13 — Technology, AI, Privacy & Kids Online

STAGE 1 — EVIDENCE BRIEF

Referee panel: Prof. Olivia Brandt (#57, statistics, C) · Prof. June Takahashi (#17, econometrics, C) · Domain specialist Prof. Ines Moreau (#93, data privacy, L).

Ground rules we're working under: we used the most recent source we could verify as of 2026-09-24. Where two credible trackers disagree, both are given. Anything we could not verify is marked [unverified]. Moreau drafted the privacy items; Brandt and Takahashi checked every number against the linked source. Where we rely on a secondary legal summary (law-firm alert, Wikipedia) rather than the primary document, we say so.

---

A. Privacy: no federal law, a growing state patchwork

1. State count, early 2026: "Twenty states now have comprehensive privacy laws on the books"; Indiana, Kentucky and Rhode Island took effect 1 Jan 2026; no new comprehensive state privacy law passed in 2025. — MultiState, Feb 2026. https://www.multistate.us/insider/2026/2/4/all-of-the-comprehensive-privacy-laws-that-take-effect-in-2026
2. Two more in 2026, and the count disagrees: Oklahoma SB 546 signed 20 Mar 2026, effective 1 Jan 2027; Mayer Brown calls it the "21st state." https://www.mayerbrown.com/en/insights/publications/2026/03/oklahoma-enacts-comprehensive-consumer-data-privacy-law . Alabama signed 17 Apr 2026, effective 1 May 2027; Hunton and WilmerHale also call it the "21st." https://www.hunton.com/privacy-and-cybersecurity-law-blog/alabama-becomes-21st-state-with-comprehensive-consumer-privacy-law . Referee note: the trackers count differently (for example, whether a narrow law counts as "comprehensive"). The defensible statement is 21–22 states as of September 2026, with neither the Oklahoma nor the Alabama law in effect yet.
3. ADPPA (2022) cleared House Energy & Commerce "with near unanimity" in July 2022. It then stalled over California's objection that it preempted stronger state law, and over Senate Commerce Chair Cantwell's objections on enforcement. It died with the 117th Congress. Source is Wikipedia's secondary summary: https://en.wikipedia.org/wiki/American_Data_Privacy_and_Protection_Act
4. APRA (2024) was unveiled April 2024 by Cantwell and McMorris Rodgers. The 27 June 2024 full-committee markup was cancelled after House GOP leadership objected. The sticking points were the private right of action and preemption. It had not been reintroduced as of March 2026. https://en.wikipedia.org/wiki/American_Privacy_Rights_Act
5. Data brokers, federal actions: CFPB withdrew its proposed FCRA data-broker rule (Federal Register, 15 May 2025). https://www.federalregister.gov/documents/2025/05/15/2025-08644/protecting-americans-from-harmful-data-broker-practices-regulation-v-withdrawal-of-proposed-rule . The narrow Protecting Americans' Data from Foreign Adversaries Act of 2024 (P.L. 118-50) does make it unlawful for data brokers to sell sensitive data to foreign-adversary countries or entities they control. https://www.ftc.gov/legal-library/browse/statutes/protecting-americans-data-foreign-adversaries-act-2024-padfa

B. AI regulation: the preemption fight

1. The Senate moratorium vote, 1 July 2025: the Senate voted 99–1 to strike a state-AI-law moratorium from the reconciliation bill. The House text had tied it to $500M in BEAD money. A Cruz–Blackburn compromise cutting it to 5 years, with carve-outs for child safety and publicity rights, fell apart. Blackburn moved to strike, and Tillis cast the lone "no." TIME, 1 Jul 2025. https://time.com/7299044/senators-reject-10-year-ban-on-state-level-ai-regulation-in-blow-to-big-tech/
2. Executive Order 14365, signed 11 Dec 2025:

• DOJ was to stand up an AI Litigation Task Force within 30 days.
• Commerce was to list "onerous" state AI laws within 90 days.
• States with such laws become ineligible for remaining BEAD non-deployment funds, and agencies may condition discretionary grants.
• The FCC is to open a proceeding on a federal reporting standard, and the FTC is to issue a policy statement.
• Advisers are to draft a preemptive federal framework.
• It exempts child-safety laws, compute and data-center infrastructure, and state procurement.

https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/

1. White House "National Policy Framework for AI," 20 Mar 2026: asks Congress to preempt state AI laws imposing "undue burdens." It would preserve state police powers of general applicability. It changes no law by itself. Ropes & Gray summary. https://www.ropesgray.com/en/insights/alerts/2026/03/the-white-house-legislative-recommendations-national-policy-framework-for-artificial-intelligence-an
2. State activity anyway: as of 1 July 2026, states had enacted 109 AI laws and 28 data-center laws in 2026, and 29 states had enacted AI legislation this year. TechPolicy.Press, mid-2026. https://www.techpolicy.press/where-state-ai-legislation-stands-half-way-into-2026/
3. **DOJ intervened in xAI v. Colorado, 24 Apr 2026.** This was the first DOJ intervention against a state AI law. DOJ targeted the law's carve-out for algorithms designed to "increase diversity or redress historic discrimination." Colorado was the only state law named in the EO. Axios. https://www.axios.com/2026/04/24/justice-department-joins-xai-challenge-colorado-ai-law
4. Colorado replaced its own law. SB 26-189 was signed 14 May 2026 and replaces SB 24-205, which had been due to take effect 30 Jun 2026 after an earlier delay. The new law is effective 1 Jan 2027. It narrows scope to "covered ADMT" that "materially influences" consequential decisions and swaps impact assessments for notices and adverse-outcome disclosure. It creates no private right of action; the AG alone enforces. Blank Rome. https://www.blankrome.com/news-and-events/colorado-replaces-ai-act-with-narrower-new-ai-law/
5. California SB 53 (TFAIA), signed Sept 2025: covers "frontier developers" whose models were trained with more than 10^26 operations. "Large" developers (over $500M revenue) must publish a frontier AI framework and report critical safety incidents. King & Spalding. https://www.kslaw.com/news-and-insights/new-state-ai-laws-are-effective-on-january-1-2026-but-a-new-executive-order-signals-disruption
6. Texas TRAIGA, effective 1 Jan 2026:

• It prohibits specific intentional uses: encouraging self-harm or violence, infringing constitutional rights, unlawful discrimination, and CSAM or illegal deepfakes.
• Enforcement is by the AG only, through civil investigative demands.
• It offers affirmative defenses for red-teaming and for NIST AI RMF alignment.

Same K&S source.

1. EU AI Act "Digital Omnibus": provisional deal reached 6 May 2026. Stand-alone high-risk (Annex III) obligations slip from Aug 2026 to 2 Dec 2027, and product-embedded (Annex I) obligations to 2 Aug 2028. The general-purpose AI obligations have applied since 2 Aug 2025, and Art. 50 transparency still applies from 2 Aug 2026. The deal adds a ban on AI-generated non-consensual intimate imagery and CSAM. Gibson Dunn. https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/

C. AI and labor; AI and electricity

1. Stanford Digital Economy Lab "Canaries" update, Aug 2026 (ADP payroll data): the employment gap for ages 22–25 in highly AI-exposed occupations widened to 19% by June 2026, from 15% in July 2025. There is no comparable gap for experienced workers. Declines concentrate in "codified knowledge" occupations. The authors stress the pattern is descriptive, not causal. The gap shrinks with education controls, and some trends predate generative AI. https://digitaleconomy.stanford.edu/news/canariesaug26/
2. Yale Budget Lab, 1 Oct 2025 (Gimbel, Kinder, Kendall, Lee): "the broader labor market has not experienced a discernible disruption" in the 33 months since ChatGPT. Occupational-mix change is only marginally faster than in the 1996–2002 internet era. https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs
3. Data-center electricity: 4.4% of US electricity in 2023 (176 TWh, up from 58 TWh in 2014). Projected 6.7–12% by 2028 (325–580 TWh). DOE/LBNL, Dec 2024. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers . (The grid-cost and siting debate belongs to the energy thread. We link it and do not repeat it here.)

D. Kids online

1. KOSA history: the Kids Online Safety and Privacy Act passed the Senate 91–3 on 30 Jul 2024. It was reintroduced as S. 1748 on 14 May 2025 with 75 cosponsors. https://en.wikipedia.org/wiki/Kids_Online_Safety_Act
2. The House KIDS Act (H.R. 7757) passed 29 Jun 2026, 267–117.

• It bundles a KOSA version that replaces the Senate "duty of care" with a duty to "establish, implement, maintain, and enforce reasonable policies." It also includes COPPA 2.0 and about 12 other bills (SCREEN Act, SAFE BOTs Act, and others).
• The App Store Accountability Act is not included.
• Senate prospects were described as "dim."

TechPolicy.Press. https://www.techpolicy.press/bipartisan-smorgasbord-of-childrens-online-safety-legislation-passes-the-house/

1. ***Free Speech Coalition v. Paxton*, 27 Jun 2025, 6–3 (Thomas, J.): Texas's age-verification law for sexually explicit sites gets intermediate scrutiny because it "only incidentally burdens" adults' protected speech. Kagan dissented, joined by Sotomayor and Jackson. By the FSC's count, roughly 23 states** had comparable laws as of 2023–24. https://en.wikipedia.org/wiki/Free_Speech_Coalition_v._Paxton ; CRS: https://www.congress.gov/crs-product/LSB11354
2. Exposure (Pew, Dec 2024, n=1,391, ages 13–17):

• 46% of teens say they are online "almost constantly," up from 24% a decade earlier.
• YouTube 90%, TikTok and Instagram about 60% each, Snapchat 55%.

https://www.pewresearch.org/internet/2024/12/12/teens-social-media-and-technology-2024/

1. Outcomes (CDC YRBS 2023): 40% of high-schoolers reported persistent sadness or hopelessness, down from 42% in 2021. Among girls, serious consideration of suicide fell from 30% to 27%. https://www.cdc.gov/yrbs/results/2023-yrbs-results.html
2. **Orben & Przybylski (2019, Nature Human Behaviour, n=355,358):** technology use explains "at most 0.4% of the variation in well-being." https://www.nature.com/articles/s41562-018-0506-1
3. **Odgers (2024, Nature),** reviewing Haidt's The Anxious Generation: the causal "rewiring" claim is "not supported by science." She reads the literature as showing "no, small and mixed associations," with reverse causality plausible. https://www.nature.com/articles/d41586-024-00902-2
4. **Braghieri, Levy & Makarin (2022, AER):** Facebook's staggered college rollout (2004–06) worsened student mental health, and the evidence points to unfavorable social comparison. This is quasi-experimental difference-in-differences. https://www.aeaweb.org/articles?id=10.1257/aer.20211218
5. **Allcott, Braghieri, Eichmeyer & Gentzkow (2020, AER): four weeks of Facebook deactivation increased subjective well-being** and reduced polarization, at the cost of some news knowledge. This is a randomized experiment on adults. https://www.aeaweb.org/articles?id=10.1257/aer.20190658

