Sources and fact-check

Showing all 71 claims.

Supported

Accurate as checked against the source.

Corrected

Core is right; wording was corrected to match the source.

Updated

Was true once; updated to the status as of October 2026.

Simulation

Result of this project's simulated feed (S28); not checked against outside sources.

Added later

Added after the main fact-check; checked against the cited source on October 18, 2026.

1. 1

   In November 2017, Molly Russell, a 14-year-old from Harrow, north-west London, died from an act of self-harm while suffering from depression and the negative effects of online content, after months of viewing self-harm and suicide content on Instagram and Pinterest.

   Fact-check note

   Facts (age 14, died 21 Nov 2017, Harrow, Instagram/Pinterest) confirmed. "Took her own life" does not match the legal finding: the coroner said suicide was not a safe conclusion. Use the coroner's wording.

   Site: Hook; Poster: Hook

   Corrected

   S29, S1, S2

2. 2

   During her inquest, it was revealed that of the 16,300 posts Molly saved, shared or liked on Instagram in the six months before her death, 2,100 were related to depression, self-harm or suicide.

   Fact-check note

   16,300 and 2,100 confirmed. The "130 a day" figure is a different count (3,500 shares + 11,000 likes + 5,000 saves, about 19,500) from pre-inquest PA reporting, not an average of the 16,300 (that would be about 89/day). Dropped from this sentence; the stat callout now attributes it separately.

   Site: Hook; Poster: Hook stat callouts

   Corrected

   S30, S1

3. 3

   Pinterest went as far as emailing her recommendations with subject lines such as "10 depression pins you might like."

   Fact-check note

   Confirmed: the email was shown to Pinterest's Judson Hoffman at the inquest. Caveat: Ian Russell's evidence quotes an "18 depression pins" email and says some emails arrived after Molly's death. S1 blocks automated fetches; S30 added.

   Site: Hook; Poster: Hook

   Supported

   S30, S1

4. 4

   These posts did not reach Molly by chance. The Children's Commissioner for England said the platforms' algorithms, "designed to maximise users' attention and engagement," recommended "an increasing volume of harrowing content to Molly," and the coroner found that the platforms' algorithms produced, in some circumstances, "binge periods of images, video-clips and text," some of it "selected and provided without Molly requesting them."

   Fact-check note

   "Which is why Molly continued to receive this harmful content" was a causal claim neither source makes. Now attributed to the Commissioner (advocate) and the coroner (finding).

   Site: Hook; Poster: Hook

   Corrected

   S2, S29

5. 5

   Molly's case is not isolated. The U.S. Surgeon General's 2023 advisory reports that up to 95% of youth aged 13 to 17 say they use a social media platform. It also cites a study of U.S. adolescents aged 12 to 15 in which those who spent more than three hours a day on social media faced double the risk of poor mental health outcomes, including symptoms of depression and anxiety. As of 2021, 8th and 10th graders spent an average of 3.5 hours a day on social media.

   Fact-check note

   Numbers right; wording fixed: the 2x figure comes from an outside cohort study (ages 12-15, association not causation) cited by the advisory, not a Surgeon General finding; 3.5 h is 8th/10th graders in 2021. Source switched from News-Medical to the HHS advisory itself.

   Site: Hook; Poster: Hook stat callouts

   Corrected

   S3

6. 6

   The harms produced by engagement-based ranking cannot be understood solely as technical failures, as they reflect value choices embedded in how these systems are designed and what they are built to prioritize. The four principles of biomedical ethics (respect for autonomy, beneficence, nonmaleficence, and justice), set out by Beauchamp and Childress, provide a framework for evaluating these choices and identifying where they fall short.

   Fact-check note

   Correct. Small precision edit (book's term "respect for autonomy"; authors named). Check which edition of Beauchamp and Childress you used: a 9th edition came out Feb. 2026.

   Site: Philosophical lens intro

   Supported

   S25, S26

7. 7

   Engagement-based ranking is built to predict which content a user is most likely to engage with and to keep supplying it, which can undermine the conditions that make voluntary choice possible, particularly for adolescents, whose capacity for impulse control and self-regulation is still developing. This builds on the argument, developed in earlier work on AI companions, that autonomy is not merely a choice but the set of conditions that make choice possible.

   Fact-check note

   "Least able to look away from" overstated the mechanism: rankers predict engagement (and include negative signals). "Undermines" softened to "can undermine". Adolescent development premise now cited to the Surgeon General's advisory.

