Executive Order 14110 on Safe, Secure, Trustworthy AI

Operative mechanics

EO 14110 (Exec. Order No. 14110, 88 Fed. Reg. 75191 (Nov. 1, 2023)) operated chiefly as a tasking instrument that directed federal agencies to produce binding sub-regulation, rather than imposing duties directly on developers. Its most cited operative lever is §4.2, which invokes the Defense Production Act of 1950 to require companies developing or possessing a "dual-use foundation model" to report training activity, physical and cybersecurity protections for model weights, ownership of those weights, and the results of red-team safety testing conducted per NIST guidance (§4.2(a)). The trigger is a compute proxy: models trained on more than 10^26 integer or floating-point operations, or 10^23 operations where the model uses primarily biological-sequence data (§4.2(b)) — a carve-out that tracks documented dual-use biosecurity risk at the AI-synthetic-biology interface ¹. A parallel reporting duty attaches to any computing cluster physically co-located in one datacenter with networking over 100 Gbit/s and a theoretical maximum of 10^20 operations per second (§4.2(b)). Separately, §4.2(c)-(d) directs Commerce to require U.S. IaaS providers to file Know-Your-Customer reports when a foreign person trains a large model with potential malicious cyber capability, and to verify foreign-customer identity through resellers. Content provisions sit in §4.5, which set a 240-day deadline for a Commerce report on authentication, provenance tracking, watermarking, and synthetic-content detection — duties whose practical bite is unclear given that audits find only ~38% of image generators implement adequate watermarking ². The science-standards backbone is §4.1, requiring NIST within 270 days to issue red-teaming guidelines and a generative-AI companion to the AI Risk Management Framework (NIST AI 100-1).

Cross-jurisdiction position

As a compute-threshold instrument, EO 14110 is most directly comparable to the EU AI Act's general-purpose AI (GPAI) regime, but diverges sharply in legal form and stringency. The Act presumes systemic risk for GPAI models trained above 10^25 FLOP (Art. 51(2), Reg. (EU) 2024/1689) — an order of magnitude below the Order's 10^26 trigger — and attaches substantive obligations (model evaluation, adversarial testing, systemic-risk mitigation, incident reporting) under Art. 55, where EO 14110 mandated only disclosure of training and red-team results to government. The comparison is complicated by definitional instability in the Act's own categories: the legal text shifted across versions among "AI system, general purpose AI system, foundation model, and generative AI" ³, so the two instruments do not cleanly target the same object. The thresholds were nonetheless set close in time and both trace to the same scholarly lineage; the 10^26 figure echoes the frontier-model framing of Anderljung et al., "Frontier AI Regulation" ⁴. The Order is also far thinner than the Act in durability: it was a self-executing tasking memo resting on existing statutory authority (chiefly the DPA), whereas the Act is binding primary legislation. Relative to the Council of Europe Framework Convention on AI (CETS No. 225, 2024), EO 14110 was narrower, focused on a national-security and standards agenda rather than human-rights treaty obligations. Against China's algorithm- and generative-AI registration rules (e.g., the 2023 Interim Measures for Generative AI Services), the Order shared a provenance/labelling concern (§4.5) but relied on voluntary standards rather than pre-deployment filing and content control.

Key fault lines and critiques

The central scholarly debate concerns whether training compute is a defensible regulatory trigger. Critics argue the 10^26 FLOP line is under-justified and is at best an imperfect proxy for risk: capability and harm correlate only loosely with pre-training compute, so high-compute models may be benign while lower-compute systems (e.g., narrow biological-design or toxicity models) can be more dangerous — a mismatch examined in Heim and Koessler, "Training Compute Thresholds" ⁵ and in Anderljung et al. ⁴, and underscored by biosecurity work showing acute dual-use danger from comparatively small synthetic-biology models ¹. A related loophole critique notes that the Order counts only training compute and ignores inference-time scaling and post-training elicitation, inviting threshold-gaming; defenses against such circumvention are canvassed in Pistillo and Villalobos, "Defending Compute Thresholds Against Legal Loopholes" ⁶. A second fault line is institutional legitimacy: commentators questioned grounding economy-wide AI reporting in the Defense Production Act, a Korean-War-era statute, rather than tailored legislation (CRS Report R47843, 2023). A third concerns capacity and durability — the Stanford HAI implementation tracking found agencies broadly met early §4 deadlines, yet the Order's reliance on executive discretion left it politically fragile, and observers (e.g., TIME, 2023) flagged that it "only goes so far" absent statutory backing. Practitioners also debated the §4.5 watermarking mandate as technically immature, since robust provenance detection remained unsolved ².

Implementation and trajectory

Implementation proceeded rapidly through 2024. NIST delivered the §4.1 deliverables, releasing the Generative AI Profile companion to the AI RMF (NIST AI 600-1) and red-teaming/secure-software guidance in July 2024, and the U.S. AI Safety Institute was stood up within NIST. The §4.2 reporting mandate was operationalized when the Bureau of Industry and Security issued a proposed rule, "Establishment of Reporting Requirements for the Development of Advanced Artificial Intelligence Models and Computing Clusters" (Sept. 11, 2024), which would have required quarterly filings on training runs above 10^26 operations and on large clusters — a compute-governance lever whose efficacy is contested, since chokepoint controls on the same compute supply chain have proven leakier than intended ⁷. The trajectory then inverted: President Trump revoked EO 14110 on January 20, 2025, and issued EO 14179, "Removing Barriers to American Leadership in Artificial Intelligence," 90 Fed. Reg. 8741 (Jan. 31, 2025), directing agencies to suspend, revise, or rescind actions taken under the prior Order. BIS did not finalize its DPA rule before revocation, leaving the reporting regime in abeyance (Skadden, 2024), and the open compute-threshold loopholes it would have inherited remained unaddressed ⁶. This rescission explains the instrument's "partial" catalog status: NIST guidance artifacts persist, but the binding reporting backbone lapsed. The successor policy was reframed around innovation and classified cyber-capability benchmarking — see EO of June 2026, "Promoting Advanced Artificial Intelligence Innovation and Security," 91 Fed. Reg. (June 5, 2026) — replacing a fixed compute threshold with a discretionary, classified "covered frontier model" designation (Wiley, 2026; Greenberg Traurig, 2026).

Enforcement record

Documented enforcement actions catalogued against Executive Order 14110 on Safe, Secure, Trustworthy AI (or against rules that this instrument now subsumes).

