Records · Verification table
AI pages — source verification, 2026-11-05
The working record behind the pages, published as written. Source file source/verify/ai-2026-11-05.md. All records.
Pages checked: content/ai.html, content/ai-learn.html, content/ai-do.html, content/ai-practice.html, and the "ai" entry of CORRECTIONS in tools/essays.py. Research notes: source/research/ai.md (its section 6 flagged the Bastani sample/effect wording and the Kosmyna numbers as never read; those were done first).
Method
- Every checkable claim was numbered and traced to its source: the URL on the page, else the URL in the research file, else the primary document (paper on arXiv/PMC/publisher, the law's text on nysenate.gov or the Commission site, the vendor's own help page).
- Each source was fetched fresh on 2026-11-05 with WebFetch. Quotes below are the source's words as returned by the fetch; where the fetch tool capped quote length, quotes are given in the pieces it returned. PDFs that WebFetch saved were read locally with
pdftotext. PNAS and Science both refuse automated reads, so their records were read through the Europe PMC REST API (the PNAS paper is open access in PMC as PMC12232635 and its full text was read; for Science only the abstract record is available). Vaccari & Chadwick 2020 was read through the Semantic Scholar API record of its abstract because both the publisher and the Loughborough repository returned 403. - Nothing was substituted from memory. Where a source could not be read, the verdict is UNFETCHABLE, full stop.
- The session's web-search budget ran out part-way (shared limit); after that only direct fetches were possible, so some vendor help pages whose URLs could not be found by search are marked UNFETCHABLE rather than guessed.
- Hosts that refused or did not render today: pnas.org (403), science.org (403), journals.sagepub.com (403), repository.lboro.ac.uk (403), cetas.turing.ac.uk HTML (403; the PDF downloaded), committees.parliament.uk (403), commerce.senate.gov (403), senate.gov/legislative LIS roll-call pages (403; senate.gov contact page loaded), help.openai.com / openai.com / privacy.openai.com (403), help.x.com and help.twitter.com (403), coe.int (403 ×3) and conventions.coe.int (https→http redirect loop), consilium.europa.eu (403), eur-lex.europa.eu (blank render ×3), meetings.unoda.org Review-Conference URLs (404 ×3), reachingcriticalwill.org / dig.watch / epic.org (403), facebook.com/privacy/genai (sections render empty), help.instagram.com and tiktok.com/support (JS shells, title only).
Verdicts: CONFIRMED (the source's words support the page), CORRECTED (fix wording given), NOT IN SOURCE (the cited source doesn't say it; what it does say is given), UNFETCHABLE (could not read the source today), NEWER DATA (a later version, edition or status exists). Grades on the site's scale: primary · research (peer-reviewed; preprints labelled) · charity's own figures · company's own figures · news/secondary · industry.
---
1. Overview — content/ai.html
| # | Words on the page | Source | Fetch | Source's words | Verdict | Grade |
|---|---|---|---|---|---|---|
| O1 | "Students who let a chatbot hand them answers scored 17% worse once it was taken away; students given hints lost nothing." | Bastani et al., PNAS 122(26), 2025 — pnas.org/doi/10.1073/pnas.2422633122 (403); read via Europe PMC full text PMC12232635 | OK (full text) | "students in the GPT Base arm perform statistically significantly worse than students in the control arm by 17%"; "Student performance in the GPT Tutor arm was statistically indistinguishable from that of the control arm"; "the prompt instructs GPT-4 to provide hints to the student without directly giving them the answer, to encourage learning" | CONFIRMED | research (peer-reviewed) |
| O2 | "A cloned voice can sound exactly like your kid." | FBI IC3 PSA I-120324-PSA, 3 Dec 2024 — ic3.gov/PSA/2024/PSA241203 | OK | "Criminals generate short audio clips containing a loved one's voice to impersonate a close relative in a crisis situation, asking for immediate financial assistance or demanding a ransom." | CONFIRMED (paraphrase; "exactly" is the page's gloss) | primary |
| O3 | "30 minutes · $0", "Daily · 2 minutes", "Thirty seconds", "Two minutes" | site's own estimates | n/a | — | n/a (site's own) | — |
| O4 | Essay title "Towards a more loving, hopeful and faithful humanity" | source/essays (Isaac's text) | not fetched (internal) | — | n/a (internal) | — |
2. Learn — content/ai-learn.html
| # | Words on the page | Source | Fetch | Source's words | Verdict | Grade | |
|---|---|---|---|---|---|---|---|
| L1 | "Nine facts … six words" | the page itself | count | 9 <li> in ol.facts; 6 <dt> in dl.words | CONFIRMED (arithmetic) | — | |
| L2 | Fact 1: "About a thousand high-school students in Turkey practised maths with a chatbot." | Bastani et al. (as O1) | OK | "In collaboration with a high school in Turkey, we conducted a large-scale randomized controlled trial"; "We conducted four 90-min sessions for about fifty 9th, 10th, and 11th-grade classes, comprising nearly 1,000 students." | CONFIRMED (note: one high school; the section-6 "from memory" flag is now cleared) | research | |
| L3 | Fact 1: "Those with unrestricted access scored 17% worse on the exam, once the bot was removed, than students with no bot." | Bastani et al. | OK | "students in the GPT Base arm perform statistically significantly worse than students in the control arm by 17%"; practice: "students in the GPT Tutor (resp., GPT Base) arm perform 127% (resp., 48%) better … on the practice problems compared to students in the control arm" | CONFIRMED | research | |
| L4 | Fact 1: "Those whose bot gave only hints scored the same as the no-bot group." | Bastani et al. | OK | "Student performance in the GPT Tutor arm was statistically indistinguishable from that of the control arm"; "provide hints to the student without directly giving them the answer" | CONFIRMED | research | |
| L5 | Fact 1: "The students with the answers thought they'd learned more." | Bastani et al. | OK | "While students in the GPT Base arm performed worse on the exam (relative to the control arm), they did not perceive that they performed worse or learned less." | CORRECTED → "The students with the answers didn't notice they'd learned less." (The paper says they did not perceive doing worse or learning less; it does not say they believed they learned more.) | research | |
| L6 | Fact 2: "Engineers who delegated a task to AI learned less of it than those who worked it through (52 people)." | Anthropic, 29 Jan 2026 — anthropic.com/research/AI-assistance-coding-skills | OK | "Published Jan 29, 2026"; "We recruited 52 (mostly junior) software engineers"; "a randomized controlled trial"; "the AI group averaged 50% on the quiz, compared to 67% in the hand-coding group" (Cohen's d=0.738, p=0.01); low-scoring pattern "AI delegation": "Participants in this group wholly relied on AI to write code and complete the task." | CONFIRMED (precision: the randomised comparison was AI-assisted vs hand-coding; "delegated" names the lowest-scoring usage pattern within the AI group) | company's own research, not peer-reviewed (page says "research" — see L41) | |
| L7 | Fact 2: "Office workers who trusted AI more reported doing less critical thinking of their own (319 people, self-reported)." | Lee et al., CHI 2025 — microsoft.com/en-us/research/publication/… and the PDF | OK (page + PDF) | "We survey 319 knowledge workers"; "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking." | CONFIRMED | research (peer-reviewed) | |