   Site: Philosophical lens, Autonomy; Poster: Philosophical quadrant

   Corrected

   S25, S26, S3

8. 8

   These systems are built to predict and reward engagement, and the Surgeon General's advisory notes that social media platforms are often designed to maximize user engagement. On this essay's reading, any good the user receives is therefore incidental to the system's engagement objective. Meta itself describes its ranking as predicting how valuable a piece of content might be to a user so it can be shown sooner, combining a wide variety of predictions, some based on behavior (such as whether a user is likely to share a post) and some based on user feedback received through surveys.

   Fact-check note

   Meta sentence verified verbatim against the live page. The first sentence was the author's inference stated as fact ("time spent" is not on Meta's page); now marked as the essay's reading and backed by the advisory.

   Site: Philosophical lens, Beneficence; Poster: Philosophical quadrant

   Corrected

   S25, S18, S3

9. 9

   When platforms are aware that their ranking systems amplify harmful content to vulnerable users and continue to deploy them without meaningful safeguards, the resulting harm is foreseeable and therefore preventable. In Molly Russell's case, the coroner found that the platforms' use of algorithms resulted, "in some circumstances," in "binge periods" of images, video clips and text, some of it "selected and provided without Molly requesting them," and that the material she viewed contributed to her death "in a more than minimal way." The Children's Commissioner went further, saying the algorithms, "designed to maximise users' attention and engagement," recommended "an increasing volume of harrowing content" to her.

   Fact-check note

   The original put the Children's Commissioner's words ("engagement", "increasing volume") in the coroner's mouth. The coroner's PFD report never says "engagement" or "recommended". Now each phrase is attributed correctly; PFD report added as primary source.

   Site: Philosophical lens, Nonmaleficence; Poster: Philosophical quadrant

   Corrected

   S25, S31, S2, S26

10. 10

    The consequences of engagement-driven design do not fall evenly, as they are concentrated on those least able to protect themselves, including minors, users experiencing psychological distress, and populations in regions with limited regulatory oversight. In Myanmar, Facebook's engagement-based algorithms operated in a country where it held near-total market dominance by 2017 and had become, for many people, their primary news source, and Amnesty International concluded that Meta "substantially contributed" to the atrocities against the Rohingya.

    Fact-check note

    "Deployed the same systems it used everywhere else" is not in the Amnesty report and was removed. "Primary source of information" narrowed to Amnesty's "for many people, their primary news source". "Catastrophic consequences" replaced with Amnesty's finding.

    Site: Philosophical lens, Justice; Poster: Philosophical quadrant

    Corrected

    S25, S23, S64, S24

11. 11

    In effect since December 10, 2025.

    Fact-check note

    Confirmed by eSafety. A High Court challenge (Digital Freedom Project v Commonwealth; Reddit) was pending with no injunction as of the latest sources. Source S5 replaced: the AMES blog was unreadable (403) and is not a primary source.

    Site: Policy, Australia, status; Poster: Policy table

    Supported

    S5

12. 12

    Account access for users under 16 on platforms including Facebook, Instagram, Reddit, Snapchat, Threads, TikTok, X, YouTube, Kick, and Twitch; standalone messaging, online gaming, and most health and education services are not covered.

    Fact-check note

    Accurate ("including"), but eSafety's list has 10 platforms; Kick and Twitch added. Messaging services with social-media-style features can be covered, hence "standalone".

    Site: Policy, Australia, targets

    Supported

    S33, S5

13. 13

    Platforms must take reasonable steps to keep under-16s off their services and cannot rely on users self-reporting their age alone, with fines of up to AUD 49.5 million.

    Fact-check note

    Confirmed (150,000 penalty units for corporations = A$49.5M per eSafety's Sept. 2025 guidance). One checker saw a later eSafety page quoting A$109.2M (300,000 units); this was not confirmed in the adversarial pass. Verify the current figure on eSafety's site before printing.

    Site: Policy, Australia, engineering obligation

    Supported

    S34, S4

14. 14

    Moderate to low. Age assurance raises accuracy and privacy concerns, and only early outcome data have been reported. eSafety's three-month evaluation (July 2026) found that the share of 10-to-15-year-olds holding an age-restricted account fell from 52.4% to 42.1%. Many children kept their accounts, most often because the platform had not yet asked them to confirm their age or had estimated it incorrectly. Most children reported little to no change in their wellbeing. eSafety says it is too early to tell whether the ban's intended outcomes will be achieved.

    Fact-check note

    "No measured outcome data has yet been reported" is false as of Oct. 2026: eSafety published a three-month evaluation in July 2026.