• FTC investigation of OpenAIUS · 2023 · ongoing
  Federal Trade Commission v. OpenAI — Civil Investigative Demand alleging consumer-protection violations: misleading claims about ChatGPT capabilities, training-data privacy, and consumer harm from hallucinations.
  Lesson: First US federal enforcement action against a frontier-AI developer. Establishes that pre-AI-statute consumer-protection authority (FTC §5) can be applied to AI services — supports the US 'sectoral / ex-post liability' regime (vs EU's ex-ante AIA). Action remains pending; no judgment yet.
  Source record →news secondary
• New York Times v. OpenAI + MicrosoftUS · 2023 · ongoing
  New York Times Company (private civil litigation) v. OpenAI Inc. + Microsoft Corp. — Unauthorised reproduction of NYT-copyrighted articles in GPT training corpora; output of substantially similar text on prompted query; removal of copyright-management information.
  Lesson: First major frontier-foundation-model copyright lawsuit by a primary news source. Discovery has surfaced disclosure of training-data composition that the EU AIA Art. 53 transparency requirements would have surfaced ex-ante. The case is the highest-stakes ex-post-liability action testing whether US sectoral approach can substitute for ex-ante regulation on training-data rights — outcome will inform 2025-2027 regulatory debates.
  Source record →
• Mobley v. Workday (US AI-hiring class action)US · 2023 · ongoing
  Mobley v. Workday, Inc., No. 3:23-cv-00770 (N.D. Cal.)
  Private civil class action; EEOC amicus participation v. Workday Inc. — Workday's algorithmic hiring tools allegedly screened out applicants on disability, age, and race. Class action seeks to certify Workday as an 'employment agency' under Title VII so disparate-impact theory applies to the algorithm's outputs rather than only its developers.
  Lesson: First major US AI-hiring class action with EEOC amicus support. If Workday is certified as an 'employment agency', US sectoral approach (EEOC + Title VII) substantially expands AI-hiring liability without requiring an AI statute. This is the load-bearing test of whether US 'principles + ex-post liability' approach can substitute for EU AIA Annex III §4 (high-risk employment AI obligations).
  Source record →regulator landing
• EEOC v. iTutorGroup (AI age-discrimination consent decree)US · 2022–2023
  EEOC v. iTutorGroup, Inc., No. 1:22-cv-02565 (E.D.N.Y.)
  Equal Employment Opportunity Commission (EEOC) v. iTutorGroup, Inc. — iTutorGroup's recruiting software automatically rejected female applicants aged 55 and older, and male applicants aged 60 and older, regardless of qualifications.
  Lesson: First US EEOC-as-party suit against an AI-mediated hiring tool resolved by consent decree ($365,000 settlement + 5-year monitoring; required revised non-discriminatory application processes; mandatory anti-discrimination training; right to re-apply for rejected applicants). Establishes that pre-AI civil-rights statutes (ADEA, Title VII, ADA) can be applied to algorithmic hiring outputs without requiring a dedicated AI statute — the load-bearing precedent for the US 'sectoral / ex-post liability' regime in employment AI.
  Source record →regulator landing
• HUD / DOJ v. Facebook (ad-targeting Fair Housing Act)US · 2018–2022
  US Department of Housing and Urban Development (HUD) + Department of Justice (DOJ) v. Meta Platforms, Inc. (Facebook) — Facebook's ad-delivery and ad-targeting tools (including 'Special Ad Audience' / Lookalike Audiences) allowed advertisers to exclude users on the basis of protected classes (race, colour, religion, sex, familial status, national origin, disability), and the platform's algorithmic delivery further skewed ad reach.
  Lesson: First major US federal settlement holding a platform liable for discriminatory algorithmic delivery under a pre-AI civil-rights statute. DOJ settlement (June 2022) required Meta to develop a new 'Variance Reduction System' to redress racially-skewed ad delivery + sunset the Special Ad Audience tool. Establishes that algorithmic-delivery discrimination — not just user-facing targeting options — is reachable under FHA. Subsequently cited as the template for analogous reasoning under ECOA (lending) and ADEA (employment).
  Source record →regulator landing

Cross-jurisdiction comparison

How peer instruments treat the topics Executive Order 14110 on Safe, Secure, Trustworthy AI governs.

°= industry self-imposed voluntary framework. Comparing a voluntary code's "governs" tint with a binding regulation's "governs" tint flattens the legal-force distinction; use the instrument-page banner for the operative status of each.

Does this instrument’s approach work? — the social-science evidence

Aggregated over the 18 topics this instrument governs: whether each harm is empirically real, and whether the peer-reviewed evidence shows governance reduces it. The badge is the epistemic status of the evidence— “thin”/“absent” efficacy evidence is itself a finding (the “second silence”). Each epistemic-status label is Policy Window's editorial assessment of the cited evidence base (a structured classification), not a verdict any single source issues.

Of the 18 governed topics with a social-science evidence review, evidence that governance reduces the harm is established for 0, contested for 0, thin for 6, and absent for 12 — for most, no replicated study yet shows this instrument's approach works (the "second silence").

• AI-Driven Worker Displacement

  • Is the harm real?evidence: contested

    AI-driven labour displacement is demonstrably real but localized rather than economy-wide as of 2025-2026. Causal microdata find measurable harm in directly exposed segments: a difference-in-differences study of the Upwork freelance market found that after ChatGPT's release, freelancers in more AI-exposed occupations (e.g. writing) saw ~2% fewer contracts and ~5% lower monthly earnings, with larger losses among previously high-skilled workers (Hui, Reshef & Zhou 2024). Effects concentrate in entry-level and highly-automatable roles while aggregate US employment and wages show little disruption through 2024-2025 — so macro-level harm remains genuinely contested even as targeted-segment harm is established; much deployment to date augments rather than substitutes, raising novice productivity ~34% in call-center work (Brynjolfsson, Li & Raymond 2025).

    Sources: Hui, Reshef & Zhou 2024 ('The Short-Term Effects of Generative AI on Employment', Organization Science); Brynjolfsson, Li & Raymond 2025 ('Generative AI at Work', Quarterly Journal of Economics 140(2):889); Acemoglu 2024 ('The Simple Macroeconomics of AI', NBER WP 32487); Autor 2024 ('Applying AI to Rebuild Middle Class Jobs', NBER WP 32140)

  • Does governance work?evidence: thin

    There are essentially no impact evaluations of governance specifically targeting AI-driven displacement; current responses (OECD/GPAI guidance, reskilling initiatives, safety-net proposals) are at the recommendation stage, so 'does AI-displacement policy work' is answered only by extrapolation from the broader displaced-worker literature. That analogue base is robust but shows modest, mixed results: Card, Kluve & Weber's (2018) meta-analysis of 200+ active-labour-market evaluations finds training has small/insignificant short-run effects that improve only over the medium-to-long run, US Trade Adjustment Assistance evaluations find largely neutral-to-negative earnings effects (Schochet et al. 2012), and the JTPA randomized evaluation found weak earnings effects for the dislocated-worker stream. Recent syntheses note retraining yields smaller gains precisely when workers move into high-AI-exposure occupations — so the evidence that standard tools reduce AI-displacement harm is thin and early.

    Sources: Card, Kluve & Weber 2018 ('What Works? A Meta-Analysis of ... Active Labor Market Program Evaluations', JEEA 16(3):894); Schochet et al. 2012 (Trade Adjustment Assistance Program impacts, Mathematica/USDOL); Bloom et al. 1997 (National JTPA Study, Journal of Human Resources); Brookings 2025 ('AI Labor Displacement and the Limits of Worker Retraining'); OECD 2023-2025 Employment Outlook

• Biometric Identification

  • Is the harm real?evidence: established

    Demographic accuracy disparities in facial recognition are robust and replicated. NIST's Face Recognition Vendor Test (189 algorithms, 18.27M images) found one-to-one false-positive rates for Asian and African-American faces elevated 10-100x over white males, with the highest one-to-many false positives for African-American women; Buolamwini & Gebru's Gender Shades found commercial gender-classification error up to 34.7% for darker-skinned women vs 0.8% for lighter-skinned men. Documented downstream harm includes at least 8-15 US wrongful arrests, nearly all of Black people. Honest caveat: magnitude is highly algorithm-dependent — the most accurate algorithms show small or statistically undetectable differentials — so the harm is real but not uniform across systems.

    Sources: Grother, Ngan & Hanaoka 2019 (NISTIR 8280, FRVT Part 3: Demographic Effects); Buolamwini & Gebru 2018 (Gender Shades, PMLR 81); Hill 2020 / Williams v. City of Detroit (ACLU 2021)

  • Does governance work?evidence: thin

    Rigorous evidence that GOVERNANCE of biometric ID reduces the documented harms is sparse. The one quantitative impact evaluation of police facial-recognition policy (Johnson et al. 2024, difference-in-differences across 268 US cities) studies effects on violent crime — a crime-control outcome, not misidentification harm — from a single research group, and does not establish that any safeguard regime curbs wrongful identification. Direct evidence on procedural safeguards points the other way: in the known wrongful-arrest cases police are reported to have bypassed required corroboration/probable-cause standards, and the strongest documented enforcement levers are private-sector biometric-privacy laws — Illinois BIPA (e.g. Meta's $650M settlement) and the separate Texas CUBI law (a $1.4B Meta settlement) — which govern private actors, not the law-enforcement context where the arrests occur. No replicated study shows a specific regulatory regime measurably reduces demographic misidentification harm.

    Sources: Johnson et al. 2024 (Cities, 'Police facial recognition applications and violent crime control in U.S. cities'); Harwell & Schaffer 2025 (Washington Post, 'Arrested by AI'); Illinois BIPA (Rosenbach v. Six Flags 2019; Meta $650M settlement 2021); Texas CUBI (Meta $1.4B settlement 2024)

• Catastrophic & Existential Risk

  • Is the harm real?evidence: contested

    The catastrophic-uplift premise is genuinely contested: the empirical uplift studies that exist find current frontier models add little. RAND's red-team study found no statistically significant difference in the viability of bioweapon-attack plans produced with vs. without LLMs (Mouton, Lucas & Guest 2024), and OpenAI's 100-participant trial found GPT-4 gave at most a mild, non-significant accuracy uplift (mean +0.88 out of 10 for PhD experts, +0.25 for students; Patwardhan et al. 2024). Honest caveat: the harm is forward-looking, not yet observed — expert opinion on the catastrophic tail is sharply split (median AI researcher puts ~5% on extremely-bad/extinction outcomes, mean ~9-16% across differently-framed questions, n=2,778; Grace et al. 2024), and forecasters underestimated how fast risk-relevant capabilities (e.g. virology troubleshooting) actually arrived (Forecasting Research Institute 2025), so the relevant capabilities are a moving target rather than a settled magnitude.