| L8 | Fact 3: "'Your Brain on ChatGPT' is a preprint with 54 participants and 18 in its key session" | Kosmyna et al., arXiv 2506.08872 | OK | "A total of 54 participants took part in Sessions 1-3, with 18 completing session 4."; v1 10 Jun 2025, v2 31 Dec 2025; no journal reference | CONFIRMED (section-6 flag cleared) | preprint | |
| L9 | Fact 3: "a published comment disputes its methods" | Stankovic et al., arXiv 2601.00856 | OK | "[Submitted on 29 Dec 2025]", only v1; Comments: "Comment on arXiv:2506.08872"; Journal reference: none; criticisms: "(i) study design considerations, including the limited sample size; (ii) the reproducibility of the analyses"; "(iii) methodological issues related to the EEG analysis; (iv) inconsistencies in the reporting of results" | CORRECTED → "a comment posted on arXiv (itself unreviewed) disputes its methods". It is not published in a journal. Same fix in Do page s3. | preprint | |
| L10 | Fact 4: "They judged AI text to be human 77% of the time." | International AI Safety Report 2026, extended summary | OK (web page) | "In one study, participants misidentified AI-generated text as human-written 77% of the time." | CONFIRMED (one study, as reported by the IASR) | expert report (not peer-reviewed; page labels it "research") | |
| L11 | Fact 4: "Free deepfake detectors lose about half their accuracy on fakes found in the wild." | Deepfake-Eval-2024, arXiv 2503.02857 | OK | "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." | CONFIRMED, wording: "open-source" (not "free") and the measure is AUC, not accuracy → suggest "Open-source deepfake detectors lose about half their discriminating power on fakes found in the wild." | preprint (v5 27 May 2026; no journal reference — see L43) | |
| L12 | Fact 5: "In experiments they caused more uncertainty than deception, and that uncertainty lowered trust in news generally." | Vaccari & Chadwick, Social Media + Society 2020 — Sage 403, Loughborough 403; abstract read via api.semanticscholar.org record | OK (abstract via index) | "We find that people are more likely to feel uncertain than to be misled by deepfakes, but this resulting uncertainty, in turn, reduces trust in news on social media." | CONFIRMED, wording: "lowered trust in news on social media" (not "generally") | research, abstract read via index | |
| L13 | Fact 5: "Labelling true headlines 'AI-generated' made people believe them less." | Altay & Gilardi, PNAS Nexus 3(10), 1 Oct 2024 | OK | "two preregistered online experiments among US and UK participants (N = 4,976)"; "labeling headlines as AI-generated lowered their perceived accuracy and participants' willingness to share them" … "regardless of whether the headlines were true or false, and created by humans or AI."; "three times smaller than labeling them as false." | CONFIRMED | research | |
| L14 | Fact 5: "In the 2024 UK, French and EU elections, researchers found viral AI disinformation but no evidence it changed a result." | CETaS (Alan Turing Institute), Stockwell, "AI-Enabled Influence Operations: Threat Analysis of the 2024 UK and European Elections", Sep 2024 — HTML 403; PDF read | OK (PDF) | "We identified just 16 confirmed viral cases of AI-enabled disinformation or deepfakes during the UK general election, while only 11 viral cases were identified in the EU and French elections combined."; "there is no evidence that AI-enabled disinformation or deepfakes meaningfully impacted UK or European election results." | CONFIRMED (now read; s5's "via summary" label can go) | research-institute briefing (not peer-reviewed) | |
| L15 | Fact 6: "On Twitter, engagement ranking amplified out-group hostility and anger more than any other emotion, and users rated its political picks lower than a time-ordered feed." | Milli et al., PNAS Nexus 4(3), 5 Mar 2025 | OK | n=806 Twitter users, "preregistered algorithmic audit"; "anger was by far the predominant emotion amplified by the engagement-based algorithm"; partisan tweets and out-group animosity each amplified by 0.24 SD (P<0.001); engagement-selected political tweets "led to significantly lower user value than the political tweets in the reverse-chronological timeline" (−0.18 SD, P=0.005) | CONFIRMED | research | |
| L16 | Fact 6: "In Meta's three-month experiment, time-ordered feeds cut time on the apps and didn't move anyone's polarisation." | Guess et al., Science 381(6656), 27 Jul 2023 — science.org 403; abstract via Europe PMC | OK (abstract) | "Moving users out of algorithmic feeds substantially decreased the time they spent on the platforms and their activity."; "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." | CONFIRMED (wording: "anyone's" → "average"; it is a null average effect) | research | |
| L17 | Fact 7: "On one music service, fully AI-generated tracks were over half of daily uploads on peak days in 2026 and one to three percent of streams." | Deezer newsroom, 21 Jul 2026 | OK | "Paris, July 21, 2026"; "accounting for more than 50% of total daily deliveries on peak days"; "fully AI-generated music currently accounts for only a small fraction of streams on Deezer, 'between 1-3%'"; "up to 85% of the streams generated by fully AI-generated tracks were in fact fraudulent in 2025" | CONFIRMED | company's own figures | |
| L18 | Fact 7: "Of the first 500 Shorts shown to a new YouTube account, 104 were AI-generated." | Kapwing, "AI Slop Report", 28 Nov 2025 | OK | "we created a new YouTube account and recorded the number of AI slop and brainrot videos among the first 500 YouTube Shorts"; "in total, 104 (21%) of the first 500 videos were AI-generated"; "Data is correct as of October 2025." | CONFIRMED | industry (company blog) | |
| L19 | Fact 8: "AI companions are rated unacceptable for anyone under 18. Common Sense Media found age gates easy to pass and bots that gave harmful advice and claimed to have feelings." | Common Sense Media press release, 30 Apr 2025 | OK | "pose unacceptable risks to children and teens under age 18 and should not be used by minors."; "Age gates and terms of service-based restrictions on use by teens were easily circumvented"; "suggestions that users harm themselves or others."; "AI companions routinely claimed to be real, and to possess emotions, consciousness, and sentience." | CONFIRMED | charity's own (primary) | |
| L20 | Fact 8: "Pediatricians' warning signs: pulling away from friends, turning to the bot when upset." | AAP / HealthyChildren.org, updated 8/27/2025 | OK | No list headed "warning signs" exists. Nearest: "Talk with your child's doctor if they seem withdrawn and prefer talking with chatbots rather than people."; "Kids may be drawn to chatbots when they feel frustrated with the people in their lives."; "This can create an unhealthy cycle of hanging out alone with a chatbot while growing more distant from friends, classmates, mentors and family." | CORRECTED → "The pediatricians' advice: talk to your child's doctor if they seem withdrawn and prefer the bot to people." (Turning to the bot when frustrated is described as the start of an unhealthy cycle, not listed as a sign.) | primary | |