    Site: Policy, Australia, feasibility

    Updated

    S35, S4

15. 15

    Status as of Oct. 18, 2026: Malaysia (under 16) in force since June 1, 2026, with age checks for existing users phased in over up to six months; Indonesia (under 16) restricting major high-risk platforms, with under-16 accounts deactivated in phases from March 28, 2026; Denmark (under 15) draft bill issued in September 2026, to go to Parliament in early 2027 with a target start of July 1, 2027; France (under 15) passed Parliament on July 21, but the Constitutional Council struck the ban down on Aug. 14; Greece (under 15) bill before Parliament, final vote scheduled for Oct. 22, starting Jan. 1, 2027 if passed; Spain (under 16) bill still in Congress and could stall now that Parliament is dissolved before the Nov. 29 snap election; the UK announced on June 15 that it will ban under-16s, with the first regulations due by the end of 2026 and the ban expected in early to spring 2027.

    Fact-check note

    Every country status had changed. Biggest: France's ban was struck down by the Constitutional Council (Aug. 14, 2026); Malaysia and Indonesia are now in force; the UK is no longer "consulting". The Greek vote on Oct. 22 will date this line if the poster is used after that.

    Site: Policy, planned bans, status; Poster: Policy table

    Updated

    S4, S36, S37, S38, S39, S40, S41, S42

16. 16

    Account access by age. Greece cites anxiety, sleep problems, and addictive platform design as its reasons, yet its proposed under-15 ban regulates access, through an age limit and age verification, rather than the design itself.

    Fact-check note

    Reasons confirmed. Added "proposed" (vote not yet held). Neither checker read the Greek bill text itself.

    Site: Policy, planned bans, targets

    Supported

    S4, S43

17. 17

    Uncertain. Critics including Amnesty Tech argue that such bans are ineffective and out of step with how young people live online, and there are concerns about privacy and invasive age verification.

    Fact-check note

    Confirmed. Minor wording: TechCrunch does not name "privacy advocates"; Amnesty's own statement raises the privacy point. Amnesty's primary statement added.

    Site: Policy, planned bans, feasibility

    Supported

    S4, S44

18. 18

    Final rules released in July 2026 (July 28); the law takes effect January 25, 2027.

    Fact-check note

    Confirmed (published in State Register July 29, effective 180 days later). No lawsuit or injunction found as of Oct. 2026 (search only).

    Site: Policy, New York, status; Poster: Policy table

    Supported

    S6, S45

19. 19

    Algorithmically personalized "addictive feeds", and notifications about those feeds between midnight and 6 a.m. Eastern, for New York users under 18 (in practice, any user the platform has not confirmed as an adult through age assurance), unless a parent gives verifiable consent; applies to platforms where at least 20% of monthly active users' time is spent on such feeds, with an exemption for smaller platforms.

    Fact-check note

    Night ban covers only notifications about the addictive feed; the 20% test is aggregate, not per user; small platforms are exempt. The rule applies by default to every user not confirmed as an adult, not just known minors.

    Site: Policy, New York, targets

    Corrected

    S45, S6

20. 20

    Age assurance with at least one option other than government ID, certified (tested) annually with results retained for 10 years; a default feed for minors limited to accounts they follow or select, in a set sequence such as chronological order, instead of the algorithmically personalized feed; data minimization and deletion for age and consent data.

    Fact-check note

    Each element confirmed in the rule text. "Non-personalized feed" made precise (follow-or-select, set sequence).

    Site: Policy, New York, engineering obligation

    Supported

    S6, S45

21. 21

    Partly in effect. On September 9, 2025, the Ninth Circuit refused to block the addictive-feed and private-by-default provisions at the preliminary-injunction stage, so they can be enforced. It also held that the limit on showing like counts is likely unconstitutional. Litigation continues. In August 2026 a federal judge again refused to block the feed provisions, this time in new suits by Meta, TikTok and YouTube, and those companies are appealing. The age-assurance requirements begin January 1, 2027.

    Fact-check note

    The 2025 ruling was a preliminary-injunction decision, not a final "upheld/struck"; the opinion never mentions a trial (that was Hunton's gloss); notification provisions remain enjoined. Added the Aug. 2026 district court ruling.

    Site: Policy, California, status; Poster: Policy table

    Corrected

    S46, S47, S7

22. 22

    Addictive feeds, defined as feeds where user-generated or user-shared content is recommended, selected, or prioritized based on information provided by or associated with the user, offered to minors without verifiable parental consent; private-by-default accounts.

    Fact-check note

    Fair compression of § 27000.5; small precision edit.

    Site: Policy, California, targets

    Supported

    S48, S46, S7

23. 23

    Contested in court; in September 2025 the Ninth Circuit held that the default barring minors' accounts from showing like counts was likely unconstitutional (a content-based rule that fails strict scrutiny) and ordered it enjoined, while letting the private-by-default setting and, for now, the addictive-feed limits stand. This suggests the legal limit falls on content-based design rules rather than on regulating interface design as such.