    Sources: Mouton, Lucas & Guest 2024 (RAND RR-A2977-2, Operational Risks of AI in Large-Scale Biological Attacks: Results of a Red-Team Study); Patwardhan et al. 2024 (OpenAI, Building an Early Warning System for LLM-aided Biological Threat Creation); Grace et al. 2024 (Thousands of AI Authors on the Future of AI, arXiv:2401.02843); Forecasting Research Institute 2025 (Forecasting LLM-enabled Biorisk and the Efficacy of Safeguards)

  • Does governance work?evidence: absent

    There is essentially no impact evidence that catastrophic-risk governance reduces catastrophic risk, and structurally there cannot yet be: the harm is a low-probability civilisational tail event, so no controlled trial or before/after evaluation of a realised catastrophe is possible. The dominant instruments are recent, voluntary developer frameworks (Anthropic's Responsible Scaling Policy 2023; OpenAI's Preparedness Framework 2023) built on if-then capability thresholds the developers themselves describe as speculative and qualitative rather than validated risk thresholds. The closest evidence is adjacent and indirect: trained-in deceptive behaviours can persist through standard safety training (Hubinger et al. 2024) — a demonstration that current mitigation may be insufficient, not that any governance regime works — and Anthropic's documented loosening of earlier commitments (RSP 2025 dropped the original pledge to define higher-tier ASL evaluations before developing the corresponding models) illustrates that even the strongest voluntary regimes lack external enforcement or measured efficacy.

    Sources: Anthropic 2023 (Responsible Scaling Policy); OpenAI 2023 (Preparedness Framework); Hubinger et al. 2024 (Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training, arXiv:2401.05566); Hendrycks, Mazeika & Woodside 2023 (An Overview of Catastrophic AI Risks, arXiv:2306.12001)

• Compute + Model-Weight Export Controls

  • Is the harm real?evidence: contested

    That compute controls materially constrain China's frontier-AI hardware access is empirically real and measured: by mid-2025 the US hosted ~75% of catalogued AI-supercomputer performance versus China's ~15% (Pilz, Sanders, Rahman & Heim 2025), corroborated by the Federal Reserve's estimate of ~74% US vs ~14% China high-end AI compute share (Haag, FEDS Notes, Oct 2025), and US prosecutions document large diversion networks (e.g. the ~$160M Alan Hao Hsu / Hao Global H100/H200 case — the first 'AI diversion' conviction, guilty plea Oct 2025, SDTX). The honest caveat is that the magnitude of the binding constraint is genuinely contested: DeepSeek's V3/R1 reached near-frontier capability at an order of magnitude less reported compute via algorithmic efficiency, and analysts argue the controls have simultaneously accelerated Chinese state-backed indigenization, so whether the controls slow capability (the actual aim) versus merely shift its cost structure remains unsettled.

    Sources: Pilz, Sanders, Rahman & Heim 2025 (Trends in AI Supercomputers, arXiv:2504.16026 / Epoch AI) — VERIFIED, gives US ~75% / China ~15%; Haag 2025 (FEDS Notes, 'The State of AI Competition in Advanced Economies', Federal Reserve, 6 Oct 2025) — VERIFIED, gives US 74% / China 14% / EU 4.8% of high-end AI compute; US v. Alan Hao Hsu / Hao Global 2025 (DOJ, SDTX; ~$160M H100/H200 diversion, guilty plea Oct 2025, first AI-diversion conviction) — VERIFIED

  • Does governance work?evidence: absent

    There is no rigorous impact evaluation showing that compute or model-weight export controls achieve their stated strategic aim of durably slowing frontier-AI capability diffusion to China — the regime is too recent, the counterfactual is unidentified, and the most ambitious instrument (the Jan 2025 BIS 'Framework for AI Diffusion', ECCN 4E091 covering model weights of closed models trained on >10^26 operations) was rescinded on 12-13 May 2025 before it ever took effect (its enforcement date was 15 May 2025), so the evidence that the rule works is itself missing. The closest analogue evidence base, the economic-sanctions evaluation literature, is sobering: Hufbauer, Schott, Elliott & Oegg (2007) coded roughly a third of their historical cases as 'successful' (their database covers ~170-200 cases since WWI; the disputed coding was ~40 of 115, ~34%), but Pape's reanalysis (1997/1998) argued the genuinely sanctions-attributable success rate was far lower (he recoded it to ~5 of 115, under 5%), and the broader literature finds efficacy decays as targets adapt and substitute — the precise dynamic export-control critics attribute to Chinese indigenization and smuggling. This is an analogue, not direct evidence on export controls.

    Sources: Hufbauer, Schott, Elliott & Oegg 2007 (Economic Sanctions Reconsidered, 3rd ed., Peterson Institute for International Economics) — VERIFIED; Pape 1997/1998 (Why Economic Sanctions Do Not Work, International Security 22(2), 1997; Why Economic Sanctions Still Do Not Work, International Security 23(1), 1998) — VERIFIED; BIS 2025 (Framework for Artificial Intelligence Diffusion, Federal Register doc 2025-00636 / 90 FR, eff. 13 Jan 2025; ECCN 4E091 model-weight control; rescinded 12-13 May 2025 before its 15 May effective date) — VERIFIED

• Compute-Threshold Reporting

  • Is the harm real?evidence: contested

    Whether training-compute (FLOP) is a defensible proxy for governance-relevant capability is genuinely contested in the literature. The strongest empirical pressure against it is algorithmic efficiency: Ho, Besiroglu, Erdil et al. (2024) estimate the compute needed to reach a fixed language-model performance level has halved roughly every eight months (95% CI ~5-14 months, i.e. ~3x/year), so any static FLOP-to-capability mapping decays quickly; Hooker (2024) argues FLOP measures operations rather than end-performance, since techniques such as fine-tuning, retrieval, chain-of-thought and tool use can add large capability gains without proportional training compute, and Ord (2025) shows inference-time scaling further decouples deployed capability from training compute. Honest caveat: defenders (Heim & Koessler 2024; Pilz, Heim & Brown 2025) note compute remains the most quantifiable, externally verifiable, and ex-ante measurable correlate of frontier capability currently available, while themselves conceding it is an imperfect proxy that should not be used in isolation — the disagreement is about durability and precision, not whether any correlation exists.

    Sources: Ho, Besiroglu, Erdil, Owen, Rahman, Guo, Atkinson, Thompson & Sevilla 2024, Algorithmic progress in language models, NeurIPS 2024 (arXiv:2403.05812; Epoch AI); Hooker 2024, On the Limitations of Compute Thresholds as a Governance Strategy (arXiv:2407.05694); Ord 2025, Inference Scaling Reshapes AI Governance (arXiv:2503.05705); Heim & Koessler 2024, Training Compute Thresholds: Features and Functions in AI Regulation (arXiv:2405.10799); Pilz, Heim & Brown 2025, Increased Compute Efficiency and the Diffusion of AI Capabilities (AAAI 2025; arXiv:2311.15377)

  • Does governance work?evidence: absent

    There is no rigorous evidence that compute-threshold reporting reduces harm or achieves its stated aim, because the regimes have not produced an evaluable record. The US 10^26-FLOP reporting obligation (Executive Order 14110, invoking the Defense Production Act) was revoked on 20 January 2025 (by EO 14148) before its recurring binding reporting rule was finalized — the implementing BIS notice of proposed rulemaking (Sept 2024) never took effect, so no durable reporting record materialized; and the EU AI Act's 10^25-FLOP systemic-risk obligations for general-purpose models only became applicable on 2 August 2025 (with transitional periods into 2027), so no outcome evaluation yet exists. Moreover the 10^25 figure is a rebuttable presumption sitting alongside qualitative high-impact criteria (Art. 51(1)(a) and (2), rebuttable under Art. 52(2)), not a validated risk cutoff. The closest analogue is the broader regulatory-disclosure-mandate literature (Fung, Graham & Weil 2007), which documents that transparency policies' effects on outcomes are highly heterogeneous and frequently ineffective or counterproductive absent enforcement and downstream use — implying that the reporting trigger working as intended is an open empirical question, not a documented result.

