as B13 | "volume 110, issue 3, pages 629-76 of American Economic Review, March 2020"; figures from the authors' 8 Nov 2019 version (published PDF 403) | CONFIRMED | "in the authors' 2019 version" added |
| K10 | s6 Allcott et al., "The effects of Facebook and Instagram on the 2020 election", PNAS, 2024 (about 35,000) | PMC11126999 | "The effects of Facebook and Instagram on the 2020 election: A deactivation experiment"; "2024 May 13;121(21)"; numbers as G1 | CONFIRMED | "(19,857 Facebook and 15,585 Instagram users; …)" |
| K11 | s6 Castelo et al., "Blocking mobile internet on smartphones", PNAS Nexus, February 2025 | Europe PMC record | "Blocking mobile internet on smartphones improves sustained attention, mental health, and subjective well-being."; 2025-02-18 | CORRECTED | full title |
| K12 | s6 Guess et al., "How do social media feed algorithms affect attitudes?", Science, 2023 | Europe PMC record | "How do social media feed algorithms affect attitudes and behavior in an election campaign?"; Science 381(6656), 2023-07-27 | CORRECTED | full title |
| K13 | s6 Milli et al., "Engagement, user satisfaction and divisive content", PNAS Nexus, March 2025 | Europe PMC record | "Engagement, user satisfaction, and the amplification of divisive content on social media."; 2025-03-05 | CORRECTED | full title |
| K14 | s11 "Diel et al., meta-analysis of 56 studies …, 2024 (pooled accuracy about 55%; record read through an index, not the paper) research, unread" | OpenAlex/Crossref records; OSF manuscript read in full; journal page BLOCKED (ScienceDirect, IU repository) | as F2–F3; "Computers in Human Behavior Reports", 2024-11-30 | NEWER DATA (now read in the authors' manuscript) | title and venue added; "(deepfakes spotted 55.54% of the time, not significantly above chance; read in the authors' preprint and the journal's abstract)"; grade "research" |
| K15 | s11 IASR "(77%)" | as F4 | "In one study" | CORRECTED | "(77%, from one study it cites)" |
| K16 | s11 Deepfake-Eval "(detectors' accuracy about halves on real-world fakes)" | as F5 | AUC −50%/−48%/−45% | CORRECTED | "(open-source detectors' AUC fell by 45% to 50% on real-world fakes)" |
| K17 | s13 Liang et al., "GPT detectors are biased against non-native English writers", Patterns, 2023 (61% of TOEFL essays flagged) | PMC10382961 | same title; "2023 Jul 10;4(7):100779"; 61.3% | CONFIRMED | — |
| K18 | s15 CSM institute page "read 5 Nov 2026" | as H1 | page live, sentence unchanged | CONFIRMED | "read 6 Nov 2026" |
| K19 | s17 Beland and Murphy, "Ill communication …", CEP, 2015 (four English cities; +6.4%; +14.2% lowest quartile) | CEP DP1350 PDF | "Ill Communication: Technology, Distraction & Student Performance"; "CEP Discussion Paper No 1350 May 2015"; "surveying schools in four English cities"; 6.41%; 14.23% | CONFIRMED | — |
| K20 | s17 Beland and Murphy (version) | Crossref | "Ill Communication: Technology, distraction & student performance", Labour Economics 41, 61–76, 2016-08, DOI 10.1016/j.labeco.2016.04.004 (journal text not read) | NEWER DATA | none (proposal) |
| K21 | s17 Abrahamsson, JHR, 2026, doi:10.3368/jhr.0224-13403r2 (Norway; abstract read) | jhr.uwpress.org; NHH | "Smartphone Bans, Student Outcomes and Mental Health"; "Published online before print March 11, 2026"; Norway as H10 | CONFIRMED | — |
| K22 | s17 Goodyear et al., 2025 (30 schools, cross-sectional, no difference) | PMC11984610 | "School phone policies and their association with mental wellbeing, phone use, and social media use (SMART Schools): a cross-sectional observational study"; "2025 Feb 4"; "30 English secondary schools"; no difference in wellbeing or attainment (H8, H11) | CONFIRMED | "(30 English schools, cross-sectional; no difference in wellbeing or attainment)" |