    Fact-check note

    The original inference ("limits of regulating design features one at a time") was not what the court held: two of three design provisions survived, and the line was content-based vs content-neutral.

    Site: Policy, California, feasibility

    Corrected

    S46, S47

24. 24

    Children's safety duties and Ofcom's Protection of Children Codes in force since July 25, 2025.

    Fact-check note

    Date correct; clarified that it is the children's duties and Codes, not the whole 2023 Act, that started that day.

    Site: Policy, UK, status; Poster: Policy table

    Supported

    S8

25. 25

    Under Ofcom's Protection of Children Code (in force 25 July 2025), services likely to be accessed by children that run a content recommender system and are assessed as medium or high risk for certain harmful content are recommended to design their feeds so that such content is kept out of children's feeds or given low prominence. Primary priority content (e.g., pornography, suicide and self-harm material) should be excluded outright. Other priority and identified non-designated content may be excluded or demoted. The Act requires a children's risk assessment and a duty to keep it up to date, and Ofcom recommends reviewing it at least annually.

    Fact-check note

    "Must" overstated: the Codes are recommended (safe-harbour) measures. Primary priority content must be excluded, not just demoted. "Annual" assessment is Ofcom guidance, not statute. Poster version can be shortened, but keep "recommended".

    Site: Policy, UK, targets

    Corrected

    S49, S50, S8

26. 26

    Recommender systems set up to filter the most harmful content out of children's feeds and reduce the prominence of other harmful content (Ofcom does not require classifiers or any specific technology), a children's risk assessment kept up to date (Ofcom recommends reviewing it at least annually), highly effective age assurance on services that allow priority harmful content, including higher-risk services with recommender feeds, and tools for reporting and complaints.

    Fact-check note

    "Classifiers" was the author's engineering gloss; the Code says it does not recommend any specific technology. "Yearly" is a recommendation.

    Site: Policy, UK, engineering obligation

    Corrected

    S49, S50, S8

27. 27

    In force: the Digital Services Act has applied to online platforms since 17 February 2024 (micro and small enterprises are exempt from Article 27). Article 38 applies to very large online platforms and very large online search engines, with obligations starting four months after the Commission designates them.

    Fact-check note

    Correct; added dates and scope. Old Presencis sources are dead (HTTP 410), replaced with EUR-Lex.

    Site: Policy, DSA, status

    Supported

    S9, S10

28. 28

    Online platforms that use recommender systems must set out in their terms and conditions, in plain and intelligible language, the main parameters of their recommender systems, including the criteria most significant in determining the information suggested and the reasons for the relative importance of those parameters; very large online platforms must also offer at least one option for each recommender system that is not based on profiling.

    Fact-check note

    Faithful paraphrase of Arts. 27 and 38; precision edit only.

    Site: Policy, DSA, targets; Poster: Policy table

    Supported

    S9, S10

29. 29

    Yes, but only at the level of the system as a whole: because the "main parameters" must be set out in the terms and conditions, the requirement works as a global explanation requirement (our reading; the DSA does not use that term). However, the requirement can be interpreted in multiple ways, and commentators report that the first round of DSA audits showed variation in how platforms defined it.

    Fact-check note

    "Global explanation" is the author's reading, not the law's label (Art. 27(2) even says the parameters "shall explain why certain information is suggested"). The audit finding is KGI's claim.

    Site: Policy, DSA, XAI requirement; Poster: Policy table

    Corrected

    S9, S11

30. 30

    Feasible, but Article 27's disclosure requirements can be read in multiple ways, and without consistent disclosures it is hard to compare how recommender systems are optimized across platforms; the Knight-Georgetown Institute's Better Feeds guidelines call for platforms to disclose specific input data and weights.

    Fact-check note

    Source says consistent disclosures "would allow" comparison, not that disclosures "cannot be compared". "Researchers" is specifically KGI.

    Site: Policy, DSA, feasibility

    Corrected

    S11

31. 31

    The Commission adopted the delegated act on researcher data access on 1 July 2025 and published it on 2 July 2025. It appeared in the Official Journal as Delegated Regulation (EU) 2025/2050 and entered into force on 29 October 2025, when vetted researchers could start requesting platform data. Vetting is slower than the Commission's "as of October" target: in May 2026 national Digital Services Coordinators were still assessing the first 49 applications.

    Fact-check note

    "Expected" framing out of date. Act dated 1 July 2025, in force 29 Oct 2025; first 49 applications still under assessment in May 2026.