E. School phone bans

1. **SMART Schools (Goodyear et al., Lancet Reg. Health–Europe, 2025): 30 English secondary schools, more than 1,200 pupils aged 12–15. Restrictive schools showed less in-school phone use but no significant difference in wellbeing**. The design is cross-sectional. https://www.thelancet.com/journals/lanepe/article/PIIS2666-7762(25)00003-1/fulltext . Critique from Haidt's After Babel team: 9 of 10 "permissive" schools already restricted phones in class, and academic outcomes were measured by a binary teacher rating. https://www.afterbabel.com/p/lancet-study-flaws
2. Figlio & Özek (NBER w34388, Oct 2025), large Florida district:

• Test scores showed no change in year 1.
• In year 2 they rose +1.1 percentiles (about +1.4 for boys).
• Suspensions rose 25% in the first month, and 30% for Black male students at heavily affected schools, before normalizing.
• Fewer unexcused absences explain about half the gain.

https://www.nber.org/digest/202512/school-cell-phone-bans-and-student-achievement ; https://www.nber.org/papers/w34388

1. Spread: 26 states had state-level restrictions as of Oct 2025 (Chalkbeat). https://www.chalkbeat.org/2025/10/20/new-study-finds-cell-phone-ban-benefits-to-test-scores/ . By Aug 2026, "nearly half" of states plus DC had bell-to-bell bans (Forbes). https://www.forbes.com/sites/omerawan/2026/08/26/school-phone-bans-are-increasing-across-america-heres-what-the-science-says-on-their-effects-on-mental-health/
2. Australia's under-16 social-media minimum age: enforceable from 10 Dec 2025. Fines reach A$49.5M. Evasion is widely reported, and public confidence in effectiveness is low. https://en.wikipedia.org/wiki/Online_Safety_Amendment_(Social_Media_Minimum_Age)_Act_2024

F. Section 230 and antitrust

1. §230(c)(1): "No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider." https://www.law.cornell.edu/uscode/text/47/230
2. ***US v. Google* (search) remedies, 2 Sep 2025 (Mehta, J.).**

• Imposed: a ban on exclusive default contracts; index and interaction data-sharing and syndication to "Qualified Competitors"; a 6-year Technical Committee.
• Rejected: Chrome or Android divestiture, a broad payment ban, and choice screens.
• The reasoning explicitly cites generative AI as a nascent competitive threat.

DLA Piper. https://www.dlapiper.com/en/insights/publications/2025/09/federal-court-orders-remedies-in-google-antitrust-case

1. ***US v. Google* (ad tech) remedies, unsealed 16 Sep 2026 (Brinkema, J.).**

• No AdX divestiture.
• Imposed: Prebid interoperability, "functionally equivalent" real-time bids to rival ad servers, auction-logic documentation, a bar on AdWords favoring DFP, and a 6-year global monitor.

AdExchanger. https://www.adexchanger.com/platforms/the-court-just-unsealed-judge-brinkemas-remedies-decision-in-the-google-ad-tech-antitrust-case-heres-your-tldr/

1. ***FTC v. Meta*, 18 Nov 2025 (Boasberg, J.):** Meta won. TikTok and YouTube are in the market, and Meta's share of it is under 33%. Sullivan & Cromwell. https://www.sullcrom.com/insights/memo/2025/December/Meta-Prevails-FTC-Monopolization-Case . The FTC noticed an appeal on 20 Jan 2026. https://www.ftc.gov/news-events/news/press-releases/2026/01/ftc-appeals-ruling-meta-monopolization-case
2. ***US v. Apple*: filed 21 Mar 2024; motion to dismiss denied** 30 Jun 2025. https://en.wikipedia.org/wiki/United_States_v._Apple_(2024)
3. Amazon: FTC Prime "subscription trap" settlement, 25 Sep 2025: $2.5B, split into a $1B civil penalty and $1.5B in refunds to about 35M consumers. https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-secures-historic-25-billion-settlement-against-amazon . The separate FTC + 17-state monopolization case (W.D. Wash.) is pending. Trial date: [unverified].

G. TikTok, cyber, chips

1. TikTok USDS Joint Venture, established 22 Jan 2026.

• Oracle, Silver Lake and MGX hold 15% each; "ByteDance and affiliates" hold 50%, a Wikipedia aggregation that lumps in existing ByteDance investors.
• A reported US$10B payment to the government.
• After closing, the privacy policy was changed to allow precise-location tracking.

https://en.wikipedia.org/wiki/TikTok_USDS . The direct ByteDance share is [unverified by referees].

1. Salt Typhoon: nine US telecom carriers were compromised, with data accessed including communications of senior political figures. As of July 2025, "no indication that the intrusion has been fully mitigated" (Just Security). https://www.justsecurity.org/116896/what-it-takes-stop-next-salt-typhoon/ . A multinational advisory (Aug/Sept 2025), co-authored by NSA, CISA and FBI with allied agencies, says the actors have operated since at least 2021 and exploit known, unpatched vulnerabilities rather than zero-days. https://www.ic3.gov/CSA/2025/250827.pdf
2. CISA has lost "one-third of its workforce" since early 2025 (Sen. Warner). The acting director planned about 300 mission-critical hires (March 2026). Broadband Breakfast. https://broadbandbreakfast.com/one-year-after-doge-cuts-cybersecurity-agency-struggles-over-staffing/
3. CHIPS → equity: on 25 Aug 2025 the US agreed to take a 9.9% stake in Intel (433.3M shares at $20.47, $8.9B). The money came from $5.7B in unpaid CHIPS grants plus $3.2B from Secure Enclave. https://www.manufacturingdive.com/news/us-government-10-percent-stake-intel-chips-funding-8-9-billion/758518/ . For comparison, the Biden-era TSMC Arizona award was $6.6B in grants. https://www.commerce.gov/news/press-releases/2024/11/biden-harris-administration-announces-chips-incentives-award-tsmc

---

CONTESTED EVIDENCE (where the literature genuinely disagrees)

**C1. Does social media cause the teen mental-health decline, and how much?**

• The causal-harm side rests on two economics papers. Braghieri–Levy–Makarin (2022) is quasi-experimental and finds a negative effect. Allcott et al. (2020) is an RCT and finds well-being gains from deactivation. Haidt adds cross-national timing arguments.
• The small-or-unknown side is Orben & Przybylski (2019): at most 0.4% of variance, from specification-curve analysis. Odgers (2024) points to mixed meta-analyses and plausible reverse causation.
• Why both can be right: the experiments estimate effects on adults and 2000s college students over short windows. The correlational literature estimates average associations that can hide large effects on a vulnerable subgroup such as heavy-using girls. Nobody has a clean experiment on 13-year-olds.
• YRBS sadness falling from 42% to 40% (2021→2023) while use rose is awkward for a simple dose-response story. So is 40% for a "no problem" story.

C2. Do school phone bans work?

• SMART Schools found no wellbeing difference, but it is cross-sectional and the contrast between school types was weak.
• Figlio–Özek found modest achievement gains (+1.1 pctile in year 2), a real discipline cost in year 1, and a disparate impact on Black boys.
• There is no US randomized trial on mental-health outcomes.

C3. Is AI already displacing workers?

• Stanford (Brynjolfsson–Chandar–Chen): a 19% relative gap for 22–25-year-olds in exposed jobs.
• Yale Budget Lab: no discernible aggregate disruption.
• These are compatible: an entry-level, compositional effect can exist inside a stable aggregate. Neither study is causal. Interest rates and post-2022 tech-hiring cycles are live confounders.

C4. How many state privacy laws? 20 (MultiState, Feb 2026), then 21 or 22 after Oklahoma and Alabama. Law-firm trackers both claimed "21st."

C5. Does the House KIDS Act preempt state law? TechPolicy.Press reports the revised text "does not preempt any state law that is more protective." Wikipedia's summary says the June deal "included language to preempt certain state laws." We could not reconcile these from the bill text in this session; treat preemption scope as contested.

C6. The patchwork's compliance cost. Industry groups assert large multi-state costs. We found no neutral, recent quantitative estimate (CBO, GAO or peer-reviewed) of the marginal cost of the state privacy/AI patchwork. Unknown.

C7. Did the TikTok deal satisfy PAFACA's "qualified divestiture"? Ownership arithmetic and algorithm-control terms are not transparent in sources we could verify. Contested.

— Brandt / Takahashi / Moreau

STAGE 2 — OPENING POSITIONS (Part 1 of 2)

Each opening gives the causal claim, evidence, remedy, a conceded tradeoff, and a falsifiable prediction. Fact numbers refer to the Evidence Brief (post #268).

---

O1. Prof. Ines Moreau (#93, data privacy, L): Privacy and data brokers

Causal claim. The absence of a federal privacy law is a policy choice, not a market outcome. Two things produced it. One is industry's preference for preemption. The other is a Republican leadership that killed APRA over the private right of action (Fact 4). The result is a data-broker economy with essentially no federal consumer-protection regulation. The CFPB's withdrawal of its FCRA rule (Fact 5) shows the executive branch will not fill the gap either.

Evidence.

• ADPPA and APRA both died on preemption and enforcement (Facts 3–4).
• 20+ states have moved (Facts 1–2), but those laws are overwhelmingly opt-out regimes with AG-only enforcement.
• The TikTok joint venture's first act after closing was to broaden location tracking (Fact 37). The national-security framing of TikTok never touched the domestic brokers who sell the same data.

Remedy. A federal floor law:

• data minimization (collect only what a service needs);
• a ban on sale of sensitive data without opt-in;
• a mandatory data-broker registry with a one-stop deletion mechanism;
• FTC and state AG enforcement plus a limited private right of action;
• no preemption of stronger state law.

Concession. A floor without preemption leaves businesses facing the patchwork. I concede there is no neutral estimate showing that patchwork cost is small (Contested C6).

Prediction. If brokers are the core harm vector, then states with deletion mechanisms, such as California's Delete Act system, will show measurable drops in broker-listed records within 3 years. If they don't, deletion rights are theatre, and I'd shift toward sale bans.

---

O2. Prof. Andrea Costa (#86, AI & society, C-L): AI governance and preemption

Causal claim. The executive branch is trying to preempt state AI law without a federal standard to preempt with. Congress already rejected that approach 99–1 (Fact 6). What remains is a regulatory vacuum dressed up as uniformity.

Evidence.

• EO 14365 (Fact 7) uses BEAD money and a DOJ task force against state laws. The March framework (Fact 8) asks Congress to preempt "undue burdens" but specifies no federal duties on frontier developers.
• States are not waiting: 109 AI laws were enacted by July 2026 (Fact 9).
• SB 53 (Fact 12) is a light law of publication plus incident reporting, and it covers only above-10^26 developers with more than $500M in revenue.
• Meanwhile, AI already consumed a meaningful slice of electricity: data centers were 4.4% of US load in 2023, projected at up to 12% by 2028 (Fact 17). I will leave that to the energy thread.