| L21 | Fact 9: "There is no federal AI law." | EO 14365, 11 Dec 2025 (whitehouse.gov) | OK | §1: "My Administration must act with the Congress to ensure that there is a minimally burdensome national standard"; "Until such a national standard exists" | CONFIRMED (indirect: a negative claim; the EO presupposes no national standard) | primary | |
| L22 | Fact 9: "The states are writing them, and Washington is suing to stop some." | EO 14365; DOJ page (justice.gov/ai-litigation-task-force → 404) | EO OK; DOJ action not fetched | EO §3: "Within 30 days of the date of this order, the Attorney General shall establish an AI Litigation Task Force" "whose sole responsibility shall be to challenge State AI laws". No fetched source confirms a suit or intervention actually filed (the research file's DOJ-joins-xAI-v-Colorado item was a search summary). | UNFETCHABLE (the suing itself). Safe wording: "a federal task force exists to sue over some of them" | primary (EO) | |
| L23 | Fact 9: "The EU's labelling rules took effect on 2 August 2026" | European Commission AI Act page (updated 19 Oct 2026); AI Act Explorer Art. 50 | OK | Commission: "The AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with some exceptions:"; "The transparency rules of the AI Act will come into effect in August 2026." Explorer: "Comes into force 2 August 2026, according to Article 113." | CONFIRMED | primary + secondary rendering of the law | |
| L24 | Fact 9: "its strictest rules slipped to 2027 and 2028" | Commission AI Act page | OK | "Starting on 2 December 2027, high-risk AI systems will be subject to strict obligations"; "For systems integrated into products such as lifts or toys, the rules will apply from 2 August 2028."; "have been extended to 2 December 2027 as a result of the political agreement on the proposal to simplify the AI Act"; AI Omnibus "entered into force on 27 July 2026" | CONFIRMED | primary | |
| L25 | Fact 9: "New York's law on the largest developers takes effect on 1 January 2027, with penalties cut tenfold on the way." | nysenate.gov S8828 (chapter amendment) and S6953B (original RAISE Act) | OK (both) | S8828: "signed chap.96" (Mar 27, 2026); "This act shall take effect [on the ninetieth day after it shall have become a law] JANUARY 1, 2027."; "NOT TO EXCEED ONE MILLION DOLLARS FOR A FIRST VIOLATION" / "NOT TO EXCEED THREE MILLION DOLLARS PER SUBSEQUENT VIOLATION". S6953B: "Dec 19, 2025 … signed chap.699"; "NOT EXCEEDING TEN MILLION DOLLARS FOR A FIRST VIOLATION" / "NOT EXCEEDING THIRTY MILLION DOLLARS FOR ANY SUBSEQUENT VIOLATION." "Large developer": "A PERSON THAT HAS TRAINED AT LEAST ONE FRONTIER MODEL AND HAS SPENT OVER ONE HUNDRED MILLION DOLLARS IN COMPUTE COSTS" | CONFIRMED; arithmetic $10M→$1M and $30M→$3M = tenfold ✓ | primary | |
| L26 | Fact 9: "A December 2025 executive order set up a task force to challenge state AI laws" | EO 14365 | OK | "Executive Order 14365"; "December 11, 2025"; §3 quote as L22 | CONFIRMED | primary | |
| L27 | Fact 9: "the Senate had voted 99 to 1 against banning them" | Inside Privacy (Covington), 7 Jul 2025 — insideprivacy.com/?p=15667; primary roll call (senate.gov LIS) and Senate Commerce release both 403 | secondary OK; primary 403 | "The Senate voted 99–1 to strike the moratorium language from the bill" during a "marathon 27-hour 'vote-a-rama' on July 1."; "The amendment passed 99–1, with only Senator Thom Tillis (R-NC) voting 'no.'"; House version: "imposing a ten-year moratorium on states" | CONFIRMED against the cited secondary; primary record UNFETCHABLE today | secondary (law-firm blog), as the page labels it | |
| L28 | Words: "Training vs retention … 72 hours to several years, depending on the service and what you clicked" | Google Gemini Apps Privacy Hub (updated 26 Oct 2026); Anthropic privacy centre | OK | Google: "Future chats are still saved for 72 hours."; human-reviewed chats "are retained for up to three years." Anthropic: "we may retain your data in a de-identified format for up to 5 years in our model training pipelines."; feedback: "we retain data associated with that submission for 5 years." | CONFIRMED | primary | |
| L29 | Words: "Deepfake vs cheapfake … Most viral political fakery is still the cheap kind." | no source on the page or in the research file; nearest fetched: CETaS Sep 2024 | — | CETaS: "Generative AI played less of a role in boosting the virality of disinformation compared to traditional interference methods and human influencers. Malicious actors continued to rely on well-established disinformation tactics, such as flooding comments sections on social media with bot accounts (astroturfing), and exploiting politically-aligned social media influencers". The word "cheapfake" and the claim "most … is still the cheap kind" appear in no fetched source. | NOT IN SOURCE → either drop the last sentence or replace with: "In the 2024 elections, viral disinformation leaned on old methods (bot comments, influencers) more than on generative AI." citing CETaS | — | |
| L30 | Words: "Engagement-ranked vs chronological … The first is what every major feed defaults to" | Instagram blog 31 May 2023; Meta newsroom 21 Jul 2022; X/TikTok help pages 403/JS | Meta sources OK | Instagram: Following is an option — "Tap 'Following' from the Feed dropdown menu to see content from accounts you follow in chronological order."; Facebook: "Home is the new name of the tab you first see when you open Facebook to discover new content." | CONFIRMED for Meta's apps; "every" not verifiable for X and TikTok today (vendor pages blocked) | primary | |
| L31a | Words: "Content Credentials … It proves the file hasn't changed, not that it's true." | C2PA FAQ; Content Authenticity Initiative "How it works" | OK (both) | C2PA: "Any tampering with the asset or its metadata invalidates the cryptographic hash and signature". CAI: Content Credentials "aren't intended to prescriptively indicate whether a piece of content is 'real.'" | CONFIRMED (page cites only C2PA; add the CAI page for the "not that it's true" half) | primary | |
| L31b | Words: "Most sites strip it on upload, so its absence proves nothing." | C2PA FAQ | OK | "While C2PA Manifests are typically embedded in the asset, they can be separated."; "While metadata can be removed inadvertently or intentionally, it is vital information for consumers,"; "providing resilience in cases where metadata is accidentally or intentionally stripped." Nothing about "most sites". | NOT IN SOURCE ("most sites") → "Credentials can be separated from the file and are often stripped when it's re-uploaded, so their absence proves nothing." | primary | |
| L32 | Words: "Frontier model … New York's and California's new laws apply to these, not to the apps built on them." | S6953B; gov.ca.gov SB 53 release, 29 Sep 2025 | OK | NY: frontier model = "AN ARTIFICIAL INTELLIGENCE MODEL TRAINED USING GREATER THAN 10^26 COMPUTATIONAL OPERATIONS … THE COMPUTE COST OF WHICH EXCEEDS ONE HUNDRED MILLION DOLLARS". CA: "Requires large frontier developers to publicly publish a framework on its website" | CONFIRMED | primary | |
| L33 | Contest 1: "one 2026 preprint found a lasting gain from tutoring-style use" | Contractor & Reyes, arXiv 2607.08849 (v1 9 Jul 2026, v2 4 Oct 2026) | OK | "AI access raises test scores, and a smaller gain persists a week later."; "Gains remain for students who use AI as a tutor ('augmentation')" and "fade for students who have AI write for them ('automation')." | CONFIRMED (wording: "lasting" = one week later; sample size not in the abstract) | preprint | |