    Site: Policy, DSA risk, status

    Corrected

    S51, S14, S52

32. 32

    Very large platforms must assess and mitigate systemic risks, undergo independent audits, and give vetted researchers access to internal data for studying systemic risks in the EU.

    Fact-check note

    Matches DSA Arts. 34, 35, 37, 40(4). Nuance: data access is on reasoned request via the Digital Services Coordinator, for vetted researchers only.

    Site: Policy, DSA risk, targets

    Supported

    S9, S14

33. 33

    Promising but still being put into practice; a 2025 Knight-Georgetown Institute analysis says platforms have given inconsistent access to publicly available data, and calls for validated, consistent metrics of harm, particularly for at-risk populations.

    Fact-check note

    Source is a KGI policy analysis, not researchers reporting their experience; it calls for validated metrics rather than stating they are lacking. The Commission's Oct. 2025 preliminary findings against TikTok and Meta on researcher access back up the "inconsistent access" point.

    Site: Policy, DSA risk, feasibility

    Corrected

    S11, S53

34. 34

    Adopted by the European Commission in July 2025 (adopted 2 July, presented 14 July, published in the Official Journal 10 October 2025).

    Fact-check note

    Correct; added precise milestones. Pick one date consistently if you show a single date (14 July = publication).

    Site: Policy, DSA minors, status; Poster: Policy table

    Supported

    S54, S12

35. 35

    Under the European Commission's 2025 guidelines on protecting minors under the DSA, minors should be able to reset their recommended feeds completely and permanently, get a meaningful explanation of why each piece of content was recommended, and modify or influence how their recommender systems work. Platforms should not expose minors to persuasive design aimed mainly at engagement, such as indefinite scrolling and virtual rewards for repeated actions, and features such as streaks should be off by default.

    Fact-check note

    "Daily rewards" is CADE's paraphrase; the guidelines say "virtual rewards for performing (repeated) actions". Streaks appear in the default-off list, not the "avoid" list. Now cited to the guidelines themselves.

    Site: Policy, DSA minors, targets; Poster: Policy table

    Corrected

    S54, S12

36. 36

    Prohibitions have applied since February 2, 2025, with national enforcement powers from August 2, 2025 and fines of up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher.

    Fact-check note

    Correct and current. The 2026 Digital Omnibus (Reg. 2026/1744) did not change these dates or amounts; it added new Art. 5 prohibitions from 2 Dec 2026.

    Site: Policy, AI Act, status

    Supported

    S55, S13

37. 37

    AI systems that exploit vulnerabilities linked to age, disability, or a specific social or economic situation, with the objective or effect of materially distorting a person's behaviour in a manner that causes, or is reasonably likely to cause, significant harm (AI Act, Art. 5(1)(b)).

    Fact-check note

    Statute says "materially distorting" (not "alter") and covers objective OR effect. Now quoted from the Act.

    Site: Policy, AI Act, targets; Poster: Policy table

    Corrected

    S55, S13

38. 38

    Approved by the Senate Commerce Committee by voice vote on August 5, 2026 and awaiting a full Senate vote; a prior version passed the Senate 91-3 in July 2024 but stalled in the House.

    Fact-check note

    Confirmed. Merged with sweep item U2 (year and House outcome). See U1 for the Sept. 30 Rand Paul block.

    Site: Policy, KOSA, status; Poster: Policy table

    Supported

    S15, S56

39. 39

    A duty of care requiring platforms to exercise reasonable care in the design of their features to prevent and mitigate specified harms to minors (such as eating disorders, suicidal behaviors and compulsive usage), limits on addictive design features by default, and protective default settings for minors.

    Fact-check note

    The duty is design-focused and limited to listed harms, not a general duty to avoid harm to children. Checked against the introduced bill text; the committee substitute was not retrieved.

    Site: Policy, KOSA, targets

    Corrected

    S57, S15

40. 40

    In March 2026, a Los Angeles jury found Meta and YouTube negligent; it awarded $3 million in compensatory damages and recommended a further $3 million in punitive damages. Meta filed a notice of appeal in July 2026, and YouTube filed its own less than a week later.

    Fact-check note

    Matches AP. Added YouTube's appeal (AP, July 15). No appellate ruling as of Oct. 2026.

    Site: Policy, KGM, status; Site: Public lens; Poster: Policy table

    Supported

    S16, S58

41. 41

    Design features such as infinite scroll, endless feeds, and autoplay, which plaintiffs targeted to sidestep Section 230 protections for third-party content.

    Fact-check note

    Matches AP almost verbatim (infinite scroll, endless feeds, autoplay). Appeals do not affect a claim about plaintiffs' strategy.