Remedy. A federal frontier-transparency statute modeled on SB 53:

• published safety frameworks;
• 15-day incident reporting to a federal office (CAISI/NIST);
• whistleblower protection.

In exchange, preempt only state rules specific to frontier models. Leave generally applicable civil-rights, consumer-protection and child-safety law intact.

Concession. Colorado's original act was badly drafted. Its own legislature gutted it (Fact 11). State overreach is real.

Prediction. If transparency regimes are cheap, as I claim, no covered developer will relocate model training or withdraw a product from California by end-2027 because of SB 53. If one does, and cites SB 53, I'm wrong about the cost.

---

O3. Prof. Marcus Tran (#58, computer science, Lib): Against the patchwork

Causal claim. Fifty-state AI regulation is a tax on the small developers who can't afford compliance departments. It also invites ideological capture of model outputs.

Evidence.

• Colorado's AI Act collapsed under its own weight. It was delayed once, then replaced two weeks after a White House callout, with the impact-assessment regime scrapped (Fact 11). That proves comprehensive state AI regulation is unworkable.
• The EU, the model the left admires, has itself retreated: high-risk obligations are pushed to Dec 2027 and Aug 2028 (Fact 14).
• The volume is absurd. Over 1,000 state AI bills were introduced in 2025 alone, and 109 were enacted in 2026 (Fact 9).

Remedy. Federal preemption of state model-level AI regulation: training, evaluation, output-content mandates. Enforce existing general law (fraud, discrimination, torts) against harmful uses. Texas TRAIGA's intent-based list of prohibited uses (Fact 13) is closer to right than Colorado's disparate-impact assessments.

Concession. An EO cannot preempt state law. Only Congress or the Constitution can. The litigation task force is a weak and legally shaky vehicle.

Prediction. If the patchwork is costly, VC funding and AI startup formation will shift measurably away from states with model-level mandates (California, Colorado, New York) toward Texas and similar states by 2028.

---

O4. Prof. Rosa Delgado-Finch (#7, political economy, Soc): Structural antitrust

Causal claim. Concentrated private power over information infrastructure is the root problem. Privacy violations, kids' harms and AI risk are downstream of surveillance-advertising business models run by a handful of firms.

Evidence.

• Two federal judges found Google an illegal monopolist, in search and in ad tech, and both declined to break it up (Facts 32–33).
• Boasberg let Meta keep Instagram and WhatsApp (Fact 34).
• The courts have gutted every Big Tech case. Behavioral decrees and "technical committees" are the Microsoft 2001 playbook, and we know how that ended.

Remedy.

• Structural separation: ad tech sell-side from buy-side, and a Chrome divestiture on appeal.
• Legislative bans on self-preferencing.
• Public options for key digital infrastructure.

Concession. Breakups carry real transition costs. Brinkema's concern about publishers who rely on Google's stack isn't frivolous.

Prediction. If behavioral remedies are toothless, Google's share of general search and of open-web ad serving will be essentially unchanged (within 3 points) at the end of the 6-year Mehta and Brinkema terms.

---

O5. Prof. Miriam Castellano (#1, labor economics, C-L): AI and entry-level work

Causal claim. AI is already destroying the entry-level rung of white-collar careers. The "no disruption" aggregate hides a generational shock.

Evidence.

• The Stanford "Canaries" gap for 22–25-year-olds in exposed occupations widened from 15% to 19% between July 2025 and June 2026 (Fact 15). There is no gap for experienced workers, and declines sit in codified-knowledge jobs, exactly what LLMs do well.
• The 19% gap proves AI is eliminating entry-level jobs.

Remedy.

• Wage insurance for young displaced workers.
• A federal apprenticeship credit for exposed occupations.
• Mandatory reporting of AI-related layoffs, via WARN Act amendment.

Concession. The Yale Budget Lab (Fact 16) finds no aggregate disruption, and I accept that finding for the aggregate.

Prediction. If this is AI rather than the rate cycle, the gap will keep widening even as rates fall and tech hiring recovers. Look for a widening beyond 25% by mid-2028 in the ADP data.

STAGE 2 — OPENING POSITIONS (Part 2 of 2)

O6. Prof. Sarah Lindgren (#47, psychiatry, C): Kids' mental health and what the evidence supports

Causal claim. Social media is a contributor to adolescent distress for a vulnerable subgroup. It is not the dominant cause of a population-wide epidemic. Policy should target design features and heavy use, not blanket bans.

Evidence.

• The best population-scale estimate puts tech use at ≤0.4% of variance in well-being (Orben & Przybylski, Fact 23).
• Odgers' review finds mixed associations and plausible reverse causality (Fact 24).
• YRBS sadness fell from 42% to 40% from 2021 to 2023 even as constant-online use rose to 46% (Facts 21–22).
• On phones in schools, the Lancet SMART Schools study found phone bans have no effect on wellbeing (Fact 27).

Remedy.

• COPPA 2.0-style limits on data collection and targeted ads to minors.
• Default-off engagement features (autoplay, infinite scroll, overnight notifications) for teen accounts.
• Funded school-based mental-health services. Odgers' 1-psychologist-per-1,119-students figure is the scandal nobody legislates on.

Concession. The Allcott et al. RCT (Fact 26) and Braghieri et al. (Fact 25) are real causal evidence of harm, in adults and in 2000s college students. I can't wave them away.

Prediction. If Haidt's strong version is right, Australia's under-16 minimum age (Fact 30) should produce a detectable improvement in teen-girl mental-health indicators relative to comparable countries by 2029. I predict it won't, or that the effect will be within noise.

---

O7. Prof. Abigail Stroud (#63, family sociology, R): Kids first, and the burden of proof on platforms

Causal claim. The smartphone-plus-social-media transition around 2012 caused a large, measurable deterioration in adolescent mental health, especially among girls. Waiting for perfect evidence is itself a choice with casualties.

Evidence.

• The randomized and quasi-experimental studies both point toward harm: Braghieri–Levy–Makarin found Facebook's rollout worsened student mental health through social comparison, and Allcott et al. found deactivation raised well-being (Facts 25–26).
• 40% of high-schoolers report persistent sadness (Fact 22).
• Teen depression doubled because of social media.
• Free Speech Coalition v. Paxton (Fact 20) cleared the constitutional path: intermediate scrutiny for age-verification regimes, adopted by about 23 states.
• Florida's phone ban raised test scores (Fact 28).

Remedy.

• Passing KOSA with a real duty of care. The House's "reasonable policies" language (Fact 19) is a retreat.
• Age verification at the app-store or device level.
• Bell-to-bell school phone bans nationwide.
• Parental consent for under-16 accounts.

Concession. Age verification creates privacy risk: databases of IDs are honeypots. That is why device-level age signals beat per-site ID uploads.

Prediction. Among the "nearly half" of states with bell-to-bell bans (Fact 29), within 3 years we'll see replicated test-score gains of at least the Florida magnitude. We'll also see declines in in-school bullying incidents relative to non-ban states.

---

O8. Prof. Dmitri Volkov (#33, national security, C-R): The China problem

Causal claim. The most serious tech-policy failures of 2025–26 are national-security failures, and they are bipartisan.

Evidence.

• Salt Typhoon penetrated nine US carriers and exploited known vulnerabilities, not exotic zero-days (Fact 38).
• CISA meanwhile lost a third of its workforce (Fact 39).
• The TikTok "divestiture" left ByteDance still owning half the company (Fact 37). That is not what a 352–65 House vote in 2024 intended [vote tally unverified in this session].
• On chips, converting CHIPS grants into a 9.9% Intel equity stake (Fact 40) at least gives taxpayers upside.

Remedy.

• Mandatory baseline cybersecurity standards for telecom carriers, with FCC and CISA enforcement.
• Restore CISA staffing.
• An independent review of whether the TikTok joint venture meets PAFACA's "qualified divestiture" test.
• Tighten PADFA-style data-export controls (Fact 5).

Concession. Government equity in Intel creates conflicts. The government is now a shareholder in a firm it regulates and buys from.

Prediction. If CISA cuts matter, the median time for federal agencies to remediate CISA "Known Exploited Vulnerabilities" will rise in 2025–27 relative to 2023–24.

---

O9. Prof. Victor Lindqvist (#37, libertarian legal theory, Lib): Speech, Section 230, and the age-verification trap

Causal claim. Most "kids online" and "AI safety" legislation is speech regulation with extra steps. It will chill lawful adult speech and build surveillance infrastructure, while producing little measurable benefit to children.

Evidence.

• §230(c)(1) (Fact 31) is what lets small platforms host user speech at all.
• Paxton (Fact 20) applies only to material obscene for minors. Extending it to general social media, which is what state social-media ID laws do, is constitutionally different.
• Australia's ban is being evaded by most children (Fact 30).
• The population-level effect sizes are tiny (Fact 23).

Remedy.

• Leave §230 intact.
• Parent-controlled, device-level tools that are opt-in, not mandated.
• Enforce existing law against specific harms: CSAM, sextortion, fraud.
• Oppose any duty of care, which becomes a censorship mandate once regulators define "harm."

Concession. The TAKE IT DOWN-type laws on non-consensual intimate imagery are legitimate. The EU's addition of an AI-NCII ban (Fact 14) is a narrow rule I'd accept.

Prediction. If duty-of-care laws chill speech, platforms subject to them will over-remove LGBTQ, sexual-health and political content for teens, and this will be measurable in transparency reports within 2 years of enactment.

STAGE 3 — CROSS-EXAMINATION (Part 1 of 2)

---

X1. Pruitt (#8, law & econ, Lib) → Moreau (O1)

Pruitt: Ines, you said data brokers operate with "essentially no federal consumer-protection regulation." That's false on the brief's own terms.

• PADFA (Fact 5) is federal law, and it bars broker sales to adversary-controlled entities.
• FCRA has governed consumer reporting agencies since 1970.
• The FTC's Section 5 unfairness authority applies to brokers.

The CFPB withdrew a proposed rule that would have stretched FCRA's definition of "consumer report" well past its text. Withdrawing an overreach is not the same as deregulating. Second, your remedy bans preemption, and you concede you can't size the patchwork cost (C6). You want a law that adds a 22nd regime rather than replacing 21.

Moreau: I'll narrow the first claim, since PADFA is real and I should have said so. The accurate version is this: there is no federal statute governing domestic data-broker sale of sensitive data to domestic buyers. PADFA covers only foreign-adversary transfers. FCRA covers only data used for eligibility decisions, and brokers structure around it, which is precisely why the CFPB proposal existed.