| L34 | Contest 1: "The dramatic claims rest on tiny samples." | editorial, resting on L8 (54/18) | — | — | n/a (supported by L8) | — | |
| L35 | Contest 2: "preregistered audits"; "critics note the platform changed its own systems during the study" | Milli (as L15); Bagchi et al., arXiv 2409.18393 = Science eLetter 26 Sep 2024 | OK | Milli: "preregistered algorithmic audit". Bagchi: "its reporting and conclusions did not account for a series of temporary emergency changes to Facebook's news feed algorithm" "in the wake of the 2020 U.S. presidential election"; "these emergency measures systematically reduced the amount of misinformation in the control group of the study" | CONFIRMED (note: the critique concerns misinformation exposure, not the polarisation null) | research / eLetter | |
| L36 | Contest 3: "EU high-risk rules pushed back, Colorado's law replaced before it started, New York's penalties cut … the labelling rules started on time, a dozen states passed chatbot-safety laws in 2026, and the attempt to ban state laws lost 99 to 1." | Commission page (L24); Arnold & Porter 23 Jul 2026; S8828/S6953B (L25); Multistate 26 Jun 2026; Transparency Coalition 15 Jul 2026; Inside Privacy (L27) | OK (all) | Arnold & Porter: "replacing its original AI law, SB 24-205, with a lighter disclosure-focused framework, SB 26-189" "weeks before that original AI act was set to take effect on June 30, 2026"; new law "taking effect on January 1, 2027". Multistate: "Twelve states have enacted companion chatbot laws through the first half of 2026". Transparency Coalition: "14 chatbot safety measures passed and/or enacted in 13 states". | CONFIRMED (counts disagree 12 vs 13 states; keep "a dozen"). The Learn page cites no source for Colorado or the state count (s9 lists only the Commission, S8828, the EO and Inside Privacy): add Arnold & Porter and Multistate/Transparency Coalition to s9. | secondary trackers | |
| L37 | Contest 4: "modest for mindfulness across 47 trials, real in one two-week trial of blocking mobile internet, and strong for social ties" | Goyal 2014; Castelo 2025; Holt-Lunstad 2010 | OK (all) | Goyal: 47 trials, 3,515 participants; "moderate evidence" anxiety ES 0.38 (8 weeks), depression 0.30, pain 0.33; "We found no evidence that meditation programs were better than any active treatment". Castelo: "The intervention improved mental health, subjective well-being, and objectively measured ability to sustain attention." Holt-Lunstad: "OR = 1.50 (95% CI 1.42 to 1.59)" | CONFIRMED | research | |
| L38 | "optimising for engagement produced measurably angrier feeds" | Milli (L15) | OK | as L15 | CONFIRMED | research | |
| L39 | Read the originals: IASR 2026 "The consensus document, 100+ experts." | IASR extended summary | OK | "3 February 2026 — Summary"; "Over 100 independent experts contributed to this report." | CONFIRMED | report | |
| L40 | s1: "Bastani et al., PNAS 2025 (about 1,000 students, read through an index)" | Europe PMC record + full text PMC12232635 | OK | Record: PNAS 122(26), first published 2025-06-25, PMID 40560616, PMCID PMC12232635, open access; full text read | NEWER DATA (status): the paper has now been read in full — drop "read through an index" (also in Do page s1) | research | |
| L41 | s2: "Anthropic … 29 January 2026 (52 engineers) research" | Anthropic page | OK | "Published Jan 29, 2026"; "52 (mostly junior) software engineers"; a company blog post of an internal RCT, no journal | CORRECTED (grade): label "company's own research, not peer-reviewed" rather than "research" | company's own | |
| L42 | s3: "Kosmyna et al., preprint, 2025, not peer-reviewed" | arXiv | OK | v1 10 Jun 2025; v2 31 Dec 2025; no journal reference | CONFIRMED | preprint | |
| L43 | s4: IASR "3 February 2026 research"; "Deepfake-Eval-2024, 2025–26 research"; Altay & Gilardi "October 2024 (4,976 people) research" | as above | OK | IASR date ✓; Deepfake-Eval versions v1 4 Mar 2025 … v5 27 May 2026, no journal reference; Altay "Published 01 October 2024", N = 4,976 | CORRECTED (grade): Deepfake-Eval-2024 is a preprint — label it "preprint" (also Do page s7, Practice s2); IASR is a report, not peer-reviewed research | — | |
| L44 | s5: "Vaccari and Chadwick, Social Media + Society, 2020 — research, via summary"; "Schiff, Schiff and Bueno, APSR, 2024 (15,000+ adults) — via summary"; "CETaS, September 2024 — via summary" | Semantic Scholar record; Cambridge Core page; CETaS PDF | OK | Vaccari: venue "Social Media + Society", 2020. Schiff: "We administer five survey experiments to over 15,000 American adults"; APSR Volume 119, Issue 1, pp. 71–90; "Published online 20 February 2024"; issue dated February 2025. CETaS: "September 2024". | NEWER DATA: Schiff et al. is now in print as APSR 119(1), 2025 (online 2024) — cite "APSR 2025 (online 2024)"; all three abstracts/texts have now been read, so "via summary" can become "abstract read" / "read" | research | |
| L45 | s6: Milli "March 2025 (806 users, preregistered)"; Guess "Science 2023"; Piccardi "Science, November 2025 (1,256 users)" | OUP; Europe PMC; arXiv 2411.14652 | OK | Milli "Published: 05 March 2025", n=806. Guess: Science 381(6656), 2023-07-27. Piccardi: "a preregistered 10-day field experiment with 1,256 participants on X"; Journal reference "Science; Volume 390 | Issue 6776; 27 November 2025" | CONFIRMED | research |
| L46 | s7: "Deezer, 2026 statements"; "Kapwing, YouTube Shorts audit, October–November 2025 (one account, manual coding)" | Deezer; Kapwing | OK | Kapwing "Nov 28, 2025"; "Data is correct as of October 2025."; "we created a new YouTube account and recorded the number…" | CONFIRMED (note: Kapwing says "recorded"; "manual coding" is the page's description, defensible) | company / industry | |
| L47 | s8: "AAP, Are AI chatbots safe for kids?, 27 August 2025" | healthychildren.org | OK | Title on page: "How AI Chatbots Affect Kids: Benefits, Risks & What Parents Need to Know"; "By: Joanna Parga-Belinkie, MD, FAAP"; "Last updated: 8/27/2025" | CORRECTED (title): the article is titled "How AI Chatbots Affect Kids: Benefits, Risks & What Parents Need to Know" (the URL slug is "are-ai-chatbots-safe-for-kids"). Same in Do s12 and Practice s3. | primary | |
| L48 | s9: Commission page "updated 19 October 2026"; S8828 "chapter 96 of 2026"; EO 14365 "11 December 2025"; Inside Privacy "July 2025"; "the earlier RAISE penalties ($10M/$30M) were seen only in a search summary — unread" | as above; S6953B | OK | Commission "19 October 2026" ✓; "signed chap.96" ✓; EO date ✓; Inside Privacy "July 7, 2025" ✓; S6953B text: "NOT EXCEEDING TEN MILLION DOLLARS … NOT EXCEEDING THIRTY MILLION DOLLARS" | NEWER DATA (status): the $10M/$30M figures are now read in the signed bill text (S6953B, chap.699, 19 Dec 2025) — replace the "unread" clause with a link to nysenate.gov/legislation/bills/2025/S6953/amendment/B | primary | |
| L49 | s10: Goyal "2014 (47 trials)"; Castelo "February 2025"; Holt-Lunstad "2010 (148 studies)" | as L37 | OK | Goyal "Published Online: January 6, 2014", 47 trials; Castelo "Published: 18 February 2025"; Holt-Lunstad "July 27, 2010", 148 studies | CONFIRMED | research |
3. Do — content/ai-do.html
| # | Words on the page | Source | Fetch | Source's words | Verdict | Grade | |
|---|---|---|---|---|---|---|---|
| D1 | Lede: 17% / hints lost nothing | Bastani (O1) | OK | as O1 | CONFIRMED | research | |