    Site: Policy, KGM, targets

    Supported

    S16

42. 42

    On March 24, 2026, a New Mexico jury found that Meta violated the state's Unfair Practices Act and imposed a $375 million civil penalty (75,000 violations at the $5,000 statutory maximum). On August 6, 2026, after a separate bench trial, the judge found that Meta's platforms created a public nuisance and ordered Meta to pay a further $567 million into an abatement fund, bringing the total to about $942 million. Meta has asked the court to amend the judgment or grant a new trial, and says it will appeal.

    Fact-check note

    The March facts are right, but the remedies phase finished in August 2026 ($567M abatement fund, ~$942M total). The judge refused to order changes to infinite scroll or autoplay, saying that needs legislative or executive action. That directly supports the thesis.

    Site: Policy, New Mexico, status; Poster: Policy table

    Updated

    S59, S17, S60

43. 43

    Among the practices at issue were the role of Meta's algorithms in prioritizing sensational or harmful content, allegedly addictive features, and statements about platform safety, which the jury found false or misleading.

    Fact-check note

    "Potentially misleading" understated the verdict: the jury found the statements false or misleading. The jury did not find the ranking design itself unlawful.

    Site: Policy, New Mexico, targets

    Corrected

    S17

44. 44

    Healthcare offers a useful comparison. The Coalition for Health AI's (CHAI) voluntary draft guide sets out a six-stage lifecycle in which designing the AI system comes before it is engineered, assessed, piloted, or deployed, and it stresses independent review to uncover risks before deployment. In the United States, social media ranking systems face no comparable general requirement to be evaluated at the design stage, so scrutiny has largely come after harm has occurred, often through litigation. Europe has started to move: the EU's Digital Services Act requires very large platforms to assess systemic risks from their algorithmic systems before deploying high-impact features, and the UK's Online Safety Act requires a new children's risk assessment before any significant change to a service's design or operation.

    Fact-check note

    CHAI is voluntary and still a draft. "No equivalent requirement" was false outside the US: DSA Art. 34(1) and UK OSA s.11(4) require pre-deployment risk assessment. S27 pointed at CHAI's homepage; now the guide itself.

    Site: Policy, Toward design-first policy; Poster: Policy quadrant

    Corrected

    S27, S9, S50

45. 45

    A design-first approach would require platforms to answer the same questions an auditor would ask: what requirement applies, what evidence proves it, who reviewed the evidence, what failed, and what changed because of the finding. An overall accuracy figure, verbal assurances, or informal promises to fix issues later count as weak evidence.

    Fact-check note

    Five audit questions match the workbook verbatim; calling the three items "weak evidence" is a fair reading.

    Site: Policy, Toward design-first policy

    Supported

    S22

46. 46

    Because platforms generally do not make their trained ranking models available for outside inspection, this project uses a simulated feed to demonstrate how explainable AI methods could be applied to engagement-based ranking and where they fall short. A gradient-boosted model was trained on 20,000 simulated post impressions to predict engagement, and then used to rank the feed of one simulated 14-year-old who has recently engaged with distress-related content. Because the data is simulated, the true drivers of engagement are known, which makes it possible to check the explanations against the truth.

    Fact-check note

    External premise softened: X open-sourced its For You ranking code and blend weights in 2026 (not its trained models or data). Simulation numbers pending.

    Site: Technical lens intro; Poster: Technical quadrant

    Simulation

    S28, S67

47. 47

    For the top-ranked post, LIME named the user's past engagement with the topic as the main reason in 39 of 40 reruns, but its simple model reproduced only about 62% of the ranker's local behavior, and it ranked "hours since posted" as the third most important factor, which SHAP did not.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, LIME; Poster: Technical quadrant

    Simulation

    S28

48. 48

    SHAP named the emotional intensity of the post as the largest contributor, nearly tied with the user's past engagement with the topic, and its contributions summed exactly to the model's score. Across the whole dataset, SHAP's ranking of features closely matched the true drivers of engagement (rank correlation 0.964).

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, SHAP; Poster: Technical quadrant

    Simulation

    S28

49. 49

    Describe the smallest meaningful change that would produce a different outcome, which, under Dr. Bent's framing, must be actionable: within the person's control and achievable within the relevant time frame, and never involving attributes the person cannot change.

    Fact-check note

    Every element appears on the workbook page.

