The Information Machine

US AI Regulation: Federal Retreat vs. State Intervention · history

Version 26

2026-06-27 02:22 UTC · 358 items

What

The Trump administration has asked OpenAI to release GPT-5.6 as a controlled, staggered preview with government-approved access on a customer-by-customer basis, citing cybersecurity concerns about the model's autonomous cyberattack capabilities [11]. The administration is simultaneously pressing Meta — the only major US lab outside its voluntary review framework — toward participation, with the White House saying 'We hope to sign the agreement soon' [9]. State governments continue advancing binding AI law independently: Illinois has enacted mandatory audit requirements [19], a 42-state coalition has subpoenaed OpenAI over consumer practices [22], and Florida's lawsuit now names CEO Sam Altman personally [23].

Why it matters

The GPT-5.6 customer-by-customer access approval, if implemented, would move the administration beyond pre-release review into active post-deployment access management — a more granular form of control with no statutory basis. Combined with the Anthropic Fable export restriction and Meta pressure, the administration is constructing a coercive oversight regime for frontier AI through executive action alone, while simultaneously opposing binding domestic consumer safety requirements.

Open questions

  • Will the customer-by-customer GPT-5.6 access approval [11] set a precedent that other frontier labs face for models with dual-use cybersecurity capabilities?

  • Will Meta comply with the White House's 'We hope to sign the agreement soon' framing [9], and will it negotiate any limits on what the government can require during review?

  • Sacks reportedly plans to reframe his Anthropic critique around Mythos-model security needs [17] — does this represent a shift in the administration's AI safety doctrine or a communications adjustment?

  • Will the 42-state subpoena against OpenAI [22] and Florida's personal lawsuit against Altman [23] converge on coordinated settlement terms or remain parallel enforcement tracks?

Narrative

The Trump administration signed an AI executive order on June 3, 2026, establishing a voluntary 30-day pre-release review for frontier models with NSA oversight and classified compliance thresholds, explicitly disclaiming binding requirements [1][2][3]. The order's limited form reflects lobbying by Musk, Zuckerberg, and David Sacks, who stripped binding requirements and cut the review window from 90 to 30 days [4][5]. In practice, the administration has applied the framework coercively: it forced Anthropic to disable its Fable-class models for all foreign nationals [6][7], directed the DOJ to protect xAI's Mississippi data center as national security infrastructure [8], and is pushing Meta — the only major US lab not inside the review system — to join, with the White House saying 'We hope to sign the agreement soon' [9]. Former Google CEO Eric Schmidt provided the clearest public articulation of the underlying doctrine: the US controls approximately 75% of global AI compute versus China's 15%, a gap he attributes to three years of coordinated export controls [10].

The most recent expansion of this posture came in late June 2026, when the US government asked OpenAI to release GPT-5.6 as a controlled, staggered preview rather than a public launch, with the government approving access customer-by-customer [11]. Sam Altman told OpenAI staff that the primary concern is the model's capability for automated high-skill cyber work — the same capabilities that help defenders find software vulnerabilities also help attackers test exploits faster [11]. This customer-level access approval moves the administration's oversight beyond pre-release review into active post-deployment access management, a step not contemplated in the June 3 executive order.

The central public argument over federal AI regulation runs between Anthropic CEO Dario Amodei and the administration's accelerationist wing. Amodei called for mandatory pre-release testing and government-backed independent auditing [12][13]. White House AI czar David Sacks accused Anthropic of 'a sophisticated regulatory capture strategy based on fear-mongering' [14], a characterization VP Vance echoed publicly [15]. Bloomberg reported that Anthropic's AI principles have made it a White House target [16], consistent with the Fable export control action. Sacks is reportedly now seeking to reframe his critique around security needs tied to Mythos-class models [17]. Separately, legal theorist Bruce Schneier has argued that AI systems should be treated as agents of the organizations that deploy them, extending employer liability doctrine to AI errors [18] — a framework closer to the liability theories state AGs are currently advancing than to the administration's national-security-first approach.