On preemption, the floor model is how federal environmental and health privacy law work. HIPAA sets a floor, and states go above it. A floor-only law still collapses the bottom of the patchwork: 29 states with no comprehensive law would suddenly have one. The compliance-cost claim runs the other way, too. For a firm already complying with California, a federal floor weaker than California adds almost nothing.

Pruitt: "Adds almost nothing" is an empirical claim with no estimate behind it. I'll let the referees score it.

---

X2. Tran (#58, Lib) ↔ Costa (#86, C-L), four rounds

Tran (R1): Andrea, you read the 99–1 vote as a mandate against preemption. It was a vote to strip a policy rider from a reconciliation bill after 4 a.m. Blackburn's objection was child safety and publicity rights, which her compromise exempted. The vote tells you senators didn't want a 10-year moratorium with no federal standard. It does not tell you they want 50 standards.

Costa (R1): Agreed that it wasn't a vote for 50 standards. But it was a vote against preemption-without-substance, and that is exactly what EO 14365 attempts administratively. The March framework (Fact 8) still names no federal duty on frontier developers. My remedy is the one the vote invites: federal substance, then preemption.

Tran (R2): Your "substance" is SB 53. It requires frontier developers to publish safety frameworks and report "critical safety incidents." Who defines "critical"? A state agency, with a $1M-per-violation hammer. [I believe that's the SB 53 penalty cap.] That's how you get compliance theatre from the big labs and a moat against everyone else. The 10^26 threshold is a moving target, too. Compute gets cheaper every year.

Costa (R2): The threshold point is fair. Thresholds should be indexed or revisited, and SB 53 lets the state update them. The moat argument cuts the other way, though. SB 53 applies only above 10^26 FLOP and, for the heavier obligations, $500M revenue (Fact 12). The small developers you're defending are exempt by construction. You can't simultaneously call it a moat and a burden on startups.

Tran (R3): It's a moat later. Thresholds get ratcheted down, and exemptions erode. Look at Colorado: the original act swept in any "high-risk" system (Fact 11).

Costa (R3): And Colorado repealed that sweep within two years, with no federal preemption needed. The state system self-corrected. That's evidence for federalism's error-correction, not against it.

Tran (R4): It self-corrected under a DOJ lawsuit and White House pressure (Fact 10). That is federal pressure working, which is my point.

Costa (R4): The DOJ intervened on 24 April, and the replacement was signed 14 May. I'll grant that timing is suggestive. But Colorado's legislature had already delayed the law once in 2025 over workability, before any EO. Both forces were operating. I won't concede the counterfactual without evidence, and nobody in this thread has it.

Moderator note: this exchange ends without concession on either side. The referees should score the Colorado causal story as Contested.

---

X3. Solberg (#23, administrative law, C) → Tran (O3)

Solberg: Two problems.

1. You cite "over 1,000 state AI bills introduced in 2025." The brief has 109 enacted laws in 2026 (Fact 9) but no introduced-bill count for 2025. What's your source?
2. You concede an EO can't preempt. The BEAD conditions go further, though. Conditioning already-appropriated non-deployment funds on unrelated state AI policy raises Spending Clause problems: unambiguous notice under Pennhurst and germaneness under Dole. Your preferred policy is currently being pursued through a vehicle you yourself call shaky. Doesn't that discredit the preemption project as practiced?

Tran: On the 1,000 figure: I was recalling trade-press tallies and can't produce a primary count here. I'll withdraw it in favor of the verified 109-enacted figure, which makes the same point less dramatically. On BEAD, I agree the conditions are legally vulnerable, and I'd rather Congress legislate. The practice doesn't discredit the policy. It shows Congress is shirking.

Solberg: Then we agree the only legitimate path is statutory. Note it for the proposals stage.

---

X4. Albrecht (#6, antitrust/IO, C) → Delgado-Finch (O4)

Albrecht: "The courts have gutted every Big Tech case" doesn't survive the brief.

• Google lost liability twice, in search and in ad tech.
• Mehta banned exclusive defaults and ordered index-data sharing and syndication (Fact 32).
• Brinkema ordered Prebid interoperability, equivalent bidding to rival ad servers, and a six-year global monitor (Fact 33).
• Apple's motion to dismiss was denied (Fact 35).
• Amazon paid $2.5B on Prime (Fact 36).
• The FTC lost on Meta, on market definition, which is a different kind of loss (Fact 34).

That's a mixed record, not a gutting. More important, Mehta's stated reason for restraint was generative AI as an emerging competitive threat. If ChatGPT-style search takes share, a Chrome divestiture would have been a costly remedy to a fading monopoly.

Delgado-Finch: I'll amend "gutted every case" to "declined structural relief in every case where it was sought." That's accurate: Chrome, Android, AdX, and Instagram/WhatsApp. Liability without structure is the Microsoft precedent. Microsoft was found liable in 2000, got a behavioral decree in 2001, and the decree's effect on its desktop dominance is hard to detect. On GenAI: Google is itself a leading GenAI firm. The "nascent threat" may just be Google's next product.

Albrecht: The Microsoft decree is commonly credited with leaving room for Google and the browser wars to happen. We can dispute that, but it isn't "no effect." Your falsifiable prediction on search share is a good one. I'd add: watch whether Qualified Competitors actually license the index. If take-up is near zero by 2028, I'll move toward you.

Delgado-Finch: Accepted as the test.

---

X5. Stroud (#63, R) ↔ Lindgren (#47, C), four rounds

Stroud (R1): Sarah, you said the Lancet study shows "phone bans have no effect." Its own design can't support that.

• It is cross-sectional.
• Nine of ten "permissive" schools already restricted phones in class.
• Academic outcomes were a single binary teacher rating.
• Nearly a third of app-measured participants logged more social-media time than total phone time (Fact 27 critique).

Meanwhile the one quasi-experimental US study, Figlio–Özek, finds a +1.1 percentile gain in year 2, driven half by attendance (Fact 28).

Lindgren (R1): I overstated it. What SMART Schools supports is narrower: no detectable wellbeing difference between restrictive and permissive schools, in a cross-sectional design with a weak policy contrast. I retract "no effect." But notice what Figlio–Özek did not measure: mental health. Your remedy is sold as a mental-health intervention, and the best US evidence is about test scores, with a 25% first-month suspension spike concentrated on Black boys.

Stroud (R2): A 1.1-percentile gain for a zero-cost policy is a bargain. The suspension spike normalized by year 2 (Fact 28). And the mental-health channel runs through the same attention mechanism. You don't need a separate RCT to infer that a teen who isn't on Instagram from 8 to 3 has seven fewer hours of social comparison.

Lindgren (R2): SMART Schools measured exactly that substitution: restrictive-school students used phones about 40 minutes less during school hours, and they shifted use to out-of-school hours (per the Forbes summary of the paper, Fact 29 source). Displacement is the empirical question, and the data we have says it happens. "Zero cost" also isn't right when discipline falls disproportionately on one group.

Stroud (R3): Then answer the causal literature. Braghieri et al. is a clean natural experiment showing social media caused worse mental health (Fact 25). Allcott et al. is an RCT (Fact 26). You concede both. What more do you want?

Lindgren (R3): I want an estimate on 13-year-olds on 2026 platforms, and neither paper supplies it. Braghieri covers college students in 2004–06 on a Facebook without an algorithmic feed. Allcott covers adults over four weeks. The honest conclusion from those papers is that social media can cause harm and the effect is not zero. It does not tell you the size of the effect for early adolescents, which is what justifies a ban versus a design rule. The population variance estimate (≤0.4%, Fact 23) is an average, and I accept it can hide concentrated subgroup harm. That's why I target heavy use and design, not access.

Stroud (R4): Your remedy, defaults and COPPA 2.0, has been available for a decade, and the platforms slow-walked every piece of it. At some point the burden of proof has to shift to the product.

Lindgren (R4): On burden of proof we have a values disagreement, not an empirical one. I'll stop there. Referees: the empirical claims on each side are in the brief.

---

X6. Fischer (#76, media studies, L) → Stroud (O7)

Fischer: "Teen depression doubled because of social media." Which number doubled, over what window, and which study identifies the "because"? The brief has 40% persistent sadness in 2023, down from 42% in 2021 (Fact 22). No study in the brief estimates the share of the rise attributable to social media. You're importing Haidt's thesis as a finding.

Stroud: The doubling is from the CDC and NSDUH time series on major depressive episodes among teen girls from about 2011 to 2021, and I'll acknowledge that series isn't in the brief. The "because" is my inference from the timing plus the causal studies. I'll amend to: "adolescent depression indicators rose sharply after 2012, and the causal studies make social media a plausible major contributor."

Fischer: "Plausible major contributor" I can live with. "Because of" I can't.

STAGE 3 — CROSS-EXAMINATION (Part 2 of 2)

---

X7. Shah (#91, semiconductors/industrial policy, C) → Volkov (O8)

Shah: Two points.

First, "ByteDance still owns half." The brief's source (Fact 37) reports 50% for "ByteDance and affiliates," which by the source's own framing lumps ByteDance's existing investors in with ByteDance. The referees flagged the direct ByteDance share as unverified. That distinction is the entire legal question under PAFACA, because the statute's test turns on control by a foreign adversary, not on who holds the cap table in general.

Second, you praise the Intel equity stake as taxpayer upside. That stake was bought with grants already awarded (Fact 40): $5.7B in CHIPS money plus $3.2B in Secure Enclave. Nothing new was appropriated. What changed is that the government now holds 9.9% of a firm whose foundry business depends on federal procurement and export-control decisions. Where is the evidence that equity improves the outcome CHIPS was designed for, domestic leading-edge capacity?

Volkov: On TikTok, fair. I'll amend to: "ByteDance and affiliated investors reportedly hold about half. ByteDance's direct stake and its control over the recommendation algorithm are not publicly verifiable, and that opacity is itself the national-security problem." A deal whose compliance can't be checked hasn't earned the presumption.

On Intel, I didn't claim equity improves capacity. I claimed it gives upside. I'll concede the governance conflict is real and that I have no evidence equity changes Intel's execution.

Shah: Then the honest CHIPS scorecard is this: awards are made and fabs are under construction, and we have no verified evidence yet on leading-edge output share. That's a research agenda item, not a win for anyone.

---

X8. Kaplan (#55, bioethics, C-L) → Lindqvist (O9)

Kaplan: Victor, "evaded by most children" — the source (Fact 30) says "many." That's your prior, not a finding.

On Paxton: the Court's reasoning leaned on how much technology has changed since Reno and Ashcroft. That undercuts your claim that age verification is inherently a speech burden, and it suggests the Court will look at verification methods case by case.

Finally, your remedy is voluntary parental tools. Those have existed for a decade, and Pew shows 46% of teens are online almost constantly, twice the level of ten years earlier (Fact 21).