    Sources: U.S. Executive Order 14110 (2023), Sec. 4.2 (10^26 FLOP, Defense Production Act); revoked by Executive Order 14148 (Jan 20, 2025); EU AI Act, Reg. (EU) 2024/1689, Art. 51 (10^25 FLOP systemic-risk rebuttable presumption; applicable Aug 2, 2025); Fung, Graham & Weil 2007, Full Disclosure: The Perils and Promise of Transparency (Cambridge University Press)

• AI in Criminal Justice

  • Is the harm real?evidence: contested

    Whether algorithmic risk assessment reproduces racial disparity is a genuine, partly mathematically irreducible dispute rather than merely an unresolved measurement question. ProPublica's analysis of COMPAS in Broward County found Black defendants who did not reoffend were nearly twice as likely to be flagged high-risk as comparable white defendants (44.9% vs 23.5% false-positive rate; Angwin et al. 2016), and Dressel & Farid (2018) showed COMPAS is no more accurate (65.2%) than untrained laypeople (67.0%); the developer's reanalysis (Flores, Bechtel & Lowenkamp 2016) found the same tool satisfies predictive parity and calibration across race. Honest caveat: Chouldechova (2017) proved both sides can be correct simultaneously — when recidivism base rates differ across groups, equal calibration and equal error rates cannot both hold, so the disagreement is partly definitional, not merely a data dispute to be settled.

    Sources: Angwin, Larson, Mattu & Kirchner 2016 (ProPublica, 'Machine Bias'); Dressel & Farid 2018 (Science Advances 4:eaao5580); Flores, Bechtel & Lowenkamp 2016 (Federal Probation 80(2):38); Chouldechova 2017 (Big Data 5(2):153)

  • Does governance work?evidence: thin

    Rigorous evidence that governing criminal-justice algorithms — mandating, auditing, or adopting risk tools — reduces the racial-disparity harm that motivates the rules is essentially absent. The leading real-world impact evaluation, Stevenson's (2018) study of Kentucky's mandatory pretrial risk-assessment law (>1M cases), found only a small increase in pretrial release that eroded as judges reverted to prior habits, with no reduction in racial disparities in pretrial detention. The closest analogue evaluations measure operational crime outcomes, not equity, and are largely null: Chicago's Strategic Subjects List had no effect on victimization (Saunders, Hunt & Hollywood 2016) and the only randomized predictive-policing trials tested crime reduction, not disparate impact (Mohler et al. 2015) — so the evidence that any governance regime measurably reduces algorithmic racial disparity is itself missing.

    Sources: Stevenson 2018 (Minnesota Law Review 103:303); Saunders, Hunt & Hollywood 2016 (Journal of Experimental Criminology 12(3):347); Mohler et al. 2015 (JASA 110(512):1399)

• Deepfakes / Synthetic Content

  • Is the harm real?evidence: established

    The flagship harm — non-consensual sexual deepfakes — is empirically real and sharply gendered: content audits find ~96-98% of deepfake videos online are non-consensual pornography overwhelmingly depicting women, and a pre-registered 10-country survey (>16,000 people) found 2.2% reporting victimization and 1.8% perpetration of synthetic intimate imagery, with documented mental-health, career, and participation harms. By contrast, the parallel claim that political/informational deepfakes UNIQUELY deceive is contested-to-refuted: experiments find deepfakes about as (not more) credible than equivalent text/audio fakes, and a 56-paper meta-analysis (k=137, N=86,155) puts unaided human detection near chance — implying a detection problem more than an exceptional-persuasion one.

    Sources: Umbach, Henry, Beard & Berryessa 2024 (CHI '24, 'Non-Consensual Synthetic Intimate Imagery ... in 10 Countries'); Diel et al. 2024 (Computers in Human Behavior Reports 16:100538, deepfake-detection meta-analysis of 56 papers); Barari, Lucas & Munger 2025 (Journal of Politics 87(2), 'Political Deepfakes Are as Credible as Other Fake Media'); Flynn et al. 2022 (British Journal of Criminology, multi-country image-based sexual abuse study)

  • Does governance work?evidence: thin

    Direct impact evidence that deepfake governance reduces the targeted harm is sparse and, where it exists, discouraging: the one quasi-experimental evaluation (Cuevas & Horta Ribeiro 2025, synthetic-control across three platforms) found the U.S. TAKE IT DOWN Act's passage plus the MrDeepfakes shutdown did NOT suppress synthetic non-consensual imagery — posting rose above counterfactual baselines and displaced elsewhere. Technical enforcement is likewise unreliable: detectors fail to generalize to unseen generators (notably diffusion models) and are vulnerable to adversarial evasion, with in-the-wild accuracy well below benchmark figures. No rigorous evaluation yet shows a deepfake-specific law, takedown mandate, or watermarking scheme producing a sustained reduction in prevalence or harm.

    Sources: Cuevas & Horta Ribeiro 2025 ('Deepfake Pornography is Resilient to Regulatory and Platform Shocks', arXiv:2602.02754); 'Adversarial Reality for Evading Deepfake Image Detectors' (ICCVW 2025); TAKE IT DOWN Act, S.146 / Pub. L. 119-12 (2025); CRS Legal Sidebar LSB11314

• AI in Education

  • Is the harm real?evidence: established

    The documented harms of educational AI are empirically real and, for proctoring, replicated: a controlled audit of a proctoring tool used by at least ~1,500 institutions found significantly higher facial-detection failure (the trigger for 'suspicious' flags) for darker-skinned and female test-takers (Yoder-Himes et al. 2022), and a technical audit of 164 government-endorsed pandemic learning products found 89% engaged in data practices that risk or infringe children's rights, with most monitoring happening without the child's knowledge or consent (Human Rights Watch 2022). Honest caveat: the benefit side is genuine but highly sensitive to how outcomes are measured rather than uniform — Kulik & Fletcher's meta-analysis of 50 intelligent-tutoring evaluations found an overall median effect of 0.66 SD, but the average effect was 0.73 SD on locally-developed tests versus only 0.13 SD on standardized tests, so much of AI education's apparent value depends on the outcome measure used.

    Sources: Yoder-Himes et al. 2022, 'Racial, skin tone, and sex disparities in automated proctoring software', Frontiers in Education 7:881449; Human Rights Watch 2022, 'How Dare They Peep into My Private Life?' (164 EdTech products endorsed by 49 governments; 89% risked/infringed children's rights); Kulik & Fletcher 2016, 'Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review', Review of Educational Research 86(1):42-78

  • Does governance work?evidence: thin

    There are essentially no rigorous impact evaluations showing that purpose-built governance of educational AI reduces the documented harms. The student-specific regime — California's SOPIPA (SB 1177, 2014, a model that more than 20 states adopted and ~33 considered) and the FTC's May 2022 COPPA ed-tech policy statement (which the agency itself said did not change existing requirements) — has near-zero documented enforcement and no published before/after evaluation of whether it changed vendor data practices or bias outcomes. The only documented remedies came not from education-specific rules but from generic legal levers: a $6.25M biometric-privacy class settlement under Illinois BIPA (Veiga v. Respondus, 2023) and a constitutional ruling that proctoring room-scans are an unreasonable search (Ogletree v. Cleveland State University, N.D. Ohio 2022, Calabrese J.) — neither of which is a replicable evaluation, and both reach private/state actors rather than the underlying demographic-bias harm.