| K23 | s17 Vereecke et al., Frontiers in Developmental Psychology, 2026 (no improvement) | frontiersin.org | "More than disconnecting? School smartphone bans and adolescents' social wellbeing"; 2026-09-22; "not associated with clear improvements" | CORRECTED | "(no clear improvement in social wellbeing)" |
| K24 | s17 UNESCO, 2026 (58% of countries by March 2026) | unesco.org, curl | "114 education systems now have a national ban on mobile phones in schools, representing 58% of countries worldwide"; "by March 2026"; "17 March 2026", "Last update: 18 September 2026" | CONFIRMED | — |
| K25 | s17 UNESCO link "phone bans are spreading" at …/phone-bans-schools-are-spreading-worldwide-policy-debate-rages | curl -I | "HTTP/1.1 301 Moved Permanently", "location: /gem-report/en/articles/how-many-countries-have-phone-bans-school"; title "How many countries have phone bans in school?" | CORRECTED | URL → the 301 target; link text → its title |
| K26 | s18 "more than 155,000 parents" | waituntil8th.org | "More than 155,000 parents have said yes to waiting on the smartphone for their families." | CONFIRMED | "read 6 Nov 2026" |
| K27 | s18 "(no studies cited)" | as H5 | as H5 | CORRECTED | "its 'Why wait' page cites studies of phone harms, none of the pledge" |
| K28 | s18 "its 2025 filing at ProPublica" | projects.propublica.org/nonprofits/organizations/830628912 | "Wait Until 8th Inc"; "EIN: 83-0628912"; "Fiscal Year Ending Dec. 2025"; "Extracted Financial Data from 990-EZ" | CONFIRMED | — |
| K29 | s20 Bergan and Cole, "Call your legislator", Political Behavior, 2015 (field experiment on state legislators) | Springer, headless Chrome | "Call Your Legislator: A Field Experimental Study of the Impact of a Constituency Mobilization Campaign on Legislative Voting"; "Polit Behav 37, 27–42 (2015)"; Michigan; phone calls | CONFIRMED | "(a field experiment on Michigan state legislators; constituent phone calls)" |
| K30 | s20 CMF, "Perceptions of citizen advocacy" (undated) | CMF 2011 PDF | cover: "Communicating with Congress" over "Perceptions of Citizen Advocacy on Capitol Hill" (line breaks joined; the running header is "Perceptions of Citizen Advocacy on Capitol Hill"); "© 2011, Congressional Management Foundation" | CORRECTED | title, "2011", survey dates and size |
| K31 | s20 Broockman and Skovron, APSR, 2018 | cambridge.org, curl | "Bias in Perceptions of Public Opinion among Political Elites"; APSR 112(3), "2018/08" | CONFIRMED | — |
| K32 | s21 EO 14365, "Ensuring a National Policy Framework for Artificial Intelligence", 11 December 2025 | whitehouse.gov, curl | "ENSURING A NATIONAL POLICY FRAMEWORK FOR ARTIFICIAL INTELLIGENCE"; "Executive Order 14365"; "December 11, 2025" | CONFIRMED | — |
| K33 | s21 IAPP, "New York signs youth chatbot bill", October 2026 | iapp.org, curl | og:title "New York's governor signs youth AI chatbot safety bill"; "Published28 Oct. 2026" | CORRECTED | link text → the real title; "28 October 2026" |
| K34 | s21 IAPP "(chapter 368 of 2026)" | iapp.org | IAPP names "Senate Bill9051B/State Assembly Bill10379C"; no chapter number anywhere in the article. | NOT IN SOURCE | "(S9051B; in force 1 January 2027)" |
| K35 | NY chapter 368 on nysenate.gov/legislation/bills/2025/S9051/amendment/B (read 5 Nov per D37; not linked on this page) | curl 403; headless Chrome: Cloudflare "Performing security verification" | — | BLOCKED | — |
| K36 | s21 Arnold & Porter, "Colorado narrows its AI law", July 2026 (SB 26-189, effective 1 January 2027) | arnoldporter.com, curl | "Colorado Narrows Its AI Law, but the FTC Says That's Not Enough: What Should Companies Make of These Changes?"; "July 23, 2026"; "SB 26-189"; "the law taking effect on January 1, 2027" | CONFIRMED | — |