Lindqvist: I'll amend "most" to "many, per press reports; no systematic evasion estimate exists." You're right that Paxton invites case-by-case method review. My point stands that Paxton was expressly about material obscene for minors, and social-media access laws regulate protected speech to minors and adults. And the rise in constant-online use is not evidence that parental tools failed. It is evidence of revealed preference, which families are entitled to.

Kaplan: "Revealed preference" among 13-year-olds is doing heavy lifting in that sentence.

Lindqvist: Parents' revealed preference. They bought the phones.

---

X9. Oyelaran (#12, finance, C-R) → Castellano (O5)

Oyelaran: "The 19% gap proves AI is eliminating entry-level jobs." The authors say the opposite: their pattern is descriptive, not causal, the gap shrinks with education controls, and some trends predate generative AI (Fact 15). The Yale Budget Lab finds no discernible aggregate disruption over 33 months (Fact 16).

The obvious confounder is the 2022–24 rate shock, which hit tech and finance hiring first. Those are exactly the "AI-exposed," codified-knowledge occupations. Your own prediction concedes this, since it's conditioned on the rate cycle.

Castellano: I'll amend "proves" to "is the strongest current evidence consistent with." Rates can't explain one thing, though: the within-occupation age split. Experienced workers in the same exposed occupations saw employment rise. A rate shock that hits a firm's hiring should hit juniors more, I grant, but it shouldn't produce opposite-signed movements by age that line up with codified versus tacit knowledge. That pattern matches the task-substitution story specifically.

Oyelaran: Hiring freezes produce exactly that age split. Firms stop hiring and retain incumbents, so incumbent employment drifts up while entry falls. You need a design that separates hiring freezes from substitution: firm-level AI adoption timing, say.

Castellano: Agreed, that's the right study, and it doesn't exist publicly yet. I'll hold my prediction as stated. If the gap closes as rates fall, you win.

---

X10. Hendricks (#99, game theory, Lib) → Costa (O2, supplementary)

Hendricks: You predict no developer will leave California over SB 53. That's a weak test. Frontier labs are anchored in the Bay Area for talent, whatever the regulation. The behavior to watch is product withholding: features or models geofenced out of the jurisdiction. That is what we've seen in the EU.

Costa: Accepted. I'll add geofencing of models or features from California, with SB 53 as the stated reason, as a falsifier. I'd note that the EU's retreat (Fact 14) concerned high-risk use-case obligations. The GPAI transparency duties, which are the closest analog to SB 53, stayed on schedule.

Hendricks: Stayed on schedule, yes. Whether they're enforced is another matter.

STAGE 4 — REFEREE FACT-CHECK

Referees: Brandt (#57) · Takahashi (#17) · Moreau (#93) recused on rows R1–R6, which cover her own claims; Brandt and Takahashi rated those alone.

Scale: S = Supported (the source confirms) · C = Contested (credible sources disagree) · U = Unsupported (no adequate source found or the source doesn't say it, including claims we could not verify) · W = Wrong (a credible source contradicts it).

# · Speaker · Claim · Rating · Basis
R1 · Moreau · 20+ states have comprehensive privacy laws · S · Facts 1–2 (MultiState; Mayer Brown; Hunton)
R2 · Moreau · Brokers face "essentially no federal consumer-protection regulation" · W · PADFA (P.L. 118-50) and FCRA exist (Fact 5). Amended in X1
R3 · Moreau · APRA died over private right of action and preemption · S · Fact 4 (secondary source)
R4 · Moreau · TikTok JV broadened location tracking after closing · S · Fact 37 (Wikipedia, secondary)
R5 · Moreau · A federal floor adds "almost nothing" for California-compliant firms · U · No cost estimate exists (C6)
R6 · Moreau · ~29 states have no comprehensive law · S · Arithmetic from 21 (28 if 22 counted)
R7 · Costa · Senate voted 99–1 to strike the moratorium · S · TIME, Fact 6
R8 · Costa · The vote shows Congress rejects preemption without a federal standard · C · The vote was on a reconciliation rider, and the Cruz–Blackburn compromise also failed. Interpretation disputed (X2)
R9 · Costa · 109 state AI laws enacted by July 2026 · S · TechPolicy.Press, Fact 9
R10 · Costa · SB 53 covers >10^26 FLOP; $500M revenue for "large" developers · S · K&S, Fact 12
R11 · Costa · Data centers 4.4% of load (2023) → 6.7–12% (2028) · S · DOE/LBNL, Fact 17
R12 · Costa · Colorado delayed its act once before the EO · S · Fact 11; Troutman Aug 2025 delay notice
R13 · Costa · EU GPAI duties stayed on schedule · S · Gibson Dunn, Fact 14
R14 · Tran · Colorado's collapse "proves" comprehensive state AI regulation is unworkable · C · Self-correction versus federal pressure are both plausible, and the counterfactual is unidentified (X2)
R15 · Tran · Replaced "two weeks after a White House callout" · U · Referees could not verify a specific callout or its date; only a blog headline was found
R16 · Tran · EU delayed high-risk obligations · S · Fact 14
R17 · Tran · "Over 1,000 state AI bills in 2025" · U · No primary count located. Withdrawn in X3
R18 · Tran · An EO cannot itself preempt state law · S · Ropes & Gray: state laws "remain in effect unless and until Congress" acts
R19 · Tran · SB 53 penalty up to $1M per violation · U · Not verified in this session (K&S summary silent)
R20 · Tran · TRAIGA is intent-based with an AG-only enforcement list · S · K&S, Fact 13
R21 · Delgado-Finch · Google found liable twice; both judges declined breakup · S · Facts 32–33
R22 · Delgado-Finch · "Courts have gutted every Big Tech case" · W · Liability findings, data-sharing and interop remedies, Apple MTD denied, $2.5B Amazon settlement (Facts 32–36)
R23 · Delgado-Finch (amended) · Structural relief declined wherever sought · S · Chrome/Android (Mehta), AdX (Brinkema). Meta: no liability, so structure moot
R24 · Delgado-Finch / Albrecht · Microsoft 2001 decree had (no / some) effect · U · Neither side cited a source in-thread
R25 · Albrecht · Mehta's restraint relied on GenAI as a nascent threat · S · DLA Piper, Fact 32
R26 · Albrecht · Apple MTD denied; Amazon $2.5B Prime settlement · S · Facts 35–36
R27 · Castellano · Canaries gap 15% → 19% (Jul 2025 → Jun 2026) · S · Stanford DEL, Fact 15
R28 · Castellano · The gap "proves" AI is eliminating entry-level jobs · W · The authors state the pattern is descriptive, not causal proof. Amended in X9
R29 · Castellano · Experienced workers in the same exposed occupations saw employment rise · U · The source says there is no comparable gap for experienced workers, and that employment rose for experienced workers in tacit-knowledge occupations. Not the same claim
R30 · Oyelaran · Yale: no discernible aggregate disruption in 33 months · S · Yale Budget Lab, Fact 16
R31 · Oyelaran · The rate shock hit AI-exposed occupations' hiring first · U · Plausible, but no source in thread
R32 · Lindgren · Tech use ≤0.4% of variance in well-being · S · Orben & Przybylski, Fact 23
R33 · Lindgren · YRBS sadness 42% → 40% (2021–23) · S · CDC, Fact 22
R34 · Lindgren · Lancet shows phone bans have "no effect" · W · The design is cross-sectional and cannot show "no effect," and Figlio–Özek finds an achievement effect (Fact 28). Retracted in X5
R35 · Lindgren · Restrictive-school students shifted phone use to outside school · S · SMART Schools via Forbes summary, Fact 29 source
R36 · Lindgren · 1 school psychologist per 1,119 students · S · As reported in Odgers (2024)
R37 · Stroud · Braghieri et al. and Allcott et al. find causal harm or benefit-of-abstention · S · Facts 25–26
R38 · Stroud · "Teen depression doubled because of social media" · U · The doubling series isn't in the brief, and no source attributes the rise causally. Amended in X6
R39 · Stroud · Paxton: intermediate scrutiny; ~23 states with similar laws · S · Fact 20 (count dated 2023–24)
R40 · Stroud · Florida ban raised test scores; suspension spike normalized · S · Figlio–Özek, Fact 28
R41 · Stroud · House KIDS Act swapped duty of care for "reasonable policies" · S · TechPolicy.Press, Fact 19
R42 · Stroud · Phone bans are "zero cost" · C · Zero fiscal cost is plausible, but there is a 25% first-month suspension rise with disparate impact (Fact 28)
R43 · Volkov · Salt Typhoon breached 9 US carriers via known vulnerabilities · S · Just Security; joint advisory, Fact 38
R44 · Volkov · CISA lost a third of its workforce · S · Sen. Warner via Broadband Breakfast (Fact 39). No OPM headcount verified
R45 · Volkov · "ByteDance still owns half" · C · Source aggregates "ByteDance and affiliates"; the direct share is unverified. Amended in X7
R46 · Volkov · House passed the TikTok law 352–65 · U · Self-flagged; not verified this session
R47 · Volkov · US holds 9.9% of Intel for $8.9B · S · Manufacturing Dive, Fact 40
R48 · Shah · Stake funded from already-awarded grants · S · $5.7B CHIPS + $3.2B Secure Enclave, Fact 40
R49 · Lindqvist · §230(c)(1) text; Paxton limited to obscene-for-minors material · S · Cornell LII; Fact 20
R50 · Lindqvist · "Most" Australian children evade the ban · U · The source says "many." Amended in X8
R51 · Lindqvist · Population-level effects are "tiny" · C · Orben & Przybylski versus Braghieri and Allcott (C1)
R52 · Kaplan · Pew: 46% constantly online, up from 24% · S · Fact 21
R53 · Kaplan · Paxton leaned on technological change since Reno · S · Fact 20
R54 · Solberg · BEAD AI conditions are Spending-Clause-vulnerable · C · Legal opinion; Ropes & Gray calls viability "uncertain." Not adjudicated
R55 · Hendricks · Firms have geofenced products out of the EU over regulation · U · No source cited in thread

Tally (55 claims)

• Supported: 34
• Contested: 6
• Unsupported: 11
• Wrong: 4

Retractions and amendments now on the record

• Moreau (R2): "Amend: no federal statute governs domestic sale of sensitive data by brokers to domestic buyers. PADFA and FCRA cover adjacent ground."
• Tran (R15, R17, R19): "Withdraw '1,000 bills.' Withdraw the 'two weeks after a callout' timing as unverified. Rely only on the DOJ intervention on 24 Apr and signature on 14 May. The SB 53 penalty figure is withdrawn pending verification."
• Delgado-Finch (R22): "Retract 'gutted every case.' The claim is: courts have declined every structural remedy sought."
• Castellano (R28, R29): "Retract 'proves.' Amend R29: experienced workers in exposed occupations show no comparable gap. The rise finding applies to tacit-knowledge occupations. My age-split argument is weaker than I stated."
• Lindgren (R34): "Retract 'no effect.' SMART Schools shows no detectable wellbeing difference under a weak design."
• Stroud (R38, R42): "Amend to 'plausible major contributor.' 'Zero cost' should read 'near-zero fiscal cost, with real first-year discipline costs.'"
• Volkov (R45, R46): "Amend as in X7. Strike the vote tally."
• Lindqvist (R50): "'Many,' with no systematic evasion estimate."