    Site: Technical lens, Counterfactuals

    Supported

    S21

50. 50

    No change to the teen's own engagement history alone could remove the top post from her top three; even with no past engagement with the topic at all, it would still rank #2. It only fell out of the top three when the post's emotional intensity was reduced from 0.86 to 0.05, a property of the content and the ranking, not of the user.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, Counterfactuals; Poster: Technical quadrant

    Simulation

    S28, S21

51. 51

    Distress-related posts made up 15.7% of all simulated posts but 47.4% of the top 10% of rankings, even though the model was never told which posts were distress-related; it learned to favor them because they produce engagement.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, proxy finding; Poster: Technical quadrant headline

    Simulation

    S28

52. 52

    Retraining the ranker without emotional intensity as a signal lowered the share of distress-related posts in the top 10% from 47.4% to 38.0%, but the same two distress-related posts remained at the top of the teen's feed, because correlated signals such as watch completion and comment rate carried the same information.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, design-side counterfactual; Poster: Technical quadrant

    Simulation

    S28

53. 53

    Dr. Bent's evaluation framework asks that tests match the explanation's intended claim and its audience. Each test below is scored on the simulated demo.

    Fact-check note

    Matches the chapter's stated outcome. Human usefulness was judged analytically, not tested with users (the page says so).

    Site: Technical lens, five tests intro

    Supported

    S20

54. 54

    SHAP passes: its contributions reproduce the model's score exactly. LIME is partial: its local surrogate explained about 62% of the model's behavior around the post.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, Fidelity; Poster: five tests scorecard

    Simulation

    S28, S20

55. 55

    Removing the two features SHAP ranked highest lowered the post's score by 0.354, compared with an average of 0.066 when two features were removed in random order, indicating that SHAP identified features the model actually uses.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, Faithfulness; Poster: five tests scorecard

    Simulation

    S28, S20

56. 56

    SHAP's top three features accounted for 84.4% of the post's ranking. A platform-style one-line reason (a paraphrase of TikTok's "Why this video" wording), "This video is similar to content you've interacted with," accounts for only 35.1%, leaving 64.9% unexplained, including the post's emotional intensity, its single largest driver.

    Fact-check note

    The quoted reason is a paraphrase, not verbatim TikTok wording; now marked. Numbers pending.

    Site: Technical lens, Completeness; Poster: five tests scorecard

    Simulation

    S28, S20, S19

57. 57

    LIME's top reason held in 39 of 40 reruns and did not change across kernel widths, and SHAP's top three features did not change when the post was altered slightly. Stability has to be tested rather than assumed, because LIME's sampling makes every run slightly different.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, Stability; Poster: five tests scorecard

    Simulation

    S28, S20

58. 58

    Not tested with real users in this project. Judged against the intended audience, a SHAP value in probability units is not something a 14-year-old or a legislator can act on, and the only actionable counterfactual places the burden on the child rather than on the platform.

    Fact-check note

    Simulation output (S28); not fact-checked externally. Will be regenerated by the new simulation.

    Site: Technical lens, Human usefulness; Poster: five tests scorecard

    Simulation

    S28, S20, S21

59. 59

    Announced in December 2022 for rollout over the following weeks, it gives short reasons such as that a video is similar to content the user has interacted with or searched for, is from an account the user follows, or is popular or was posted recently in the user's region.

    Fact-check note

    It was announced, not confirmed launched, in December 2022; "nearby" is really "recently in your region". Present-day status of the feature not verified.

    Site: Public lens, TikTok

    Corrected

    S61, S19

60. 60

    Popular Science reported that TikTok's "Why this video" feature would not reveal any major details about how the algorithm works and would give users only a hint as to why a video was shown. TikTok itself presented the feature as a step toward "meaningful transparency" and said it would expand it to bring "more granularity and transparency."

    Fact-check note

    Misattribution: the "won't reveal major details / a hint" judgment is the Popular Science reporter's, not TikTok's.

    Site: Public lens, TikTok limits

    Corrected

    S19, S61

61. 61

    In June 2023, Meta released 22 system cards describing how AI ranks content across Facebook and Instagram, and said it would expand "Why Am I Seeing This?" to Instagram's Reels tab and Explore and to Facebook Reels in the coming weeks, letting users tap an individual reel to see how their previous activity may have informed the models that shape the reels they see.

    Fact-check note

    22 system cards and date correct. The expansion was announced ("in the coming weeks"), not shown as completed.

    Site: Public lens, Meta

    Corrected

    S18

62. 62

    Meta states that there is a limit to what it can disclose safely, including the signals it uses to keep bad content out of Feeds, because disclosing them might make it easier to circumvent its defenses, and it acknowledges that not everyone will find information just because it is published on its website.

    Fact-check note

    Both halves match Meta's post; optional rewording adds Meta's stated reason.