State governments advance AI regulation independent of federal coordination. Illinois enacted SB 315 requiring annual third-party audits, 72-hour incident reporting, and whistleblower protections [19]. California is advancing SB 947 and SB 951 requiring human oversight of automated employer discipline and 90-day advance notice of AI-driven layoffs [20][21]. New York AG Letitia James served OpenAI with a subpoena on behalf of a 42-state coalition targeting advertising practices, data handling, treatment of minors, and chatbot sycophancy [22]; Florida expanded its suit to 83 pages naming CEO Sam Altman personally [23]. Trump's preemption executive order challenged state AI laws on constitutional grounds [24], but multiple law firms found its authority limited absent congressional action [25]. OpenAI has publicly advocated for coherent federal governance over fragmented state regulation [26][27] while its PAC admitted running false-flag social media accounts [28][29].

Timeline

  • 2026-05: Trump signs preemption EO challenging state AI laws; multiple law firms find constitutional authority limited absent congressional authorization [24][59][25]
  • 2026-05: California Senate advances SB 947 requiring human oversight of automated employer discipline and SB 951 requiring 90-day advance notice of AI-driven layoffs [21][42][20][43]
  • 2026-05-21: Trump cancels AI security EO after Sacks calls Trump directly to oppose pre-deployment review; Musk and Zuckerberg also lobby against binding requirements [4][5][60]
  • 2026-05-28: Illinois Governor Pritzker signs SB 315 requiring mandatory third-party auditing, 72-hour critical incident reporting, and whistleblower protections [44][45][19]
  • 2026-06-01: Florida becomes first state to sue OpenAI and Sam Altman over ChatGPT's alleged role in violent crimes; AG Uthmeier opens a separate criminal investigation [46][47][48]
  • 2026-06-03: Trump signs revised AI executive order establishing a voluntary 30-day pre-release review for frontier models with NSA oversight and classified compliance thresholds [1][2][30][3]
  • 2026-06-03: OpenAI publishes Frontier Governance Framework calling for coherent federal architecture over fragmented state regulation [38][26][27][39]
  • 2026-06-04: Build American AI — backed by OpenAI, Andreessen/a16z, and Palantir — acknowledged running false-flag social media accounts; WIRED reports the PAC paid TikTok influencers to fear-monger about Chinese AI [28][40][61][62]
  • 2026-06-10: Anthropic CEO Dario Amodei publishes blog calling for mandatory pre-release testing and government-backed independent auditing [12][13]
  • 2026-06-12: New York AG Letitia James serves OpenAI with a subpoena on behalf of a 42-state coalition, targeting advertising practices, data handling, treatment of minors, and chatbot sycophancy [22]
  • 2026-06-14: David Sacks publicly accuses Anthropic of 'regulatory capture strategy based on fear-mongering'; Amodei responds, defending Anthropic's regulatory stance as aligned with Trump priorities [33][14][34][35][36]
  • 2026-06-14: US government forces Anthropic to disable its Fable-class AI models for all foreign nationals [6][37][7]
  • 2026-06-15: Florida expands its lawsuit against OpenAI to 83 pages naming CEO Sam Altman personally, separate from the ongoing criminal probe [23][49]
  • 2026-06-18: DOJ moves to protect xAI's Mississippi AI data center, treating high-speed AI compute as national security infrastructure [8]
  • 2026-06-19: VP Vance publicly accuses AI CEOs of using fear as a marketing strategy, echoing Sacks' critique of AI safety advocacy [15]
  • 2026-06-21: Bloomberg reports Anthropic's AI safety principles have made it a White House target; Eric Schmidt states the US controls ~75% of global AI compute vs China's 15%, attributing the gap to coordinated export controls [16][10]
  • 2026-06-24: Trump administration pressures Meta to submit AI models to government review before public release; Meta is the only major US AI lab outside the voluntary review system [31][52][32][54]
  • 2026-06-25: US government asks OpenAI to release GPT-5.6 as a controlled, staggered preview with customer-by-customer access approval, citing cybersecurity concerns about automated cyberattack capabilities [11]
  • 2026-06-25: White House says 'We hope to sign the agreement soon' about Meta joining the federal AI model review system [9]

Perspectives

Trump administration (federal)

Signed a voluntary AI EO with NSA oversight disclaiming binding requirements; simultaneously forced Anthropic to disable Fable-class models for foreign nationals, moved DOJ to protect xAI infrastructure as national security property, pressed Meta to join the review system, and asked OpenAI to limit GPT-5.6 deployment to government-approved customers.