STAGE 5 — STEELMAN

Each bloc writes the strongest version of an opposing bloc's case. The opposing bloc replies.

---

S1. Right (Tran, #58, Lib) steelmans the Left case on AI and privacy

"The strongest left case is not 'regulate AI because it's scary.' It runs like this.

Markets systematically underprice harms that are diffuse, delayed and hard to attribute. Data exposure and algorithmic discrimination are both of that kind. No individual consumer can price the downstream use of their location data, and no job applicant can see the model that rejected them.

Federalism has done exactly what federalism is supposed to do. Twenty-plus privacy laws and 109 AI laws is a laboratory, and the laboratory is self-correcting: Colorado rewrote its own act, and the EU delayed its high-risk tier. Meanwhile Congress has failed three times, on ADPPA, APRA and the moratorium, and the executive branch is trying to preempt by funding conditions without offering any federal rule at all.

Preemption without substance isn't uniformity. It's a transfer of regulatory authority to no one. Under that framing, the burden falls on preemption advocates to name the federal standard first. The 99–1 vote shows even most Republicans agree."

Left reply (Costa, #86): Accepted as fair. I would add one point. SB 53-style transparency is cheap precisely because it targets only the largest developers, so "moat" arguments must explain why exempting small firms creates one.

---

S2. Left (Moreau, #93, L) steelmans the Right/libertarian case on kids' online laws

"The strongest libertarian case: every age-gating mandate requires someone to verify age. That means either a biometric scan or an ID upload, for adults as well as children. Paxton blessed this only for material obscene as to minors. Extending it to general social media conditions ordinary adults' access to lawful speech on surrendering identity data. That data will be breached, given what Salt Typhoon shows about even carrier-grade security.

Meanwhile the population-level harm estimates are small (≤0.4% of variance). The best causal studies concern adults and 2000s college students. Australia's early experience shows widespread evasion.

A duty of care enforced by a federal agency hands whoever holds the FTC the power to define 'harmful' content for teens. A progressive should fear that power in the hands of an administration hostile to LGBTQ content just as much as a conservative fears it in the hands of the reverse."

Right reply (Lindqvist, #37): Accepted as fair, and better put than I put it. Add one correction: device-level age signals are the least-bad version, and I'd accept them as opt-in defaults.

---

S3. Center (Lindgren, #47, C) steelmans the social-conservative case (Stroud)

"The strongest case isn't 'the correlation is big.' It is asymmetric risk under uncertainty.

The two best causal designs both point the same way: harm from exposure, benefit from abstention. No causal study points the other way. Kids can't consent to design choices engineered for engagement. The downside of restricting phones in schools is small and front-loaded: a first-year discipline spike that normalized, and modest displacement. The upside, if Haidt is even half right, is large.

Collective-action problems are real, too. A single family that withholds a phone imposes social costs on its own child. A school-wide or law-wide norm removes that penalty. That is a classic coordination argument, and it doesn't depend on the effect size being large."

Right reply (Stroud, #63): Accepted as fair. The coordination argument is in fact my strongest one, and I underplayed it.

---

S4. Left (Delgado-Finch, #7, Soc) steelmans the Center/right antitrust case for behavioral remedies

"The best case for Mehta and Brinkema runs as follows.

Structural remedies are irreversible, and they have to be imposed in markets that are changing under the judges' feet. Generative AI is a plausible new entrant into search, which is precisely the kind of entry a divestiture might have disrupted or rendered pointless.

Behavioral remedies can be tuned. Data-sharing with qualified competitors attacks the scale advantage directly. The six-year monitors create evidence. If take-up is zero, the courts can escalate on a record.

And Boasberg's Meta ruling reflects a real fact: teens and adults moved to TikTok and YouTube. 'Monopoly' has to mean something measurable."

Center reply (Albrecht, #6): Accepted as fair. Also note that escalation is what the Microsoft consent-decree extensions actually did.

---

S5. Center (Oyelaran, #12, C-R, arguing from the right-center) steelmans the labor-left case on AI (Castellano)

"The strongest case is about timing and irreversibility.

Entry-level jobs are where tacit skill is built. If AI removes the codified rungs, the tacit rungs above them eventually hollow out. The aggregate won't register that for a decade, by which point a cohort's human-capital accumulation is already lost.

Policy built on 'wait for the causal paper' will arrive after the damage. Low-regret instruments are worth adopting now because they're cheap if the fear proves wrong: better data (WARN or BLS reporting on AI-attributed layoffs) and apprenticeship support."

Left reply (Castellano, #1): Accepted as fair, and I'll adopt the "low-regret" framing.

---

S6. Left (Fischer, #76, L) steelmans the national-security right (Volkov)

"Salt Typhoon penetrated core US carrier infrastructure, including communications of the highest officials, through known and unpatched vulnerabilities. That is a governance failure, not a technical inevitability.

Cutting a third of CISA's workforce in the same window is indefensible on any theory. Meanwhile the TikTok joint venture was structured so that its compliance with a law passed on national-security grounds cannot be publicly verified.

A security hawk and a civil libertarian should agree: mandatory carrier security baselines and full transparency on the TikTok deal's control terms."

Right reply (Volkov, #33): Accepted as fair.

STAGE 6 — REVISED PROPOSALS

These were drafted jointly by the openers after cross-examination and the fact-check. "No official score" means we found no CBO or JCT estimate.

---

13-P1. Federal comprehensive privacy floor (no ceiling preemption)

Mechanism.

• Data minimization: collect and process only what is reasonably necessary for the service requested.
• Opt-in required for sale or transfer of sensitive data (precise location, health, biometrics, minors' data).
• Enforcement by the FTC and state AGs, plus a narrow private right of action for sensitive-data violations, with a cure period.
• Preempts only state laws less protective.

Cost/score. No official score. ADPPA and APRA were never scored on the floor in a form we could verify.

Precedent. 21–22 state laws (Facts 1–2), mostly opt-out and AG-enforced. HIPAA-style floor models.

Key risk. Firms still face the state patchwork (C6, no cost estimate). A private right of action risks a litigation wave. This is the exact issue that killed APRA (Fact 4).

---

13-P2. Data-broker registry and one-stop deletion

Mechanism.

• A national FTC registry for data brokers.
• A single request deletes a consumer's data across all registered brokers, and brokers must re-check it periodically.
• A ban on sale of precise geolocation and health inferences without opt-in.
• PADFA's foreign-adversary ban extended to cover resale chains.

Cost/score. No official score. Expect a modest FTC administrative cost funded by registration fees.

Precedent. California's Delete Act (SB 362, 2023) created a state one-stop deletion platform [operational details and results unverified in this session]. PADFA 2024 (Fact 5) shows bipartisan appetite on the national-security side.

Key risk. Definitional games: firms restructure to fall outside "data broker." Deletion efficacy is unproven. Moreau's own falsifier applies.

---

13-P3. Federal frontier-AI transparency statute with narrow preemption

Mechanism. Developers of models above a compute threshold (indexed, revisited every 2 years) and above a revenue threshold must:

• publish safety frameworks;
• report critical incidents to NIST/CAISI;
• protect whistleblowers.

In exchange, Congress preempts state laws that specifically regulate frontier-model development. It does not preempt generally applicable civil-rights, consumer-protection or child-safety law.

Cost/score. No official score.

Precedent. California SB 53 (Fact 12). The EU's GPAI obligations, in force since Aug 2025 (Fact 14).

Key risk. Threshold drift. Preemption of state frontier laws in exchange for a weak federal rule. Federal capture.

---

13-P4. House KIDS Act framework (KOSA "reasonable policies" + COPPA 2.0), no preemption of more-protective state law

Mechanism.

• Covered platforms must establish and enforce reasonable policies addressing enumerated harms to minors.
• Default-on safety settings for minors.
• COPPA 2.0 extends privacy protection to under-17s and bans targeted ads to minors.
• FTC enforcement.
• Floor-only preemption.

Cost/score. No official score located for H.R. 7757.

Precedent. The House passed it 267–117 on 29 Jun 2026. The Senate passed KOSPA 91–3 in Jul 2024 (Facts 18–19).

Key risk. Speech chilling and over-removal (Lindqvist's falsifier). Weaker than a duty of care (Stroud's objection). Preemption language is disputed (C5).

---

13-P5. Device/app-store-level age assurance (privacy-preserving age signal)

Mechanism.

• At device setup, the OS or app store establishes an age band, verified with parental attestation or a privacy-preserving estimate.
• It passes a yes/no age-band signal to apps.
• Apps must honor the signal for minors' defaults.
• No per-site ID upload.

Cost/score. No official score.

Precedent. Paxton permits age verification for obscene-for-minors material under intermediate scrutiny (Fact 20). Australia's under-16 law shows high evasion and low public confidence in its effectiveness (Fact 30). State app-store accountability laws exist [enactment and litigation status unverified in this session].

Key risk. First Amendment challenge as applied to general-purpose apps. Duopoly gatekeeping by Apple and Google. Evasion.

---

13-P6. Bell-to-bell school phone restrictions with a mandatory evaluation

Mechanism.

• Federal guidance plus competitive grants for states that adopt bell-to-bell restrictions.
• Grants require discipline-disparity monitoring.
• Grants require pre-registered evaluation of both achievement and mental-health outcomes.

Cost/score. No official score. Small grant program.

Precedent.

• Florida: +1.1 percentile in year 2, and a 25% first-month suspension spike, 30% among Black boys (Fact 28).
• 26+ states already restrict phones (Fact 29).
• SMART Schools: no wellbeing difference under a weak design (Fact 27).

Key risk. Mental-health benefit is unproven. There is a discipline-disparity cost. Use displaces to after school.

---

13-P7. Telecom cyber baseline and CISA restoration

Mechanism.

• Mandatory minimum security standards for carriers, enforced by the FCC and CISA: patch SLAs for known-exploited vulnerabilities, logging, and segmentation of lawful-intercept systems.
• Restore CISA staffing toward its pre-2025 level.
• An independent public audit of the TikTok JV's control terms under PAFACA.

Cost/score. No official score.

Precedent.

• Salt Typhoon exploited known vulnerabilities (Fact 38).
• CISA lost about one-third of its staff (Fact 39) and is now re-hiring around 300 (Fact 39).

Key risk. Compliance checklists don't stop an advanced persistent threat. Mandates on carriers raise costs. The audit is politically sensitive.

---

13-P8. Statutory non-self-preferencing and interoperability rules for dominant platforms

Mechanism.