| D2 | Step 1: "People who asked for explanations kept their skill; people who handed over the work lost it." | Bastani; Anthropic | OK | Anthropic: stronger mastery came from "asking follow-up questions, requesting explanations, or posing conceptual questions"; "Participants in this group wholly relied on AI to write code and complete the task." Bastani as L4 | CONFIRMED | research / company | |
| D3 | Habits: "In one small study, people who drafted on their own before using a chatbot remembered more of their own work than people who started with it … that session had 18 people." | Kosmyna arXiv | OK | "Brain-to-LLM users exhibited higher memory recall and activation of occipito-parietal and prefrontal areas"; "LLM users also struggled to accurately quote their own work."; "18 completing session 4" | CONFIRMED | preprint | |
| D4 | Habits: "The hint-only tutor in the classroom trial did no harm." | Bastani | OK | as L4 | CONFIRMED | research | |
| D5 | Habits: "Office workers who trusted the AI more did less critical thinking of their own." | Lee et al. | OK | as L7 (self-reported) | CONFIRMED | research | |
| D6 | Habits: "Models invent citations." | none cited | — | — | NOT IN SOURCE (no source cited; general statement) — low priority; could cite the IASR or leave as common knowledge | — | |
| D7 | Settings intro: "It doesn't remove anything from models already trained, and companies still keep chats for a while." | Anthropic news 28 Aug 2025; privacy centre | OK | "Your data will still be included in model training that has already started"; retention quotes as L28 | CONFIRMED | primary | |
| D8 | Claude: "Settings → Privacy → Help improve Claude → off." | privacy.claude.com 10023580 (updated March 16, 2026); anthropic.com/news/updates-to-our-consumer-terms | OK (both) | Help centre: setting changed in "Privacy Settings" (links claude.ai/settings/data-privacy-controls); calls it "Model Improvement"; gives no menu path. News post: "You can change your choice in your Privacy Settings at any time."; "the model training setting". Neither page uses the label "Help improve Claude" or the path "Settings → Privacy". | NOT IN SOURCE (path and label) → "Claude: Settings → Privacy (claude.ai/settings/data-privacy-controls) → the model-training toggle → off (Anthropic's help pages call it 'Model Improvement')." | primary | |
| D9 | Claude: "Incognito chats are never used for training." | 10023580 | OK | "Your Incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your Privacy Settings." | CONFIRMED | primary | |
| D10 | Claude: "Thumbs-up or thumbs-down feedback keeps that whole chat for five years." | 10023580; 10023548 (dated July 1, 2026) | OK | "When you provide us feedback via our thumbs up/down button, we will store the entire related conversation" … "for up to 5 years"; "we retain data associated with that submission for 5 years." | CONFIRMED | primary | |
| D11 | ChatGPT: "profile icon → Settings → Data controls → Improve the model for everyone → off. For a one-off private chat, use Temporary Chat." | help.openai.com/en/articles/7730893; openai.com/consumer-privacy; privacy.openai.com | 403 ×3 | — | UNFETCHABLE (the page already says so; keep the caveat) | primary (unread) | |
| D12 | Gemini: "myactivity.google.com/product/gemini → Keep Activity → Turn off." | Gemini Apps Privacy Hub (updated October 26, 2026); support.google.com/gemini/answer/13278892 | OK (both) | Hub: "To stop your future chats from being reviewed to improve Google services, turn off your Keep Activity setting." (links myactivity.google.com/product/gemini). Steps article: "On your computer, go to gemini.google.com." → "Click Settings & help" → "Activity" → "Near the top, click On and then Turn off or Turn off and delete activity." | CONFIRMED (setting name and "Turn off" confirmed; consider adding Google's in-app route: Gemini → Settings & help → Activity → On → Turn off) | primary | |
| D13 | Gemini: "Chats are still kept 72 hours; chats people already reviewed, up to three years." | Privacy Hub | OK | "Future chats are still saved for 72 hours."; "Chats reviewed by human reviewers … are not deleted when you delete your activity. Instead, they are retained for up to three years." | CONFIRMED | primary | |
| D14 | Meta AI: "since December 2025 Meta uses your Meta AI chats to personalise ads, with no opt-out." | about.fb.com, 1 Oct 2025 | OK | "We will soon use your interactions with AI at Meta to personalize the content and ads you see"; "goes into effect on December 16, 2025"; excluded topics "religious views, sexual orientation, political views, health"; controls: "you can continue to adjust the content and ads you're seeing at any time" via Ads Preferences and feed controls; the post describes no opt-out from AI-based personalisation | NOT IN SOURCE ("with no opt-out" is not stated) → "Meta's announcement offers no opt-out, only the existing ad-preference controls." Also: this bullet carries no source footnote; add about.fb.com. | primary | |
| D15 | Meta AI: "Privacy guides say there's no US opt-out from training on public posts either; Meta's own page wouldn't load for us." | facebook.com/privacy/genai | loads, but "Where does Meta get training information?" and other sections render empty | — | UNFETCHABLE (as the page says) | primary (unread) | |
| D16 | Real: "Drag an image or video into contentcredentials.org/verify." | contentcredentials.org/verify → 301 → verify.contentauthenticity.org | redirect OK; target is a JS app (only the title "Content Credentials" rendered) | — | CONFIRMED that the tool lives at that address; the drag-and-drop instruction text could not be read | primary (tool) | |
| D17 | Real: "A credential proves the file hasn't changed, not that it's true, and most sites strip them" | C2PA FAQ; CAI | OK | as L31a/L31b | CONFIRMED (first half) / NOT IN SOURCE ("most sites") — same fix as L31b | primary | |
| D18 | Real: "Ask Gemini whether Google made it. Upload the image and ask 'Was this created with Google AI?' It only recognises Google's own images." | blog.google 20 Nov 2025; deepmind.google/technologies/synthid (current); blog.google SynthID Detector 20 May 2025 | OK (all three) | 2025 blog: "Was this created with Google AI?"; "Gemini will check for the SynthID watermark"; "Soon, we'll expand SynthID verification to support additional formats beyond images, such as video and audio"; "we will also extend our verification approach to support C2PA content credentials". DeepMind page now: "Simply upload the image, video or audio clip to your chat, and ask if it's been created or altered by Google AI."; "Gemini, Search and Chrome will check for a SynthID watermark, and let you know if it finds one."; Detector portal: "anyone can easily check if an image, video, or audio file was made with AI from Google or our partners" | NEWER DATA → "Upload an image, video or audio clip to Gemini and ask whether it was created or altered by Google AI. It checks for Google's SynthID watermark (Google's and partner models'); other companies' AI output won't be flagged." Add a source footnote (none now). | primary | |
| D19 | Real: "People take AI-written text for human 77% of the time, and free deepfake detectors lose about half their accuracy on real fakes." | IASR; Deepfake-Eval-2024 | OK | as L10, L11 | CONFIRMED (wording: "open-source"; AUC) | report / preprint | |