| K37 | s21 Multistate, 2026 (twelve states by mid-year) | multistate.ai, curl | "Jun 26" (datePublished 2026-06-26); "Weekly Update, Vol. 105."; "Twelve states have enacted companion chatbot laws through the first half of 2026: Colorado, Georgia, Iowa, New York, Rhode Island, California, Idaho, Nebraska, Oregon, Connecticut, and Washington." That list names eleven. | CONFIRMED (as stated; the source's list disagrees with its count) | "26 June 2026 (twelve states by mid-year, though its own list names eleven)" |
Not re-checked (5 November rows exist and nothing in them was flagged)
Lede 17% (D1), Bastani "about a thousand" (L2), the Claude, ChatGPT and Gemini settings (D8–D13), Meta AI wording (D14, D15), the 99–1 vote (L27, D36), RAISE Act dates and penalties (L25), EU 2 August 2026 (L23), Instagram/Facebook/X/YouTube/TikTok paths (D23–D27), AAP lines (D30–D32), CSM rating (D28), Q&A 1–6 (D39–D43), s3, s7, s9, s10, s12, s14, s16, s19, s22. The 5 November row D43a (Council of Europe treaty) and D43b (CCW dates) stay UNFETCHABLE; not retried.
Summary by verdict
One verdict per row; n/a rows (C1, the new prompt, the description) excluded.
| Verdict | Count | Rows |
|---|---|---|
| CONFIRMED | 63 | A1–A5, A7, B1–B3, B7, B8, B10, B12–B14, D1–D3, D5, D7, D9, E1, E2, F1, F3, F6, F7, F11, G1, H1–H4, H6, H7, H9, I2–I4, J1, J4–J7, K1, K2, K4, K5, K9, K10, K17–K19, K21, K22, K24, K26, K28, K29, K31, K32, K36, K37 |
| CORRECTED | 36 | A6, B4–B6, B9, B11, B15, D8, F2, F4, F5, F8, F10, G2, H5, H8, H10–H12, J2, J3, J8, K3, K6–K8, K11–K13, K15, K16, K23, K25, K27, K30, K33 |
| NOT IN SOURCE | 5 | D4, D6, F9, I1, K34 |
| NEWER DATA | 2 | K14, K20 |
| BLOCKED | 1 | K35 |
| UNREACHABLE | 0 | — |
| UNFETCHABLE | 0 | — |
Evidence labels changed: detectors strong → moderate (D8); Content Credentials weak → none but plausible (F8); Gemini weak → none but plausible (F10); phone-free school "contested on wellbeing" → "contested" (H8). Source grade changed: World Bank working paper research → preprint (K6); Diel research, unread → research (K14).
What would overturn these
- B5/B6: a trial where hint-style use lowers unaided scores, or answer-style use raises them a week on.
- H8–H12: a controlled roll-out of enforced bans that finds no attainment or bullying effect, or that finds a wellbeing gain.
- J3: a CMF survey after 2010 that splits "a lot" from "some" influence differently (the 2017 report's chart, from a 2015 survey, gives only combined bars as text: individualized email 92%, form email 56%).
- K37: Multistate correcting its list or its count.
Staleness
Readings are as of 2026-11-06 09:35–09:52 EST. Fast-moving: the 988 language FAQ (the needs file's monthly check), the Wait Until 8th count, Multistate's tally, UNESCO's ban count (updated twice in 2026), and the EU deadline, which passes on 2 December 2026 (the page already greys it with data-until).
Contest a fact, suggest something, or tell us what’s broken
If someone is in danger right now, don’t use this form. Call your local emergency number (911 in the US). For trafficking: 1-888-373-7888, or text 233733.
No name, no account. Your note goes into the site’s private records, nothing about your browser or connection is stored with it, and the email box is the only way we’d know who you are. The weekly review reads every note, re-reads the original source, and fixes the page if it should. A contested fact and its outcome are listed on Contested, in the review’s words, never yours.
No tracking on any link.Ticks stay on this device.What changedContested