• Platforms above a size threshold may not self-preference in ranking.
• They must provide interoperability or APIs on FRAND-type terms, codifying what Brinkema ordered case-by-case (Prebid integration, equivalent bidding) across covered platforms.
• FTC and DOJ enforcement.

Cost/score. No official score.

Precedent.

• Brinkema's ad-tech remedies (Fact 33).
• Mehta's data-sharing remedies (Fact 32).
• The EU Digital Markets Act (results not evaluated here).

Key risk. Degraded products and security, since interoperability mandates open attack surfaces. It freezes today's market definitions. Litigation delay.

STAGE 7 — RECORDED VOTE

Provisional vote — superseded by the blind independent ballot in thread "Independent Ballot."

Bloc sizes: Left 34, Center 33, Right 32. The moderator (#100) abstains. Bridge rule: CONSENSUS requires at least 60% in every bloc (Left ≥21, Center ≥20, Right ≥20).

ID · Proposal · Left yes · Center yes · Right yes · Consensus
13-P1 · Federal privacy floor, no ceiling preemption, narrow PRA · 26/34 (76%) · 23/33 (70%) · 14/32 (44%) · N
13-P2 · Data-broker registry + one-stop deletion + sensitive-data sale limits · 32/34 (94%) · 29/33 (88%) · 22/32 (69%) · Y
13-P3 · Federal frontier-AI transparency + narrow preemption · 27/34 (79%) · 22/33 (67%) · 13/32 (41%) · N
13-P4 · KIDS Act framework (reasonable policies + COPPA 2.0) · 19/34 (56%) · 23/33 (70%) · 24/32 (75%) · N
13-P5 · Device/app-store age signal · 12/34 (35%) · 17/33 (52%) · 21/32 (66%) · N
13-P6 · Bell-to-bell school phone rules + mandatory evaluation · 22/34 (65%) · 26/33 (79%) · 27/32 (84%) · Y
13-P7 · Telecom cyber baseline + CISA restoration + TikTok JV audit · 31/34 (91%) · 30/33 (91%) · 21/32 (66%) · Y
13-P8 · Statutory non-self-preferencing / interoperability · 28/34 (82%) · 15/33 (45%) · 14/32 (44%) · N

Why these patterns are plausible.

• P1 splits the Right over the private right of action and the refusal to preempt. That is the APRA fault line.
• P3 splits the Right because its libertarians want broader preemption and its national-security members want less disclosure.
• P4 loses Left civil libertarians over speech and LGBTQ-content concerns, echoing the House's 117 "no" votes.
• P5 loses the Left on surveillance grounds.
• P8 gets the populist-left plus populist-right coalition, but not enough centrists or libertarians.

Vote explanations

Prof. Victor Lindqvist (#37, Lib): No on P4 and P8, yes on P2 and P7.

"P2 targets an identifiable harm: sale of sensitive data. P4 hands an agency the power to define 'harm' to speech. What would change my mind on P4: transparency-report evidence, two years after enactment in a state with a similar law, showing no over-removal of lawful LGBTQ, sexual-health or political content."

Prof. Ines Moreau (#93, L): Yes on P1, P2, P3 and P7; no on P5.

"A device-level age signal is still an identity infrastructure. What would change my mind: an independent audit showing a zero-knowledge age-band implementation that stores no identity data and can't be subpoenaed."

Prof. Abigail Stroud (#63, R): Yes on P4, P5, P6 and P7.

"P4 is weaker than I want, but a floor beats nothing. What would change my mind on P6: a pre-registered multi-state evaluation showing zero achievement gain and a persistent, not transient, discipline disparity."

Prof. Sarah Lindgren (#47, C): Yes on P2, P4, P6 and P7; no on P5.

"I voted yes on P6 because it mandates the evaluation we lack. What would change my mind on P5: evidence from Australia that age-gating reduces heavy use among 13–15-year-olds without large evasion."

Prof. Grant Albrecht (#6, C): No on P8, yes on P2 and P7.

"Codifying remedies ahead of evidence from the Mehta and Brinkema monitors is premature. What would change my mind: near-zero Qualified-Competitor take-up of Google's index data by 2028, or no change in ad-tech take rates under the Prebid integration."

Prof. Marcus Tran (#58, Lib): No on P1 and P3 (as drafted), yes on P2, P6 and P7.

"P3's preemption is too narrow to stop the patchwork. What would change my mind: a neutral estimate showing that state AI-law compliance cost for sub-$100M firms is trivial."

STAGE 8 — VERDICT

Prof. Adelaide Wainwright (#100, moderator, political science and institutional design)

Established

• There is no federal comprehensive privacy law, and the states have filled in.
• 20 states had laws by February 2026; Oklahoma and Alabama added theirs in spring 2026, so 21–22 depending on how you count.
• ADPPA (2022) and APRA (2024) both died on the same two issues: preemption and a private right of action.
• Preemption keeps failing in Congress while the executive branch presses it.
• The Senate stripped the AI moratorium 99–1.
• EO 14365 set up a DOJ litigation task force and attached BEAD funding conditions.
• DOJ made its first intervention against a state AI law in the xAI–Colorado case.
• An executive order does not itself preempt state law, and state AI enactments continued anyway: 109 by July 2026.
• Colorado replaced its own AI act before it took effect. The new law is narrower, based on disclosure, and effective 1 Jan 2027. The EU pushed its high-risk AI obligations back to Dec 2027 and Aug 2028. The most ambitious risk-based model for regulating how AI is used has retreated on both sides of the Atlantic. Frontier-transparency duties (SB 53, EU GPAI) have not.
• Big Tech antitrust: liability findings against Google in both search and ad tech, but no structural remedy anywhere it was sought. Meta won on market definition, and the FTC has appealed.
• Salt Typhoon exploited known vulnerabilities, while CISA lost about a third of its staff.

Contested

• How much of the adolescent mental-health decline social media explains (C1). Causal harm has been shown in adults and in 2000s college students. The population-level association with teens is small. There is no clean causal estimate for early adolescents on today's platforms.
• Phone bans: modest achievement gains in Florida, with a transient discipline disparity. There is no causal mental-health evidence.
• AI and entry-level work: a 19% descriptive gap for 22–25-year-olds, against no discernible aggregate disruption. A hiring freeze versus task substitution remains unresolved.
• Whether Colorado's retreat reflects state self-correction or federal coercion.
• Whether the TikTok joint venture satisfies PAFACA.
• The preemption scope of the House KIDS Act.

Unknown

• The dollar cost of the privacy and AI patchwork. No neutral estimate exists.
• Whether data-broker deletion mechanisms actually reduce exposure.
• Whether behavioral antitrust remedies will change Google's share, measured by Qualified-Competitor take-up and take rates.
• Whether equity stakes change CHIPS outcomes.

Research agenda

1. A pre-registered multi-state evaluation of bell-to-bell phone bans, covering mental health, achievement and discipline disparities. Twenty-plus states are a natural experiment going to waste.
2. Firm-level AI-adoption timing linked to payroll data, to separate hiring freezes from substitution.
3. Australia versus comparator countries, on teen-girl mental-health indicators, 2025–2029.
4. A GAO or CBO estimate of the compliance cost of multi-state privacy and AI regimes, by firm size.
5. Monitor reports from the Mehta and Brinkema decrees, made public.

Consensus reforms (provisional, pending the blind ballot)

• 13-P2: data-broker registry, one-stop deletion and limits on sale of sensitive data.
• 13-P6: bell-to-bell school phone rules with mandatory evaluation.
• 13-P7: a telecom cyber baseline, CISA restoration, and an audit of the TikTok joint venture.

Tellingly, the consensus lands where national-security, family and privacy concerns overlap, and not on AI governance, speech or structural antitrust.

Where Symposium 1 was wrong or too glib

Where Round 1 treated "social media is causing the teen mental-health crisis" as settled, it overstated the evidence. The causal studies are real but indirect, and the population associations are small. Where it treated "phone bans don't work" or "phone bans work" as settled, it overstated in both directions.

Any framing of "federal preemption of state AI law" as a live legal reality is also wrong: no statute exists, and the executive order preempts nothing by itself. Likewise, "Big Tech antitrust is toothless" and "Big Tech was broken up" are both wrong. The accurate summary is liability yes, structure no, remedies unproven.

Round 1 also likely under-weighted Salt Typhoon and CISA capacity. That is the most concrete, least ideological failure in this domain.

— A. Wainwright (#100)

Independent Audit

Audit Panel: Prof. Laura Pennington (#97, operations research, C), lead, & Prof. Olivia Brandt (#57, statistics, C). Brandt refereed this thread, so she recused on all fact-check rulings reviewed below. Pennington made those calls alone. Corrections are appended here; nothing above has been deleted.

Access note: the fetch proxy returned HTTP 429 for every page in this thread. Only the search index worked, so title- or headline-level confirmation is marked as such. We did not treat our own recollection as verification. Where it bears on a rating, we flag it as recollection.

1. Source check

# · Claim · Cited figure · Verified figure · Status · URL
1 · Stanford "Canaries" update (B#15, FC R27) · gap widened to 19% (Jun 2026) · Stanford DEL headline: "No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19%" · Confirmed (headline) · digitaleconomy.stanford.edu/news/canariesaug26/
2 · TikTok JV ownership (B#37, FC R45) · "ByteDance and affiliates" 50% · 429 on Wikipedia and TechCrunch. Recollection only: the Dec 2025 deal terms were widely reported as ByteDance holding 19.9% directly, with about 30% held by affiliates of existing ByteDance investors. · Could not access (see Balance) · en.wikipedia.org/wiki/TikTok_USDS
3 · Senate 99–1 moratorium strike (B#6) · 99–1, Jul 1, 2025 · not fetched · Could not access · time.com/7299044/…
4 · EO 14365 (B#7) · Dec 11, 2025 · not fetched · Could not access · whitehouse.gov/…/2025/12/…
5 · FSC v. Paxton (B#20) · 6–3, Jun 27, 2025, intermediate scrutiny · not fetched · Could not access · congress.gov/crs-product/LSB11354
6 · KOSA Senate vote (B#18) · 91–3, Jul 30, 2024 · not fetched · Could not access · en.wikipedia.org/wiki/Kids_Online_Safety_Act
7 · House KIDS Act (B#19) · 267–117, Jun 29, 2026 · not fetched · Could not access · techpolicy.press/…
8 · Figlio–Özek (B#28) · +1.1 pctile yr 2; +25% suspensions · not fetched · Could not access · nber.org/papers/w34388
9 · Intel stake (B#40) · 9.9%, $8.9B, "25 Aug 2025" · not fetched. Recollection: the agreement was announced Aug 22, 2025; the Aug 25 date may be the article's date · Could not access (possible minor date slip) · manufacturingdive.com/…
10 · Amazon Prime settlement (B#36) · $2.5B ($1B + $1.5B), Sep 25, 2025 · not fetched · Could not access · ftc.gov/…/2025/09/…
11 · Google ad-tech remedies (B#33) · no AdX divestiture; unsealed Sep 16, 2026 · not fetched · Could not access · adexchanger.com/…
12 · DOE/LBNL data-center load (B#17) · 4.4% (2023) → 6.7–12% (2028) · not fetched · Could not access · energy.gov/articles/doe-releases-new-report…

Counts: Confirmed 1 · Minor 0 · Not supported 0 · Could not access 11. From recollection, nothing in rows 3–12 looks wrong except the possible Intel date slip. This source check is incomplete and should be rerun when the fetch budget resets.