Evolution: The administration has moved from pre-release review to post-deployment access control: the GPT-5.6 customer-by-customer approval request extends oversight beyond what the June 3 executive order contemplated.

[24][4][1][2][30][3][8][6][16][31][32][11][9]

Tech accelerationist faction (Sacks, Musk, Zuckerberg, Vance)

Opposes AI oversight mechanisms; Sacks accuses Anthropic of regulatory capture through fear-mongering; VP Vance echoed the framing publicly; Sacks is reportedly now seeking to narrow his critique to specific Mythos-model security concerns.

Evolution: Sacks is reportedly attempting to reframe his opposition around specific model security concerns rather than general AI safety advocacy, suggesting a narrowing of argument if not of substance.

[4][5][33][14][15][17]

Anthropic / Dario Amodei

Calls for mandatory pre-release testing and government-backed independent auditing; disputes the 'fear-mongering' characterization while claiming alignment with Trump administration priorities.

Evolution: Confirmed subject of a government export control action requiring it to disable Fable-class models for foreign nationals; Bloomberg characterizes its AI principles as making it a White House target — a difficult position as Amodei simultaneously defends his regulatory advocacy as pro-administration.

[13][12][34][35][36][6][37][16][7]

OpenAI

Publicly advocates for federal governance and shared evaluation standards; agreed to the voluntary pre-release review system; now subject to a government request to limit GPT-5.6 deployment to government-approved customers due to cybersecurity concerns.

Evolution: The GPT-5.6 customer-by-customer approval request is a qualitatively different form of government involvement than pre-release review — OpenAI's stated support for federal governance is being tested against an access management request with no clear statutory basis.

[38][26][27][39][28][40][29][22][41][11]

State legislators (Illinois + California)

Advancing binding AI safety and employment law independent of federal coordination; Illinois SB 315 enacted; California SB 947 and SB 951 advancing.

Evolution: Consistent; operating without federal coordination and not slowed by the Trump preemption EO.

[21][42][20][43][44][45][19]

State AGs (42-state coalition + Florida litigation)

Pursues AI accountability through civil and criminal enforcement; the 42-state coalition subpoenaed OpenAI over consumer practices; Florida expanded its suit to 83 pages naming Altman personally.

Evolution: Expanded from Florida's single-state lawsuit to a coordinated 42-state investigation. Legal theorists including Bruce Schneier have publicly argued for deployer liability frameworks consistent with the AGs' enforcement theory.

[46][47][48][22][23][49][50][51][18]

Meta

Has not joined the voluntary pre-release review system; the administration is actively pressing it to participate, with the White House saying 'We hope to sign the agreement soon.'

Evolution: The White House's public confidence framing suggests the administration expects compliance rather than confrontation — Meta's actual response remains publicly unconfirmed.

[31][52][53][32][54][55][56][57][9][58]

Eric Schmidt

Publicly confirmed that US export controls have produced a 75%-to-15% US-China compute gap, providing the clearest public articulation of the national security doctrine behind administration AI policy.

Evolution: Consistent; provides the most direct public rationale for the enforcement logic the administration has acted on but not stated so directly itself.

[10]

Tensions

  • Mandatory oversight (Amodei) vs. accelerationist opposition (Sacks, Vance): Amodei calls for mandatory pre-release testing and independent auditing [12]; Sacks and VP Vance argue AI safety advocates manufacture fear for commercial reasons [14][15] — yet the administration simultaneously disabled Anthropic's Fable models for foreign nationals and Bloomberg reports Anthropic's principles make it a White House target [6][16]. [12][14][6][15][16]
  • Federal preemption preference vs. state enforcement reality: Trump's preemption EO lacks constitutional authority absent congressional action [24][25]; binding Illinois law [19] and a 42-state coalition subpoena [22] advance without federal coordination. [24][25][19][22]
  • National security enforcement vs. resistance to domestic consumer safety requirements: the government disabled Anthropic's Fable models for foreign nationals [6] and protected xAI compute as national security infrastructure [8] while maintaining binding domestic safety requirements are unnecessary — Schmidt's compute dominance framing explains the external logic but not the domestic-safety gap [10]. [8][6][1][3][10]
  • 'Voluntary' review framework vs. coercive application: the administration labeled the review system voluntary [1][3] but is pressing Meta to join [9] and requesting customer-by-customer government approval for GPT-5.6 [11] — neither action is contemplated in the voluntary framework's stated terms. [1][3][11][9]
  • OpenAI's governance advocacy vs. its PAC's political operations and legal exposure: OpenAI calls for democratic governance and shared evaluation standards [26][41] while Build American AI admitted running sockpuppet accounts [29] and the 42-state subpoena targets the consumer practices OpenAI presents as responsible [22]. [26][28][29][22][41]
  • Government access control vs. absence of statutory authority: the administration's request that OpenAI limit GPT-5.6 to government-approved customers [11] extends executive oversight into post-deployment access management without any clear statutory basis, using cybersecurity concerns to justify a form of control the June 3 EO did not authorize [1][3]. [11][1][3]