| D20 | Real: "'AI-generated' labels already make people doubt true headlines." | Altay & Gilardi | OK | as L13 | CONFIRMED | research | |
| D21 | Real: "The FBI's advice: a secret word, and hang up and call back on a number you already have." | FBI PSA 3 Dec 2024; FTC alert 20 Mar 2023 | OK (both) | FBI: "Create a secret word or phrase with your family to verify their identity."; FBI's hang-up tip is about institutions: "Verify the identity of the person calling you by hanging up the phone, researching the contact of the bank or organization purporting to call you, and call the phone number directly." FTC: "Don't trust the voice. Call the person who supposedly contacted you and verify the story."; "Use a phone number you know is theirs."; "If you can't reach your loved one, try to get in touch with them through another family member or their friends." The FTC alert does not mention a safe word. | CORRECTED → "The FBI's advice: a secret word or phrase. The FTC's: hang up and call them back on a number you already know." | primary | |
| D22 | Feeds intro: Meta three-month experiment; "users liked the algorithm's political picks less than the time-ordered ones" | Guess; Milli | OK | as L16, L15 | CONFIRMED | research | |
| D23 | Instagram: "tap the Instagram logo at the top left → Following. It resets each time you open the app." | about.instagram.com "Instagram Ranking Explained", 31 May 2023; help.instagram.com (JS shell) | blog OK; help centre unreadable | "You can also customize your feed to only show content from your Favorites or from accounts you follow"; "(Following will show you the last 30 days of posts from those you follow in chronological order)."; caption: "Tap 'Following' from the Feed dropdown menu to see content from accounts you follow in chronological order." Nothing about the logo at top left or the choice resetting. | NOT IN SOURCE (the logo wording and the reset) → "Instagram: open the Feed dropdown at the top → Following (the last 30 days from accounts you follow, in time order). Instagram doesn't document whether the choice sticks between sessions." Following = chronological is CONFIRMED. | primary (2023) | |
| D24 | Facebook: "the Feeds tab → Friends or Favorites, newest first." | about.fb.com "Introducing Home and Feeds on Facebook", 21 Jul 2022 | OK | "We're launching the Feeds tab, a new way to find the most recent posts from your friends, Favorites, Pages and groups."; "There are no Suggested For You posts in Feeds and ads are included."; "on iOS, this bar is found at the bottom of the app, and on Android, it's found at the top." | CONFIRMED (against a 2022 announcement; current help-centre text not fetched) | primary (2022) | |
| D25 | X: "Following is ranked by default now. On the web, click the arrow next to Following → Sort by → Most recent." | help.x.com/en/using-x/x-timeline; help.twitter.com mirror | 403 ×2 | — | UNFETCHABLE (the research file's only support was a secondary blog, Feb 2026) | primary (unread) | |
| D26 | YouTube: "the Subscriptions tab. Turning off and clearing watch history empties the recommendations on Home." | support.google.com/youtube/answer/95725 | OK | "If you have no relevant or significant recent watch history and have indicated you do not wish to receive future recommendations based on that history, YouTube features like recommendations on the YouTube homepage are removed."; steps: "Go to My Activity." → "Controls" → "Select Turn off." | CONFIRMED (with Google's condition: no significant recent history) | primary | |
| D27 | TikTok: "the Following tab. There's no un-personalised feed in the US." | support.tiktok.com (redirects to tiktok.com/support/faq_detail… which renders only "TikTok Support"); newsroom DSA URL resolved to an unrelated page | unreadable | — | UNFETCHABLE | primary (unread) | |
| D28 | Kids: "Common Sense Media rates AI companion apps like Character.AI and Replika 'unacceptable' for anyone under 18: age gates were easy to get past, and the bots gave harmful advice and claimed to have feelings." | CSM 30 Apr 2025 | OK | "Character.AI, Nomi, Replika, and others"; rest as L19 | CONFIRMED | charity's own | |
| D29 | Kids: "Snapchat's My AI, Meta AI in Instagram and WhatsApp" | help.snapchat.com My AI article; about.fb.com | OK | Snapchat: "My AI is a chatbot currently available to Snapchatters." Meta post refers to "AI at Meta" across its apps | CONFIRMED | primary | |
| D30 | Kids: "Keep chatbot use in a shared room for younger kids, and read a few chats together." | AAP | OK | "Encourage them to work in a common area so you can keep loose tabs on what they're viewing."; "Suggest looking together at your child's chats with the bot." | CONFIRMED | primary | |
| D31 | Kids: "bots are built to keep you talking and to tell you what you want to hear. Private things go to real people." | AAP | OK | "the chatbot's only real goal is to tell them what they want to hear and keep them engaged."; "Private information should only be shared with parents, family members or trusted friends." | CONFIRMED | primary | |
| D32 | Kids: "Watch for pulling away from friends, turning to the bot when upset, or the bot saying sexual, violent or false things. Those are the pediatricians' warning signs." | AAP | OK | No "warning signs" list. "Talk with your child's doctor if they seem withdrawn and prefer talking with chatbots rather than people."; "Kids may be drawn to chatbots when they feel frustrated with the people in their lives. This can create an unhealthy cycle…"; "chatbots may tell kids false, harmful, highly sexual or violent things." | CORRECTED → "The AAP's advice: talk to your child's doctor if they seem withdrawn and prefer the bot to people. It warns that kids turn to bots when frustrated with people, and that bots may say false, harmful, highly sexual or violent things." | primary | |
| D33 | Inner: "An eight-week mindfulness course. Across 47 trials, it eased anxiety, depression and pain modestly, though no better than exercise, medication or therapy." | Goyal 2014 | OK | 47 trials, 3,515 participants; anxiety ES 0.38 at 8 weeks, depression 0.30, pain 0.33 — "moderate evidence"; "We found no evidence that meditation programs were better than any active treatment" (drugs, exercise, other behavioral therapies) | CONFIRMED (8 weeks is the measurement point; most programmes were 8-week MBSR-style) | research | |
| D34 | Inner: "Two weeks without mobile internet (calls and texts still work). Wellbeing, mental health and measured attention all improved, though only a quarter of people managed ten days." | Castelo 2025 | OK | 467 participants, preregistered RCT; "maintain mobile connection (through texts and calls)"; d = 0.45 / 0.56 / 0.23 (pooled); attention measured with gradCPT; 119 participants (25.5%) compliant, meaning the block was active for at least 10 of the 14 days | CONFIRMED | research | |
| D35 | Inner: "Across 148 studies, people with stronger social ties had 50% higher odds of survival." | Holt-Lunstad 2010 | OK | "Across 148 studies (308,849 participants), the random effects weighted average effect size was OR = 1.50 (95% CI 1.42 to 1.59), indicating a 50% increased likelihood of survival for participants with stronger social relationships." | CONFIRMED | research | |