    Site: Public lens, Meta limits

    Supported

    S18

63. 63

    Under the EU Digital Services Act, very large platforms must offer at least one option for each of their recommender systems that is not based on profiling (Art. 38). Users must be able to switch to it at any time from the part of the interface where the feed is ranked (Art. 27(3); Recital 94).

    Fact-check note

    First check marked this partially supported; the adversarial pass showed Art. 38 builds on Art. 27(3) and Recital 94, so the "same interface" point holds. Source S10 replaced with EUR-Lex.

    Site: Public lens, EU non-profiling option

    Supported

    S10, S9

64. 64

    Researchers have cited a civil society complaint alleging that Meta is not meaningfully implementing these recommender requirements on Facebook and Instagram. In a separate case brought by Bits of Freedom, a Dutch court found in October 2025 that Meta was not complying with DSA Articles 27(3) and 38, and the Court of Appeal upheld that ruling in March 2026.

    Fact-check note

    Matches S11. The complaint (Apr. 2025, Ireland) targets the non-profiled feed option. Added the separate Dutch court rulings (reported by DSA Observatory, July 2026; Meta's cassation appeal is pending). Do not attribute the Dutch ruling to the Irish complaint.

    Site: Public lens, EU limits

    Supported

    S11, S62, S63

65. 65

    In the simulation, the reason a platform would most likely show (paraphrasing TikTok's "Why this video" feature), "This video is similar to content you've interacted with," is accurate, but it accounts for only 35.1% of why the top post was ranked first. It does not tell the user that the post's emotional intensity was the largest single factor, and the only change it implies she can make, engaging less with similar content, would not remove the post from the top of her feed.

    Fact-check note

    Same paraphrase fix as 56. Numbers pending.

    Site: Public lens, gap; Poster: Public quadrant

    Simulation

    S28, S19

66. 66

    In March 2026, two juries reached verdicts against Meta within a day of each other: a Los Angeles jury found Meta and YouTube negligent in a case focused on design features such as infinite scroll and autoplay, and a New Mexico jury imposed a $375 million penalty in a case that included algorithms prioritizing sensational or harmful content.

    Fact-check note

    All elements confirmed (NM verdict Mar. 24, LA verdict Mar. 25). Optional context: appeals, and NM's $567M abatement order.

    Site: Public lens, stakes; Poster: Public quadrant

    Supported

    S16, S17

67. 67

    These dynamics are not limited to young users: Amnesty International found that Facebook's engagement-driven news feed, ranking and recommendation algorithms amplified anti-Rohingya hatred in Myanmar in the years before the 2017 atrocities (which the US government has determined to be genocide), and that Meta's later fixes, such as banning military leaders' accounts and adding Burmese-language moderators, focused on content moderation rather than on changing those algorithms.

    Fact-check note

    "Left the ranking algorithm untouched" overstated: Meta did demote some content. Amnesty's point is the engagement core was not changed. The genocide label is attributed to the US determination; the ICJ case had no judgment as of Oct. 2026.

    Site: Public lens, Myanmar line

    Corrected

    S64, S23, S24

68. A1

    In its final judgment (Aug. 2026), the New Mexico court ordered changes for under-18 accounts: no push notifications from 10 PM to 7 AM or during school hours (para. 168), like counts hidden by default (para. 169) and a cap of 90 hours of use a month (para. 170). It granted no relief on the design and implementation of Meta's algorithms (para. 173) and declined to order changes to infinite scroll and autoplay, which it said "require legislative or executive branch action" (para. 160), citing the risk of competitively damaging Meta when such features are industry-wide, and First Amendment and Section 230 concerns (paras. 160, 163–166).

    Site: Policy lens, New Mexico row and pull quote

    Added later

    S59

69. A2

    In eSafety's three-month evaluation, the share of under-16s holding an age-restricted account fell from 52.4% to 42.1%; platforms' failure to implement effective age assurance was cited as the main reason children kept accounts.

    Site: Policy lens, Australia row and age-check line

    Added later

    S71, S35, S72

70. A3

    In an Internet Matters survey of 1,270 UK children aged 9–16 (fieldwork Sept.–Oct. 2025, published May 2026), 32% said they had bypassed an age check in the past two months; the most common methods were a fake birthdate (13%) and someone else's login (9%).

    Site: Policy lens, age-check line

    Added later

    S73

71. A4

    The four principles come from Principles of Biomedical Ethics by Tom Beauchamp and James Childress, first published in 1979 and in its ninth edition (2026); Floridi and Cowls (2019) adopted the same four for AI ethics and added a fifth, explicability.

    Site: Philosophical lens intro

    Added later

    S74, S75, S76