    Sources: California SOPIPA (SB 1177, 2014); FTC Policy Statement on Education Technology and COPPA (adopted May 19, 2022); Veiga v. Respondus, Inc. ($6.25M BIPA class settlement, 2023; covers Illinois Respondus Monitor users Nov. 2015–June 2023); Ogletree v. Cleveland State University (N.D. Ohio 2022, Calabrese J., room-scan Fourth Amendment ruling)

• AI in Employment

  • Is the harm real?evidence: established

    Discrimination and adverse outcomes in employment decisions are empirically well-established, and AI systems demonstrably reproduce them. The foundational field-experiment literature shows robust human baseline discrimination (Bertrand & Mullainathan 2004 found White-sounding names received 50% more callbacks), and AI-specific audits confirm the pattern: Amazon scrapped a recruiting tool that penalized resumes containing 'women's' (Dastin 2018), and a controlled resume-screening audit of language-model retrieval found systems favored White-associated names ~85% of the time and never preferred Black male-associated over White male-associated names (Wilson & Caliskan 2024). On the monitoring side, a meta-analysis (k=94, N≈23,461) found electronic performance monitoring reliably raises worker stress with no evidence of improved performance (Ravid et al. 2023). Honest caveat: measured disparities are highly model-, prompt-, and context-dependent, and most evidence comes from controlled audits and one firm's internal test rather than measured outcomes in live, at-scale hiring pipelines.

    Sources: Bertrand & Mullainathan 2004 (American Economic Review 94(4):991-1013); Wilson & Caliskan 2024 (AAAI/ACM AIES; 'Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval'); Dastin 2018 (Reuters, 'Amazon scraps secret AI recruiting tool that showed bias against women'); Ravid, White, Tomczak & Behrend 2023 (Personnel Psychology 76:5-40)

  • Does governance work?evidence: absent

    There is no rigorous evidence that governing AI in employment reduces the documented harms; the central evaluated regime appears to fail at the compliance stage before any impact on bias can occur. NYC Local Law 144 — the first jurisdiction worldwide to mandate independent bias audits and public posting for automated employment decision tools — was directly studied across 391 employers and found to produce 'null compliance': the law's discretion makes it impossible to tell whether firms comply, with very few posting the required audits (Wright et al. 2024). Parallel qualitative work shows the audits themselves are undermined by missing demographic data, opaque aggregation, and 'test data' that does not reflect real use (Groves et al. 2024). No study links any AI-employment rule to a measured reduction in discriminatory hiring outcomes — the evidence that the rule works is itself missing, largely because mandated transparency artifacts (audit reports) are sparse, non-standardized, and unenforced.

    Sources: Wright, Muenster, Vecchione, Metcalf & Matias et al. 2024 ('Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability', ACM FAccT '24); Groves, Metcalf, Kennedy, Vecchione & Strait 2024 ('Auditing Work: Exploring the New York City algorithmic bias audit regime', ACM FAccT '24); Ravid, White, Tomczak & Behrend 2023 (Personnel Psychology 76:5-40, on monitoring outcomes as the closest analogue evaluation evidence)

• Environmental Impact of AI Training

  • Is the harm real?evidence: established

    The resource demands of AI compute are empirically documented at the model level: Strubell et al. (2019) quantified large-NLP training energy/carbon, Luccioni et al. (2023) estimated BLOOM's training at ~24.7 tCO2eq (dynamic power) rising to ~50.5 tCO2eq with manufacturing and deployment, Li et al. (2023) estimated GPT-3-scale training in US datacenters can evaporate on the order of hundreds of thousands of litres of freshwater (their central figure ~700,000 L), and Luccioni, Jernite & Strubell (2024) showed generative inference is markedly more energy-intensive per query than task-specific models; at the macro scale the IEA (2024) and de Vries (2023) document rapidly rising datacenter electricity demand. Honest caveat: absolute estimates vary by up to orders of magnitude with grid carbon intensity, hardware, utilisation and accounting boundaries, and cleanly attributing the AI-specific increment (versus general datacenter and crypto growth) remains genuinely contested — the IEA itself bundles AI with datacenters and crypto — so the existence of the footprint is established while its magnitude and trajectory are not.

    Sources: Strubell, Ganesh & McCallum 2019 (ACL Anthology P19-1355; 'Energy and Policy Considerations for Deep Learning in NLP'); Luccioni, Viguier & Ligozat 2023 (JMLR 24; BLOOM 176B carbon footprint, 24.7/50.5 tCO2eq; arXiv:2211.02001); Li, Yang, Islam & Ren 2023 (arXiv:2304.03271, 'Making AI Less Thirsty', later Comm. ACM 2025); Luccioni, Jernite & Strubell 2024 (ACM FAccT '24, 'Power Hungry Processing', DOI 10.1145/3630106.3658542); de Vries 2023 (Joule 7(10):2191-2194, DOI 10.1016/j.joule.2023.09.004); IEA 2024 (Electricity 2024)

  • Does governance work?evidence: absent

    There is no impact evaluation showing that any AI-specific environmental-governance instrument reduces energy, water or carbon use, because every named instrument is voluntary or non-binding and very recent: EU AI Act Art. 95 codes of conduct are explicitly optional with no sanctions, and NIST AI 600-1 and the G7 Hiroshima Code are guidance, not enforceable caps. The closest analogue evaluation literature is divided in a way that disfavours the voluntary form chosen here: rigorous reviews find voluntary environmental programs generally fail to produce significant abatement beyond business-as-usual (Koehler 2007; Morgenstern & Pizer 2007), whereas the one form with credible positive evidence is mandatory disclosure (Downar et al. 2021 found a UK carbon-reporting mandate cut emissions ~8% versus a control group) which the AI instruments do not yet impose, leaving the proposition that AI environmental governance works essentially untested.

    Sources: EU AI Act Art. 95 / Recital 142 (Reg. (EU) 2024/1689); NIST AI 600-1 (2024, GenAI Profile); G7 Hiroshima Process International Code of Conduct (30 Oct 2023); Koehler 2007 (Policy Studies Journal 35(4):689-722); Morgenstern & Pizer (eds.) 2007 (Reality Check, RFF Press); Downar, Ernstberger, Reichelstein, Schwenen & Zaklan 2021 (Review of Accounting Studies 26(3):1137-1175)

• Foundation Models / GPAI

  • Is the harm real?evidence: contested

    Whether the foundation-model category maps to a coherent capability/risk tier is genuinely contested. The original case rests on scale-driven 'emergent abilities' that appear unpredictably above a size threshold (Wei et al. 2022; Ganguli et al. 2022 documented capabilities that are smoothly predictable in aggregate loss yet locally surprising), but Schaeffer, Miranda & Koyejo (2023, a NeurIPS Outstanding Paper) showed many 'emergent' jumps are artefacts of discontinuous metrics and dissolve under linear/continuous scoring — implying capability scales more smoothly than a sharp tier would suggest. Honest caveat: this is a live empirical disagreement about measurement, not a settled finding either way, and compute (the regulatory proxy) is an imperfect stand-in for capability or risk regardless of which side is right.

    Sources: Wei et al. 2022 (Emergent Abilities of Large Language Models, TMLR; arXiv:2206.07682); Schaeffer, Miranda & Koyejo 2023 (Are Emergent Abilities of Large Language Models a Mirage?, NeurIPS 2023, Outstanding Paper; arXiv:2304.15004); Ganguli et al. 2022 (Predictability and Surprise in Large Generative Models, ACM FAccT; DOI 10.1145/3531146.3533229)

  • Does governance work?evidence: absent

    There is no impact evaluation showing that GPAI/foundation-model governance reduces harm — the rules are too new (EU AI Act GPAI obligations and the 10^25-FLOP systemic-risk presumption only began binding on 2 August 2025) and the central regulatory lever is itself contested: Hooker (2024) argues compute thresholds are a shortsighted proxy because compute does not reliably track capability or risk, and the thresholds already diverge across jurisdictions (EU 10^25 vs. the now-rescinded US EO 14110's 10^26 operations, rescinded 20 January 2025). The mandated mitigation methods also lack validated efficacy: model evaluation and red-teaming face well-documented coverage limits and an 'audit gap' in the survey/position literature (behavioural testing cannot establish the absence of untested failure modes), and adversarial red-teaming repeatedly defeats deployed safeguards — the UK AI Safety Institute reports finding universal jailbreaks for every frontier system it has tested, and a large public agent-injection competition elicited policy violations across all 22 frontier models tested from ~1.8M attacks (Zou et al. 2025). Even compliant evaluation therefore cannot yet certify the safety the rules demand. (Caveat: this is an absence-of-evidence claim — no efficacy study has been done — not evidence the rules are ineffective.)