| D36 | Law: "There's no federal AI law … a December 2025 executive order set up a Justice Department task force to challenge them, and the Senate had earlier voted 99 to 1 against a ten-year ban on state AI laws." | EO 14365; Inside Privacy | OK | as L21, L26, L27 ("the Attorney General shall establish an AI Litigation Task Force") | CONFIRMED (99–1 against the cited secondary; primary roll call unfetchable) | primary / secondary | |
| D37 | Law: "In New York, two laws take effect on 1 January 2027: the RAISE Act on the largest AI developers, and a law on AI companions for minors signed on 28 October 2026." | S8828; nysenate.gov S9051B; IAPP | OK (all) | S8828 as L25. S9051B: "signed chap.368" dated Oct 28, 2026; "This act shall take effect January 1, 2027."; "TO OBTAIN CIVIL PENALTIES OF UP TO TWENTY-FIVE THOUSAND DOLLARS PER VIOLATION". IAPP: "28 Oct. 2026"; "The law enters into force in 1 Jan. 2027" | CONFIRMED | primary | |
| D38 | Law: links to house.gov, senate.gov, pluralpolicy.com | the three pages | OK (all) | "Find Your Representative | house.gov" — "Enter your ZIP code:"; "U.S. Senate: Contacting U.S. Senators" — "Select your state"; Plural: "Find your Local Legislator." | CONFIRMED | primary |
| D39 | Q&A 1: "Letting it produce the answer cost students and engineers once it was removed; using it as a tutor didn't, and one 2026 preprint found a lasting gain. The viral MIT brain-scan study is unreviewed and small." | Bastani; Anthropic; Contractor & Reyes; Kosmyna | OK | as above | CONFIRMED (engineers were quizzed without AI right after the task) | — | |
| D40 | Q&A 2: training off does "Something narrow … Nothing comes out of models already trained" | Anthropic news | OK | "Your data will still be included in model training that has already started" | CONFIRMED | primary | |
| D41 | Q&A 3: "ranking a feed for engagement measurably raised anger, and turning hostile posts down measurably cooled feelings toward the other side." | Milli; Piccardi | OK | Piccardi: "shifted out-party partisan animosity by two points on a 100-point feeling thermometer" | CONFIRMED | research | |
| D42 | Q&A 4: "The EU delayed its strictest rules and New York cut the RAISE Act's penalties, but the EU's labelling rules started on time and a dozen states passed chatbot-safety laws in 2026." | as L36 | OK | as L36 | CONFIRMED (the state count has no footnote source on this page either — add Multistate / Transparency Coalition to s14) | — | |
| D43a | Q&A 6: "The one binding AI treaty, the Council of Europe's, isn't in force yet." | coe.int Treaty Office (CETS 225) ×3; conventions.coe.int; consilium.europa.eu; Wikipedia | all blocked / redirect loop; Wikipedia OK but silent | Wikipedia (secondary, carries an LLM-text banner): entry-into-force condition "five ratifications made by five countries, including three member states of the Council of Europe"; says nothing about current ratifications or status. | UNFETCHABLE (status could not be read from any official page today; the research file's 15 May 2026 coe.int item was itself a search summary). Keep the claim only with an "unverified since …" label, or cite the research note's date. No source footnote on the page — add one when found. | — | |
| D43b | Q&A 6: "talks on autonomous weapons, meets 16–20 November 2026 to decide whether to start negotiating." | meetings.unoda.org GGE LAWS 2026 page; three UNODA Review-Conference URLs (404); GGE rolling text PDF 5 Jun 2026; Stop Killer Robots news, 22 Oct 2026 | UNODA GGE page OK (no dates); PDF OK; SKR OK | UNODA: GGE sessions "from 2-6 March 2026 and from 31 August-4 September 2026"; mentions "the close of the mandate and the CCW Review Conference" without dates. Rolling text: "nothing has reached consensus until everything has reached consensus, the Group provisionally found consensus on the following formulations". Stop Killer Robots (22 Oct 2026): "In less than one month, states and civil society will meet for the Seventh Review Conference" "at the United Nations in Geneva"; "At the Review Conference, states will decide their next steps in how to address autonomous weapons under the CCW." The exact dates 16–20 appear only in a link slug on that page. | UNFETCHABLE (exact dates from a primary page); "November 2026, Geneva, decides next steps" CONFIRMED from a secondary. Suggest "meets in Geneva in late November 2026" unless a UNODA page is found. No footnote on the page — add one. | secondary | |
| D44 | Footer: "The autonomous-weapons line goes stale after 20 November 2026." | consistent with D43b | — | — | n/a | — | |
| D45 | Sources: s1 "PNAS 2025 (read through an index)"; s3 "a published comment"; s4 "Lee et al. (Microsoft Research and Carnegie Mellon)"; s5 Gemini hub "updated 26 October 2026"; s7 Deepfake-Eval "May 2026" graded research; s9 FBI "3 December 2024", FTC "March 2023"; s12 AAP "Are AI chatbots safe for kids?, updated 27 August 2025"; s14 IAPP "October 2026" | as above; Lee PDF author block | OK | Lee PDF: "Hao-Ping (Hank) Lee — Carnegie Mellon University, Pittsburgh"; all other authors "Microsoft Research, Cambridge, United Kingdom". Deepfake-Eval v5 "Wed, 27 May 2026". FTC "March 20, 2023". IAPP "28 Oct. 2026". | s1 NEWER DATA (now read in full); s3 CORRECTED (not published; arXiv comment); s4 CONFIRMED; s5 CONFIRMED; s7 CORRECTED grade → preprint; s9 CONFIRMED; s12 CORRECTED title (as L47); s14 CONFIRMED | — |
4. Practice — content/ai-practice.html
| # | Words on the page | Source | Fetch | Source's words | Verdict | Grade |
|---|---|---|---|---|---|---|
| P1 | Daily: "Write three bullets first … The students who kept their skill were the ones who did the work before the bot did." (cites s1 = Bastani) | Bastani | OK | Bastani compared hints vs answers; nobody in it drafted before using the bot. The write-first evidence is Kosmyna's 18-person session ("Brain-to-LLM users exhibited higher memory recall"). | NOT IN SOURCE → "The students who kept their skill were the ones whose bot made them do the work: hints, not answers." (or cite Kosmyna for write-first, with its 18-person caveat) | research |
| P2 | Drill 1: "people judge AI text human 77% of the time and free detectors miss half of real fakes." | IASR; Deepfake-Eval-2024 | OK | "AUC decreasing by 50% for video, 48% for audio, and 45% for image models" — a drop in discriminating power, not a miss rate | CORRECTED → "open-source detectors lose about half their accuracy on real-world fakes" ("miss half" is not what the benchmark measured) | report / preprint |
| P3 | Drill 2: "Pediatricians list exactly this: pulling away from friends, turning to the bot when upset. Companion apps are rated unacceptable under 18." | AAP; CSM | OK | AAP as L20 (no list; "withdrawn and prefer talking with chatbots rather than people" is the doctor-visit trigger); CSM as L19 | CORRECTED → "Pediatricians say to talk to your child's doctor when a kid seems withdrawn and prefers the bot to people. Companion apps are rated unacceptable under 18." | primary |
| P4 | Drill 3: "The hint-only tutor in the classroom trial did no harm; the answer machine cost 17% once it was gone." | Bastani | OK | as L3, L4 | CONFIRMED | research |
| P5 | Drill 4: "In a 4,976-person experiment, the label lowered belief in headlines even when they were true and human-written." | Altay & Gilardi | OK | "two preregistered online experiments among US and UK participants (N = 4,976)"; "regardless of whether the headlines were true or false, and created by humans or AI." | CORRECTED (minor) → "In two experiments with 4,976 people" | research |