Sources

  1. [1] Trump Signs Executive Order For AI Testing Prior To Frontier Model Releases — Zvi's AI Roundups (2026-06-03)
  2. [2] President Trump signs executive order to review advanced AI models. — Rohan Paul Twitter (2026-06-03)
  3. [3] Executive Order Creates Voluntary Regulatory Regime of Frontier AI ... — reactive:us-ai-policy-regulation
  4. [4] Trump abruptly cancels EO signing event after top AI firm CEOs declined to go — Ars Technica AI (2026-05-22)
  5. [5] AI #170: Lack of Executive Order — Zvi's AI Roundups (2026-05-28)
  6. [6] The US government just forced Anthropic to disable its most advanced AI models for all foreign nationals — and the compa... — reactive:us-ai-policy-regulation (2026-06-14)
  7. [7] David Sacks on the US government's shutdown of Anthropic's Fable ... — reactive:us-ai-policy-regulation
  8. [8] Large AI campuses are indeed strategic infrastructure of a nation. — Rohan Paul Twitter (2026-06-18)
  9. [9] 'We hope to sign the agreement soon': White House calls on Meta to submit AI models for review, citing abilities and vulnerabilities evaluation | TechRadar — reactive:us-ai-policy-regulation
  10. [10] Eric Schmidt just said in public what US AI policy has been built around in private for three years (Save this). — Milk Road AI Twitter (2026-06-21)
  11. [11] The Information: The US government is asking OpenAI to slow GPT-5.6 into a controlled preview instead of releasing it br… — Rohan Paul Twitter (2026-06-25)
  12. [12] Dario Amodei just published a super long blog, calling for an urgent policy overhaul because he thinks frontier AI is mo… — Rohan Paul Twitter (2026-06-10)
  13. [13] Dario Amodei argues in his new blog today, that powerful AI will soon become the dominant source of military and economi… — Rohan Paul Twitter (2026-06-11)
  14. [14] Anthropic is running a sophisticated regulatory capture strategy based on fear-mongering. It is principally responsible for the state regulatory frenzy that is damaging the startup ecosystem. — reactive:us-ai-policy-regulation
  15. [15] US VP JD Vance on AI Company's marketing strategy, some AI CEOs sell fear because fear makes their product look powerful… — Rohan Paul Twitter (2026-06-19)
  16. [16] Bloomberg, “Anthropic's AI Principles Make It a White House Target” — reactive:us-ai-policy-regulation
  17. [17] Sacks urges focus on Mythos-driven security needs, seeks to ‘correct’ the narrative around his Anthropic critique | InsideAIPolicy.com — reactive:us-ai-policy-regulation
  18. [18] AI and Liability — Simon Willison (2026-06-25)
  19. [19] Illinois Signs Frontier AI Safety Bill into Law | Chris Lehane posted ... — reactive:us-ai-policy-regulation
  20. [20] CA Senate Approves No Robo Bosses Act of 2026 to Ensure Human Oversight of AI in the Workplace | Senator Jerry McNerney — reactive:us-ai-policy-regulation
  21. [21] Bill Text: CA SB951 | 2025-2026 | Regular Session | Amended — reactive:us-ai-policy-regulation
  22. [22] 😺 42 states just subpoenaed OpenAI — The Neuron (2026-06-15)
  23. [23] Florida AG Sues OpenAI, Alleges ChatGPT Aided Mass Shooters — reactive:us-ai-policy-regulation
  24. [24] President Trump Signs Executive Order Preempting State AI Laws ... — reactive:us-ai-policy-regulation
  25. [25] President Trump Signs Executive Order Challenging State AI Laws | Paul Hastings LLP — reactive:us-ai-policy-regulation
  26. [26] OpenAI public policy agenda — OpenAI Blog (2026-06-03)