    Sources: Hooker 2024 (On the Limitations of Compute Thresholds as a Governance Strategy, arXiv:2407.05694); EU AI Act Arts. 51 & 55 (GPAI systemic-risk presumption, 10^25 FLOP; binding 2 Aug 2025); US EO 14110 (10^26-operation reporting threshold, rescinded 20 Jan 2025 by EO 14148); Zou et al. 2025 (Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition / Gray Swan Arena, arXiv:2507.20526 — 22 frontier agents, ~1.8M attacks); UK AI Safety/Security Institute, Frontier AI Trends Report (universal jailbreaks for every system tested); METR, Common Elements of Frontier AI Safety Policies (2024)

• AI in Healthcare

  • Is the harm real?evidence: established

    Both the benefit and the harm of clinical AI are empirically real and well-documented, but outcomes are highly deployment-dependent. Rigorous prospective studies show genuine clinical value in narrow tasks — the MASAI RCT (>100,000 women) found AI-supported mammography detected ~20% more cancers (6.1 vs 5.1 per 1000 screened) at comparable recall rates (Lang et al. 2023, Lancet Oncology), and IDx-DR's pivotal trial achieved 87.2% sensitivity / 90.7% specificity for diabetic retinopathy (Abramoff et al. 2018, npj Digital Medicine) — yet widely deployed models can fail or harm: the Epic Sepsis Model, live at hundreds of US hospitals, scored AUC 0.63 with 33% sensitivity on external validation (Wong et al. 2021, JAMA Internal Medicine), and a population-health algorithm covering ~200M people understated Black patients' illness because it predicted cost not need (Obermeyer et al. 2019, Science). Honest caveat: there is no single 'AI in healthcare' effect — performance ranges from life-saving to dangerous depending on task, calibration, and whether the model was prospectively validated.

    Sources: Lang K, Josefsson V, Larsson A-M, et al. 2023 (Lancet Oncology 24(8):936-944, MASAI trial clinical safety analysis; AI-supported screening detected 6.1 vs 5.1 cancers per 1000, ~20% higher, similar recall rates); Abramoff MD, Lavin PT, Birch M, Shah N, Folk JC. 2018 (npj Digital Medicine 1:39, IDx-DR pivotal trial; 87.2% sensitivity / 90.7% specificity); Wong A, Otles E, Donnelly JP, et al. 2021 (JAMA Internal Medicine 181(8):1065-1070, Epic Sepsis Model external validation; AUC 0.63, 33% sensitivity); Obermeyer Z, Powers B, Vogeli C, Mullainathan S. 2019 (Science 366(6464):447-453, racial bias from cost-as-proxy)

  • Does governance work?evidence: absent

    There is essentially no impact-evaluation evidence that the prevailing governance regime for medical AI — FDA authorization, predominantly via the 510(k) substantial-equivalence pathway — measurably reduces patient harm or improves outcomes. Analyses of authorized AI devices find that clinical validation is frequently absent or non-prospective (of 521 FDA-authorized AI devices, ~43% had no published clinical-validation data and only ~28% were prospectively validated; Chouffani El Fassi & Henderson et al. 2024) and that demographic performance is almost never reported (race/ethnicity in 3.6%, and only 9.0% of 692 510(k)/cleared AI devices carried a prospective post-market-surveillance study; Muralidharan et al. 2024). Earlier analysis of 130 cleared devices likewise found 97% were evaluated only retrospectively (Wu et al. 2021). The closest analogue evidence on the pathway itself is discouraging: the Institute of Medicine (2011) concluded the 510(k) process was not designed to assess safety and effectiveness — i.e., no direct study establishes that the rule, as written, prevents the harms it targets. Caveat: this is an absence of impact evaluation plus reporting-gap and design-critique evidence, not a study showing the regime fails to reduce harm.

    Sources: Chouffani El Fassi S, Abdullah A, Fang Y, ... Henderson GE, et al. 2024 (Nature Medicine, 'Not all AI health tools with regulatory authorization are clinically validated', s41591-024-03203-3; 521 devices, ~43% no clinical validation, ~28% prospectively validated); Muralidharan V, Adewale BA, Huang CJ, et al. 2024 (npj Digital Medicine 7:273, scoping review of reporting gaps in 692 FDA-approved AI medical devices; race/ethnicity 3.6%, prospective post-market surveillance 9.0%); Wu E, Wu K, Daneshjou R, Ouyang D, Ho DE, Zou J. 2021 (Nature Medicine 27:582-584, analysis of 130 FDA approvals; 97% retrospective-only evaluation); Institute of Medicine 2011 (Medical Devices and the Public's Health: The FDA 510(k) Clearance Process at 35 Years)

• National Security Carveouts in AI Regulation

  • Is the harm real?evidence: thin

    That civilian AI-governance instruments carve out national-security uses is black-letter and undisputed (EU AIA Art. 2(3); CoE Framework Convention Art. 3(2) on national-security activities, distinct from Art. 3(4) on national defence; US NSM-25 (Oct. 2024) as the national-security-track instrument fulfilling §4.8 of EO 14110); civil-society legal analysis argues a blanket exclusion is harder to square with a necessity-and-proportionality approach than a qualified one (Korff/ECNL 2022; Vogiatzoglou 2024). But whether the carveout itself produces concrete unredressed harm is empirically under-observed almost by construction — the secrecy it confers suppresses the very evidence needed to measure it. The closest analogue, national-security deference in the courts, shows the mechanism is real (the FISC granted all but eleven of 33,900 applications 1979-2012, a 99.97% approval rate; Sinnar 2022 documents downstream harms to securitized communities), yet Clarke (2014) shows that lopsided ex parte approval rates alone do not prove rubber-stamping, because rational case selection and pre-vetting produce similar rates in ordinary Title III wiretaps (99.93%) and delayed-notice warrants (99.6-99.8%) — so the magnitude of harm attributable to the carveout, as opposed to the legitimate secrecy of the domain, remains genuinely contested.

    Sources: Korff 2022 (ECNL Opinion on the implications of the exclusion of national security from AI legislation, Oct. 2022); Sinnar 2022 (Harvard Law Review Forum 136:59, 'A Label Covering a "Multitude of Sins": The Harm of National Security Deference'); Clarke 2014 (Stanford Law Review Online 66:125, 'Is the Foreign Intelligence Surveillance Court Really a Rubber Stamp?'); EPIC FISC statistics 1979-2012

  • Does governance work?evidence: absent

    There is no impact evaluation showing that any specific design of the national-security carveout — categorical exclusion versus parallel governance track versus civilian-compliance-with-override — measurably improves oversight or reduces harm relative to the alternatives; the question is argued doctrinally (Vogiatzoglou 2024; Korff/ECNL 2022) but has never been tested empirically. The closest analogue evaluation literature is on the parallel-track model already in use for intelligence surveillance (the FISC / FISA oversight regime), and even there the evidence that the mechanism delivers effective scrutiny is itself contested rather than established (Clarke 2014; Sinnar 2022). No direct evaluation exists because the carveouts are recent (EU AIA 2024, CoE Framework Convention 2024, US NSM-25 2024), enforcement actions are by design non-public, and private parties typically lack standing to challenge a specific exempt deployment — the structural features that make the harm hard to observe also make the governance impossible to evaluate.

    Sources: Vogiatzoglou 2024 (Verfassungsblog, 'The AI Act National Security Exception: room for manoeuvres?', 9 Dec. 2024); Korff 2022 (ECNL Opinion, exclusion of national security from AI legislation); Clarke 2014 (Stanford Law Review Online 66:125); Sinnar 2022 (Harvard Law Review Forum 136:59)

• Open-Weight Frontier Release

  • Is the harm real?evidence: contested

    The empirical picture splits into two well-separated questions. (1) The MECHANISM that distinguishes open-weight release — that safety guardrails can be cheaply and irreversibly stripped once weights are public — is established: Qi et al. (2024) removed GPT-3.5 Turbo safety alignment by fine-tuning on only ~10 adversarially designed examples for under $0.20 (and the attack generalizes to Llama-2), and even purpose-built tamper-resistant safeguards (Tamirisa et al. 2025, TAR) were subsequently shown to be defeatable by adaptive fine-tuning (Qi et al. 2024, durability critique). (2) Whether this mechanism produces real-world CATASTROPHIC uplift is genuinely contested and, for the headline biosecurity case, currently unsupported: RAND's red-team study found no statistically significant difference in the viability of bioweapon attack plans produced with versus without LLM assistance (Mouton, Lucas & Guest 2024), and OpenAI's 100-participant trial found at most mild uplift over an internet baseline (Patwardhan et al. 2024). Honest caveat: these null/mild results are time-stamped to 2023-2024 frontier capability and to biothreats specifically; the marginal-risk framework (Kapoor, Bommasani et al. 2024) concludes the evidence base is too thin to characterize marginal risk across most misuse vectors, so 'no measured harm yet' is not 'no harm.'