| P6 | Drill 5: "A cloned voice can sound exactly like her and a spoofed number can show hers. Ask for the family word. Hang up and call her on the number you have. The FBI's advice, word for word." | FBI PSA; FTC alert | OK | FBI: secret word ✓; "The PSA doesn't mention caller-ID spoofing, spoofed phone numbers, or calls appearing to come from a known number." FTC alert likewise contains no spoofing sentence. The call-back advice is the FTC's ("Use a phone number you know is theirs."); the FBI's hang-up tip concerns "the bank or organization purporting to call you". | CORRECTED → "A cloned voice can sound exactly like her. Ask for the family word (the FBI's advice). Hang up and call her on the number you have (the FTC's)." Drop "a spoofed number can show hers" (NOT IN SOURCE) or find a source for it. Drop "word for word". | primary |
| P7 | Drill 6: "The company still keeps them for a while, from 72 hours to years depending on the service and what you clicked. Nothing comes out of models already trained." | Google hub; Anthropic | OK | as L28, D7 | CONFIRMED | primary |
| P8 | "Six situations." | the page | count | 6 <details class="drill"> | CONFIRMED (arithmetic) | — |
| P9 | Rhythm: "Switch your feeds back to Following. Most reset themselves." | Instagram blog; X/TikTok unreadable | — | Instagram's text says nothing about resetting; X and TikTok pages unfetchable | NOT IN SOURCE → "Some reset themselves." | — |
| P10 | Say it in thirty seconds: 17% / hints / safe word | Bastani; FBI | OK | as above | CONFIRMED | — |
| P11 | Sources: s2 IASR "February 2026"; s3 AAP "August 2025", CSM "April 2025"; s4 FBI "3 December 2024"; s5 Google hub "October 2026", OpenAI "(read through a search index)" | as above | OK / 403 | IASR "3 February 2026"; AAP "8/27/2025" (title as L47); CSM "April 30, 2025"; FBI "December 3, 2024"; Google "October 26, 2026"; OpenAI 403 | CONFIRMED except: AAP title (CORRECTED, as L47); Deepfake-Eval graded "research" (CORRECTED → preprint); OpenAI still UNFETCHABLE | — |
5. CORRECTIONS["ai"] — tools/essays.py
| # | Words | Source | Fetch | Source's words | Verdict | Grade |
|---|---|---|---|---|---|---|
| C1 | "on one music service AI tracks were half of uploads on peak days but 1–3% of streams." | Deezer 21 Jul 2026 | OK | "more than 50% of total daily deliveries on peak days"; "'between 1-3%'" of streams | CONFIRMED | company's own figures |
| C2 | "The record since 2025 shows AI rules weakened and delayed more often than erased." | synthesis of L24, L25, L36, L27 | OK | each component confirmed above (EU delay to 2 Dec 2027 / 2 Aug 2028; Colorado replaced before 30 Jun 2026; NY penalties $10M/$30M → $1M/$3M; Art. 50 applicable 2 Aug 2026; 12–13 states' chatbot laws; 99–1) | CONFIRMED (synthesis) | mixed |
| C3 | "Nothing with the weight of the Geneva Conventions exists yet. The one binding AI treaty isn't in force." | Council of Europe Treaty Office | 403 / redirect loop | — | UNFETCHABLE (as D43a) | — |
6. Dated items in source/stale.tsv for this topic — status on 2026-11-05
| Due | Item | Status today | Evidence |
|---|---|---|---|
| 2026-11-20 | Autonomous-weapons talks 16–20 Nov 2026: did they agree to start negotiating? | Not yet happened. The Seventh CCW Review Conference meets in Geneva this month and "states will decide their next steps in how to address autonomous weapons under the CCW" (Stop Killer Robots, 22 Oct 2026). Exact dates 16–20 Nov not confirmed from an official page (UNODA pages 404). The GGE's rolling text (5 Jun 2026) is "provisionally" agreed with "nothing has reached consensus until everything has reached consensus". | secondary (SKR) + primary (UNODA rolling text PDF) |
| 2026-12-02 | EU AI Act Art. 50 marking deadline for systems already on the market | Upcoming; unchanged. nicfab (secondary, 24 Jul 2026): providers of generative systems "placed on the market before 2 August 2026 must take the necessary steps to comply with Article 50(2)" by 2 December 2026; new Art. 5 prohibitions "apply from 2 December 2026". Commission page: "Prohibition 9 comes into effect in December 2026 and was introduced as a part of the AI Omnibus." EUR-Lex text of Reg. (EU) 2026/1744 did not render in three URL forms, so the primary wording of the transitional clause remains unread. | primary (Commission) + secondary (nicfab) |
| 2027-01-01 | NY RAISE Act and youth-chatbot law take effect | Upcoming; unchanged. S8828: "shall take effect … JANUARY 1, 2027" (chap.96, 27 Mar 2026). S9051B: "This act shall take effect January 1, 2027." (chap.368, 28 Oct 2026). | primary |
7. Arithmetic the pages do themselves
| Check | Holds? |
|---|---|
"Nine facts" / "six words" (Learn) — 9 li, 6 dt | yes |
| "Six situations" (Practice) — 6 drills | yes |
| "penalties cut tenfold" — $10M→$1M, $30M→$3M | yes |
| "17% worse" — paper: 17% | yes |
| "about a thousand" — "nearly 1,000" | yes |
| "a dozen states" — 12 (Multistate) / 13 (Transparency Coalition) | yes, within the disagreement |
| "lose about half their accuracy" — AUC −50% / −48% / −45% | yes in size; measure is AUC |
| "one to three percent of streams" — "between 1-3%" | yes |
| "104 of 500" — 104 (21%); 104/500 = 20.8% | yes |
| "50% higher odds of survival" — OR 1.50 | yes |
| "only a quarter … ten days" — 25.5% compliant ≥10 of 14 days | yes |
| "47 trials", "148 studies", "54 … 18", "319", "52", "4,976", "15,000+", "806", "1,256", "100+ experts", "77%", "72 hours", "three years", "five years" | all match the sources |
| "two laws take effect on 1 January 2027" — S8828, S9051B | yes |
| "since December 2025" — 16 Dec 2025 | yes |
| "2027 and 2028" — 2 Dec 2027, 2 Aug 2028 | yes |
| "16–20 November 2026" | not confirmable today (see D43b) |
8. Summary by verdict
Counting one verdict per numbered row (split rows L31a/b, D17, D43a/b and D45's sub-items counted separately where marked; n/a rows excluded):
| Verdict | Count | Rows |
|---|---|---|
| CONFIRMED | 81 | O1, O2, L1–L4, L6–L8, L10–L19, L21, L23–L28, L30, L31a, L32, L33, L35–L39, L42, L45, L46, L49, D1–D5, D7, D9, D10, D12, D13, D16, D17 (first half), D19, D20, D22, D24, D26, D28–D31, D33–D42, D45 (s4, s5, s9, s14 = 4), P4, P7, P8, P10, P11 (dates), C1, C2 |
| CORRECTED | 15 | L5, L9, L20, L41 (grade), L43 (grade), L47 (title), D21, D32, D45 (s3, s7 grade, s12 title = 3), P2, P3, P5, P6 |
| NOT IN SOURCE | 9 | L29, L31b, D6, D8, D14, D17 (second half), D23, P1, P9 (+ the "spoofed number" clause inside P6) |
| UNFETCHABLE | 9 | L22 (the lawsuit), D11, D15, D25, D27, D43a, D43b (exact dates), C3, P11 (OpenAI) |
| NEWER DATA | 5 | L40, L44, L48, D18, D45 (s1) |
| n/a (site's own / internal / editorial) | 5 | O3, O4, L34, D44, and reading-time estimates |
Grade-label changes recommended (inside the CORRECTED count): Anthropic study → "company's own research, not peer-reviewed"; Deepfake-Eval-2024 → "preprint"; IASR → "report" rather than "research"; CETaS/Vaccari/Schiff "via summary" → "read" (abstract or full text).
Uncited-but-confirmed items that need a footnote source added: Colorado SB 26-189 (Arnold & Porter), the state chatbot-law count (Multistate; Transparency Coalition), the Meta AI bullet (about.fb.com, 1 Oct 2025), the Gemini verification bullet (blog.google 20 Nov 2025; deepmind.google SynthID page), the feed-menu bullets (Instagram blog 2023; Meta newsroom 2022; YouTube help 95725), and the Q&A 6 treaty/CCW paragraph (none today).
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