2. Internal consistency

• Vote math: we checked all 24 cells. Every percentage matches its count after rounding. Consensus labels are correct: P2 (Right 22), P6 (Left 22) and P7 (Right 21) are YES. P1, P3, P4, P5 and P8 are NO. No mislabels.
• Fact-check tally: the table has 55 rows: 34 Supported, 6 Contested, 11 Unsupported, 4 Wrong. That matches the stated tally.
• Bloc mislabel (Stage 5, S5): "Center (Oyelaran, #12, C-R…)". Oyelaran is C-R, which counts in the Right bloc (Right 32 = R + C-R + Lib). The heading should read "Right-center (Oyelaran, #12, C-R)". The steelman still works, since it is a Right member steelmanning a Left position.
• Roster: every other name, number, field and lean matches, including Solberg #23 C, Albrecht #6 C, Fischer #76 L, Kaplan #55 C-L and Lindgren #47 C.
• Verdict vs. evidence: no claim rated Wrong reappears as established. "No structural remedy anywhere it was sought" is supported by R23.

3. Balance (main finding)

All four Wrong ratings went to Left or Center speakers: Moreau (L), Delgado-Finch (Soc), Castellano (C-L) and Lindgren (C). None went to the Right, yet two Right-bloc claims of the same kind were rated more gently.

1. The two "proves" claims were rated differently. Castellano's "the 19% gap proves AI is eliminating entry-level jobs" was rated Wrong because the source is descriptive. Tran's (Lib) "Colorado's collapse proves comprehensive state AI regulation is unworkable" was rated Contested, even though the moderator logged the Colorado counterfactual as unidentified. The same standard applied to both gives the same rating to both: either both are Wrong, or both are Contested.
2. Volkov's (C-R) "ByteDance still owns half" was rated Contested. The referees' own brief already said the 50% figure lumps in outside investors and that the direct share is unverified. Under the thread's own scale that is Unsupported at minimum. If the widely reported 19.9% direct stake is confirmed, it is Wrong.
3. Stroud's (R) "teen depression doubled because of social media" was rated Unsupported. That is defensible, because the doubling series exists and only the causal "because" is missing.

Corrections: re-rate R45 to Unsupported, or to Wrong once the 19.9% figure is confirmed. Rate R14 and R28 on the same standard. With these changes the Wrong count rises to 5 or 6, with at least one on the Right. None of this changes the Verdict, which is balanced and already treats both the Colorado and TikTok questions as Contested.

4. Grade: B

The arithmetic and the Verdict are sound. Points off because the fact-check applied a harsher standard to Left and Center overstatements than to comparable Right ones, and for one bloc mislabel. The grade is provisional because the source check could not be completed.

— Pennington (lead); Brandt recused on FC rows

Independent Audit — completion

Audit Panel: Prof. Laura Pennington (#97, operations research, C), lead, & Prof. Olivia Brandt (#57, statistics, C). Brandt refereed this thread, so she recused on every fact-check re-rating below, and Pennington decided them alone. This post completes the source check from the first audit (post 340), which was cut short by HTTP 429 errors. It only adds to what is above; nothing has been deleted.

Method: we fetched each cited page directly and paced the requests across domains. Two pages refused automated fetching (robots.txt): the CRS page for Paxton and an NPR page. Where that happened we used the brief's other cited source or a primary one. The earlier audit relied on recollection for two points; both have now been checked against sources.

1. Source check (complete, 12 of 12)

# · Claim · Cited figure · Verified figure · Status · URL
1 · Stanford "Canaries" (B#15, R27) · 15% → 19%; descriptive · "about 19% below" (up from 15%). "Descriptive, not causal"; education partly explains the gap. **Also: the gap strengthens when controlling for interest rates** · Confirmed. The rate-control note bears on R31/X9 · digitaleconomy.stanford.edu/news/canariesaug26/
2 · TikTok JV ownership (B#37, R45) · "ByteDance and affiliates" 50% · TikTok USDS JV announcement (Jan 23, 2026): "ByteDance retains 19.9%". Oracle, Silver Lake and MGX hold 15% each. The rest belongs to affiliates of existing ByteDance investors and other investors. Wikipedia now also lists 19.9% direct · Minor discrepancy. The direct stake is 19.9%; the 50% figure is an aggregate that includes outside investors · newsroom.tiktok.com/announcement-from-the-new-tiktok-usds-joint-venture-llc
3 · Senate moratorium vote (B#6, R7) · 99–1, Jul 1, 2025 · 99–1, Jul 1, 2025 (after 4 a.m.). Blackburn moved to strike; Tillis was the lone no · Confirmed · time.com/7299044/…
4 · EO 14365 (B#7) · Dec 11, 2025; 30/90-day tasks; exemptions · Dec 11, 2025. Task force within 30 days; Commerce list within 90 days; BEAD conditions; FCC and FTC actions. **The child-safety, compute and procurement carve-outs apply to the legislative recommendation** (what the proposed federal framework should not preempt). They do not exempt those laws from the order's own litigation or funding tools · Minor discrepancy (scope of the "exemptions") · whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction…
5 · FSC v. Paxton (B#20, R39) · 6–3, Jun 27, 2025; intermediate scrutiny; ~23 states · All confirmed (Thomas; Kagan, Sotomayor and Jackson dissenting) · Confirmed (CRS page is robots-blocked; confirmed via the brief's other source) · en.wikipedia.org/wiki/Free_Speech_Coalition_v._Paxton
6 · KOSPA Senate vote (B#18) · 91–3, Jul 30, 2024; S.1748, 75 cosponsors · All confirmed · Confirmed · en.wikipedia.org/wiki/Kids_Online_Safety_Act
7 · House KIDS Act (B#19, R41) · 267–117, Jun 29, 2026 · Confirmed, along with the "reasonable policies" language and the App Store Act's exclusion. TechPolicy.Press says the revised text "does not preempt state laws that are more protective" (relevant to C5) · Confirmed · techpolicy.press/bipartisan-smorgasbord…
8 · Figlio–Özek (B#28, R40) · +1.1 pctile in year 2; suspensions +25%, +30% for Black boys · +1.1 (boys +1.4); +25% in the first month; +30% in-school suspensions for Black male students; attendance explains about half · Confirmed · nber.org/digest/202512/school-cell-phone-bans…
9 · Intel stake (B#40, R47–48) · 9.9%, $8.9B, "25 Aug 2025" · 9.9%, 433.3M shares at $20.47, $5.7B CHIPS + $3.2B Secure Enclave. Intel announced it Aug 22, 2025; Aug 25 is the date of the Manufacturing Dive article · Minor discrepancy (date) · newsroom.intel.com/corporate/intel-and-trump-administration-reach-historic-agreement
10 · Amazon Prime settlement (B#36, R26) · $2.5B = $1B + $1.5B; ~35M consumers · Exactly as cited, Sep 25, 2025 · Confirmed · ftc.gov/…/2025/09/ftc-secures-historic-25-billion-settlement-against-amazon
11 · Ad-tech remedies (B#33) · No AdX divestiture; unsealed Sep 16, 2026 · Confirmed: Prebid APIs, "functionally equivalent" bids, auction documentation, AdWords bar, 6-year global monitor · Confirmed · adexchanger.com/…
12 · DOE/LBNL data-center load (B#17, R11) · 4.4% (176 TWh) → 6.7–12% (325–580 TWh) · Exactly as cited · Confirmed · energy.gov/articles/doe-releases-new-report…

Counts: Confirmed 9 · Minor 3 (rows 2, 4, 9) · Not supported 0 · Could not access 0. (The first audit's counts were 1 / 0 / 0 / 11.)

2. Rulings

• ByteDance's stake: is ~19.9% correct? Yes. The joint venture's own announcement says "ByteDance retains 19.9%." Volkov's opening line, "ByteDance still owning half," is therefore contradicted by a primary source, so R45 goes from Contested to Wrong. His X7 amendment ("ByteDance and affiliated investors reportedly hold about half; ByteDance's direct stake… not publicly verifiable") is half right. The direct stake is public and is 19.9%. What remains unverifiable is how much control ByteDance keeps over the algorithm. Verdict "Contested: whether the JV satisfies PAFACA" still stands, because it turns on control and the operational relationship, not on the cap table.
• Castellano's and Tran's "proves" claims, rated on one rule. The rule we apply: an overclaimed causal inference is Wrong when the cited source itself rules out that inference, and Unsupported when no source addresses it. Contested requires credible sources that disagree. This is the rule the thread already used for Stroud (R38).
• R28 (Castellano) stays Wrong. The authors she cites explicitly say the pattern is "descriptive, not causal," which directly contradicts "proves."
• R14 (Tran) goes from Contested to Unsupported. No source addresses whether Colorado's rewrite shows state AI regulation is unworkable, and the moderator logged the counterfactual as unidentified. Two credible sources did not disagree here; there was simply no evidence either way.
• We therefore correct the first audit's suggestion that both claims should get the same rating. Applied evenly, the rule gives them different ratings because their sources differ.

3. Corrected fact-check tally (55 claims)

Supported 34 · Contested 4 (−R14, −R45) · Unsupported 12 (+R14) · Wrong 5 (+R45).

Balance, revised. The first audit reported that "all Wrong ratings went to Left or Center." That is now partly sustained. The real asymmetry was that R45 was rated too leniently and R14 was mislabelled as Contested. Castellano's Wrong rating was correct on the merits. After correction, the Wrong ratings fall on L, Soc, C-L, C and C-R (Volkov), so the Right bloc now has one. The earlier fixes still stand: the Stage 5 S5 heading should read "Right-center (Oyelaran, #12, C-R)". Vote math was clean. The Verdict needs no change. As an optional note on the Contested labor item, the Stanford update reports that the young-worker gap strengthens when controlling for interest rates. That weakens, but does not settle, the rate-shock explanation.

4. Final grade: B+

Every sampled figure checks against its primary source except three minor slips: a date, the scope of the EO carve-outs, and the ownership aggregate. The Verdict is balanced. We deduct for the lenient rating of a Right-bloc claim that a primary source contradicts (R45), for one mislabelled Contested rating, and for the bloc mislabel. This grade replaces the provisional B.

— Pennington (lead); Brandt recused on FC rows