    Sources: Kapoor, Bommasani, Klyman, Longpre et al. 2024, 'Position: On the Societal Impact of Open Foundation Models', PMLR 235 / ICML 2024 (arXiv 2403.07918); Mouton, Lucas & Guest 2024, RAND RR-A2977-2, 'The Operational Risks of AI in Large-Scale Biological Attacks: Results of a Red-Team Study'; Qi, Zeng, Xie, Chen, Jia, Mittal & Henderson 2024, 'Fine-tuning Aligned Language Models Compromises Safety', ICLR 2024 (arXiv 2310.03693); Tamirisa et al. 2025, 'Tamper-Resistant Safeguards for Open-Weight LLMs', ICLR 2025 (arXiv 2408.00761); Qi, Wei, Carlini, Huang, Xie, He, Jagielski, Nasr, Mittal & Henderson 2024, 'On Evaluating the Durability of Safeguards for Open-Weight LLMs' (arXiv 2412.07097); Patwardhan et al. 2024, 'Building an early warning system for LLM-aided biological threat creation', OpenAI

  • Does governance work?evidence: absent

    There is no impact evaluation showing that any specific weight-release governance regime reduces downstream harm, because no binding regime has been implemented and measured: California SB-1047's release-conditioning framework was vetoed in September 2024, and the EU AI Act's open-source carve-outs (Recital 102, Art. 53(2)) exempt most open-weight models (those below the systemic-risk compute threshold) from the documentation obligations that would generate evaluable conduct. The structural obstacle is also documented: Kapoor, Bommasani et al. (2024) characterize open-weight release as effectively irreversible and poorly monitorable once weights are public, so post-release governance has little to act on. The closest analogue evidence — technology export controls — is mixed and points to circumvention: commentators argue blanket export controls on freely copyable open-source models cannot work (Just Security 2024), and independent analyses of the post-2022 semiconductor controls document displacement to less-regulated channels (smuggling, threshold-tuned chip variants, cloud access) rather than disappearance of activity (e.g., CSIS, FPRI 2024), suggesting recipient-restriction regimes face the same leakage problem for weights. (Caveat: this is analogical, not direct evidence about weight-release governance, which remains unmeasured.)

    Sources: Kapoor, Bommasani, Klyman, Longpre et al. 2024, 'Position: On the Societal Impact of Open Foundation Models', PMLR 235 (arXiv 2403.07918); California SB-1047 (2024, vetoed by Gov. Newsom 29 Sep 2024); EU AI Act Regulation (EU) 2024/1689, Recital 102 & Art. 53(2) open-source exemptions; Just Security 2024, 'Export Controls on Open-Source Models Will Not Win the AI Race'; CSIS, 'The Limits of Chip Export Controls in Meeting the China Challenge' and FPRI 2024, 'Breaking the Circuit: US-China Semiconductor Controls' (export-control circumvention analogue)

• Sovereign AI Doctrine

  • Is the harm real?evidence: thin

    Sovereign-AI doctrine is post-2023 and largely aspirational, so its core empirical premise — that frontier model deployment can be meaningfully bound to a national jurisdiction — is only just beginning to be tested. What IS measurable is the underlying compute geography the doctrine reacts to: an audit of 775 non-U.S. data-center projects estimates U.S. companies operate ~48% of them when weighted by investment value (a proxy for compute capacity, and explicitly an initial public-data approximation), implying 'in-territory' hardware is frequently still subject to foreign corporate/legal control (Richardson et al. 2025). Honest caveat: there is no peer-reviewed evidence base establishing whether jurisdiction-bound frontier deployment is technically feasible at scale — the descriptive dependency (foreign operation of locally-sited hardware) is documented, but the doctrine's central feasibility claim is thin and early.

    Sources: Richardson et al. 2025 (arXiv:2508.00932, 'How Sovereign Is Sovereign Compute? A Review of 775 Non-U.S. Data Centers'); Gupta, Walker & Reddie 2024 (arXiv:2411.14425, 'Whack-a-Chip: The Futility of Hardware-Centric Export Controls', UC Berkeley Risk & Security Lab)

  • Does governance work?evidence: absent

    There is no rigorous impact evaluation showing that sovereign-AI governance achieves its stated aim of secure, contained national AI capability. The closest direct levers have measurable but mostly adverse or contested evidence: ex-ante simulations of the closest analogue — data-localization mandates — project GDP losses (EU GDP −0.4% under proposed/GDPR-style measures rising to −1.1% under economy-wide localization; Bauer, Lee-Makiyama, van der Marel & Verschelde 2014, ECIPE Occasional Paper No. 3/2014) yet quantify no realized sovereignty benefit, and chip export controls — the other main instrument — show contested efficacy: one cross-firm study finds no innovation harm to 30 leading semiconductor firms (Schumacher 2024, CSIS) while case evidence documents systematic circumvention via software/efficiency gains and chip exfiltration/smuggling (Gupta, Walker & Reddie 2024). No replicated study demonstrates that any sovereign-AI regime measurably delivers the jurisdictional control it asserts.

    Sources: Bauer, Lee-Makiyama, van der Marel & Verschelde 2014 (ECIPE Occasional Paper No. 3/2014, 'The Costs of Data Localisation: Friendly Fire on Economic Recovery'); Schumacher 2024 (CSIS, 'Did U.S. Semiconductor Export Controls Harm Innovation?'); Gupta, Walker & Reddie 2024 (arXiv:2411.14425, 'Whack-a-Chip: The Futility of Hardware-Centric Export Controls')

• Synthetic Content Provenance

  • Is the harm real?evidence: contested

    The harm provenance targets is real but concentrated, and the technical premise that the mandated signal survives is itself empirically shaky. Synthetic-media harm is well documented in two domains: non-consensual intimate imagery (Ajder et al.'s 2019 Deeptrace audit found 96% of deepfake videos were pornographic and effectively 100% targeted women) and impersonation fraud (the Arup case, ~US$25.6M / HK$200M lost via a deepfake video call). The honest caveat is twofold: a feared broad political-misinformation harm is not yet demonstrated at scale, and CS work shows invisible watermarks are removable in practice (Jiang, Zhang & Gong 2023, WEvade, evade detection via adversarial perturbation; Zhao et al. 2024 prove pixel-level watermarks are provably removable via regeneration attacks), so the provenance signal a rule would mandate is itself contested.

    Sources: Ajder, Patrini, Cavalli & Cullen 2019 (Deeptrace, 'The State of Deepfakes: Landscape, Threats, and Impact'); Jiang, Zhang & Gong 2023 ('Evading Watermark based Detection of AI-Generated Content', ACM CCS 2023); Zhao et al. 2024 (NeurIPS, 'Invisible Image Watermarks Are Provably Removable Using Generative AI'); Arup deepfake fraud (CNN Business, 2024-05-16, US$25.6M)

  • Does governance work?evidence: thin

    There is no impact evaluation showing that mandated provenance/labeling reduces synthetic-media harm; the major mandates (China's GenAI labeling Measures, effective 2025-09-01; EU AIA Art. 50, machine-readable marking) are too new and unevaluated, and the delivery layer is leaky: the C2PA spec's own Security Considerations document the strip-and-repost threat, and platform audits report C2PA/Content-Credentials metadata is stripped by essentially all major social platforms on upload (consistent with Imatag's 2018 finding that ~80% of uploaded images lose metadata, only ~15% retaining it). The closest analogue evaluation literature — Pennycook, Bear, Collins & Rand (2020), the 'implied truth effect' — gives reason for caution rather than confidence: labeling only some content can make unlabeled false content seem more credible, so a partial-coverage provenance regime could backfire.

    Sources: Pennycook, Bear, Collins & Rand 2020 (Management Science 66(11):4944-4957, 'The Implied Truth Effect'); China Measures for Labeling AI-Generated Synthetic Content (eff. 2025-09-01); EU AI Act Art. 50; Imatag 2018 metadata-stripping study (~80%); C2PA Security Considerations (spec.c2pa.org) on manifest removal

• Technological Sovereignty

  • Is the harm real?evidence: contested

    The structural fact that compute capacity is geographically concentrated is well-measured: Lehdonvirta, Wú & Hawkins find only ~33 countries host facilities with AI-accelerator hardware and roughly 24 have the capacity to train full-scale foundation models, the Stanford AI Index 2026 reports low-income countries collectively hold ~0.1% of global data-centre compute (the US hosting >10x any other nation), and Cottier et al. document amortized frontier-training cost rising 2.4x/year (95% CI 2.0-3.1x) toward $1B+ models by 2027. But this is a political-economy FRAME, not a documented harm, and the core contested claim of the topic, that the cost curve locks mid-sized economies OUT of capability, is empirically cut both ways: a feasibility study of Brazil and Mexico (Malagon et al. 2025) estimates usable (non-frontier) 10-trillion-token sovereign models are fiscally viable at roughly $8-14M on H100 hardware, and DeepSeek-style efficiency gains (V3 trained for ~$5.5M, ~11x less compute than Llama 3 405B) show frontier-adjacent performance at a fraction of prior compute, so whether domestic frontier-tier capability is foreclosed for middle powers remains genuinely unsettled.

    Sources: Lehdonvirta, Wú & Hawkins 2024 (Compute North vs. Compute South, Proceedings of the 2024 AAAI/ACM Conference on AI, Ethics & Society 7:828-838); Cottier, Rahman, Fattorini, Maslej & Owen 2024 (The Rising Costs of Training Frontier AI Models, arXiv:2405.21015); Stanford AI Index 2026 (Maslej et al., Stanford HAI); Malagon, Ulloa Ruiz, Sandoval Plaza, Rosario Bolívar, García Mesa & Alvarado Morales 2025 (The Feasibility of Training Sovereign Language Models in the Global South: A Study of Brazil and Mexico, arXiv:2510.19801)

  • Does governance work?evidence: absent

    There is no rigorous impact evaluation showing that technological-sovereignty policies (on-shore compute mandates, national foundation-model champions, talent-retention schemes such as EuroHPC AI Factories or India's IndiaAI Mission) actually deliver sustained domestic capability or strategic autonomy; these programs are recent, utilization and cost-per-GPU-hour are largely unpublished, and no counterfactual study exists. The closest analogue evidence base, the industrial-policy literature synthesized by Juhász, Lane & Rodrik, finds that properly-identified studies are more favorable than older correlational work suggested but that outcomes depend heavily on instrument design and structural context, and the older national-champion record warns of subsidized 'zombie' firms and government capture, so the closest analogue is mixed and the direct evidence that the sovereignty rule works is simply missing.

    Sources: Juhász, Lane & Rodrik 2024 (The New Economics of Industrial Policy, Annual Review of Economics 16:213-242); Ahmed & Wahed 2020 (The De-democratization of AI: Deep Learning and the Compute Divide in Artificial Intelligence Research, arXiv:2010.15581); IndiaAI Mission (Indian Cabinet, March 2024); EuroHPC Joint Undertaking AI Factories (2024 regulation amendment; no published impact evaluation)

• Transparency Obligations

  • Is the harm real?evidence: contested

    Documentation artifacts (model cards, datasheets) are well-specified as proposals and are genuinely adopted, but the empirical premise that mandated disclosure produces meaningful transparency is contested. Selbst & Barocas (2018) argue inscrutability and non-intuitiveness are distinct problems and that disclosing rules does not resolve the latter, and large-scale audits find documentation is sparsely and unevenly completed: a systematic analysis of 32,111 Hugging Face model cards (Liang et al. 2024) found environmental-impact, limitations and evaluation sections least often filled, and Bhat et al. (2023, 45 practitioners) found a substantial gap between the documentation proposal and actual practice. Honest caveat: the documentation frameworks themselves are real and adopted, so the dispute is about whether disclosure conveys decision-relevant information, not whether the artifacts exist.

    Sources: Selbst & Barocas 2018 (Fordham Law Review 87:1085-1139); Liang et al. 2024 (Nature Machine Intelligence, s42256-024-00857-z, 'Systematic analysis of 32,111 AI model cards'); Bhat et al. 2023 (CHI '23, 'Aspirations and Practice of ML Model Documentation', DOI 10.1145/3544548.3581518); Mitchell et al. 2019 (FAccT, Model Cards for Model Reporting); Gebru et al. 2021 (CACM 64(12):86-92, Datasheets for Datasets)

  • Does governance work?evidence: absent

    There is no rigorous impact evaluation showing that AI transparency mandates (model cards, training-data summaries) measurably reduce bias, misuse or accidents — the central regulatory assumption is empirically untested, partly because flagship mandates like EU AI Act Art. 53(1)(d) GPAI training-data summaries are only subject to AI Office enforcement/verification from 2 August 2026 (the obligation itself began 2 August 2025 for new models). The closest analogue, mandated consumer disclosure, shows small and context-dependent effects: Bollinger, Leslie & Sorensen (2011) found mandatory calorie posting cut average calories per transaction by about 6%, while Loewenstein, Sunstein & Golman (2014) review evidence that disclosure effects are frequently diminished or even reversed by limited attention and often change provider rather than recipient behavior. These are analogues, not AI studies; no study demonstrates that AI transparency disclosure achieves its stated downstream safety aims.

    Sources: Bollinger, Leslie & Sorensen 2011 (AEJ: Economic Policy 3(1):91-128); Loewenstein, Sunstein & Golman 2014 (Annual Review of Economics 6:391-419, 'Disclosure: Psychology Changes Everything'); EU AI Act Art. 53(1)(d) GPAI training-data summary (obligation from 2 Aug 2025; AI Office enforcement from 2 Aug 2026)

Related debates — rival interpretations & counterevidence

Structured controversies where this instrument's provisions are the locus of disagreement. Each debate page lays out the competing positions with primary-source citations.

• Pause AI vs Accelerate Capabilities — Should the global community impose temporary or capability-conditional pauses on frontier-AI development, or should development accelerate with safety work conducted in parallel?
• Pre-Deployment Red-Team vs Post-Deployment Audit — Should AI capability + safety evaluations happen primarily before deployment (red-team gating release), or primarily after (post-deployment audit + incident response)?
• Risk-Based vs Principles-Based vs Ex-Post Liability Regimes — Should AI governance work via (a) risk-based ex-ante categorisation + obligations (EU), (b) high-level principles delegated to sector regulators (UK / OECD / G7), or (c) ex-post liability + civil litigation (US sectoral)?
• Compute vs Behavioural Capability Thresholds — Should the regulatory trigger for 'frontier' / 'foundation' / 'systemic-risk' status be training-compute thresholds (objective + ex-ante observable), or behavioural-capability evaluation (more semantically meaningful but operationally costly)?

Related instruments

• EU AI Act · EU
• Interim Measures for Generative AI Service Management · CN
• G7 Hiroshima AI Process Code of Conduct · G7
• Council of Europe Framework Convention on AI · council_of_europe
• Bletchley Declaration on AI Safety · global
• Seoul Declaration on Safe, Innovative and Inclusive AI · global
• NIST AI RMF Generative AI Profile · US
• California SB-1047: Safe and Secure Innovation for Frontier AI Models Act · US
• Brazil AI Bill (PL 2338/2023) · BR
• Anthropic Responsible Scaling Policy (RSP) v2 · US
• OpenAI Preparedness Framework · US
• Google DeepMind Frontier Safety Framework · US
• Meta Frontier AI Framework · US
• White House Voluntary AI Commitments · US
• Singapore Model AI Governance Framework for Generative AI · SG
• EU General-Purpose AI Code of Practice · EU
• GSA Generative AI and Specialized Computing Infrastructure Acquisition Resource Guide · US
• California SB-53: Transparency in Frontier Artificial Intelligence Act (TFAIA) · US
• Provisions on the Administration of Deep Synthesis of Internet Information Services · CN
• New York RAISE Act: Responsible AI Safety and Education Act · US
• Italy Law No. 132/2025 on Artificial Intelligence (Legge 23 settembre 2025, n. 132) · IT