  27. [27] A blueprint for democratic governance of frontier AI — OpenAI Blog (2026-06-03)
  28. [28] AI #171: False Flag — Zvi's AI Roundups (2026-06-04)
  29. [29] Build American AI admits hiring an outside vendor to run anonymous sockpuppet accounts targeting its opponents · Digg — reactive:us-ai-policy-regulation
  30. [30] Trump AI Executive Order Gives NSA Classified Role Over Frontier Models — reactive:us-ai-policy-regulation
  31. [31] NYT: The Trump administration is pressuring Meta to submit its AI models to government review before public release. — Rohan Paul Twitter (2026-06-24)
  32. [32] Meta Faces Growing Pressure From Trump Administration To Submit Advanced AI Models For Federal Testing: R - Benzinga — reactive:us-ai-policy-regulation
  33. [33] David Sacks is done being polite about Anthropic (Save this). — Milk Road AI Twitter (2026-06-14)
  34. [34] Anthropic CEO disputes David Sacks' claims that company is 'woke' — reactive:us-ai-policy-regulation
  35. [35] Anthropic CEO defends support for AI regulations, alignment with ... — reactive:us-ai-policy-regulation
  36. [36] Anthropic CEO claps back after Trump officials accuse firm of AI fear-mongering   | TechCrunch — reactive:us-ai-policy-regulation
  37. [37] Legal Considerations Related to the Anthropic “Export Controls ... — reactive:us-ai-policy-regulation
  38. [38] OpenAI’s Frontier Governance Framework — OpenAI Blog (2026-05-28)
  39. [39] 🟡 Let’s get physical — Semafor Technology (2026-06-03)
  40. [40] The OpenAI-Andreessen-Palantir SuperPAC admits that it was “part ... — reactive:us-ai-policy-regulation
  41. [41] Helping build shared standards for advanced AI — OpenAI Blog (2026-06-23)
  42. [42] SB 951: Employment: technological displacement: notice. — reactive:us-ai-policy-regulation
  43. [43] California Set to Restrict AI Use in the Workplace With “No Robo ... — reactive:us-ai-policy-regulation
  44. [44] Trump loses more control over AI regulation as Illinois passes landmark law — Ars Technica AI (2026-05-28)
  45. [45] Illinois Legislature passes historic AI bill that would require third-party safety audits — reactive:us-ai-policy-regulation
  46. [46] ChatGPT creators knew product would cause harm, Florida argues in lawsuit • Florida Phoenix — reactive:us-ai-policy-regulation
  47. [47] Attorney General James Uthmeier Launches Criminal Investigation into OpenAI, ChatGPT | My Florida Legal — reactive:us-ai-policy-regulation
  48. [48] Florida AG brings 'first-in-the-nation state-led lawsuit' against OpenAI and its CEO — reactive:us-ai-policy-regulation
  49. [49] Florida Attorney General Sues OpenAI Over ChatGPT's Alleged ... — reactive:us-ai-policy-regulation
  50. [50] Florida attorney general sues OpenAI | Courthouse News Service — reactive:us-ai-policy-regulation
  51. [51] Florida sues OpenAI over ChatGPT risks Attorney General ... — reactive:us-ai-policy-regulation
  52. [52] Meta Faces White House Push to Join AI Pre-Release ... — reactive:us-ai-policy-regulation
  53. [53] US Government Pressures Meta to Allow Pre-Release Reviews of Its ... — reactive:us-ai-policy-regulation
  54. [54] Trump administration pushes Meta for AI model security reviews- NYT — reactive:us-ai-policy-regulation
  55. [55] The Trump administration is leaning on Meta to agree to voluntary government review of its AI models — and as of today, ... — reactive:us-ai-policy-regulation (2026-06-24)
  56. [56] Meta Under Pressure to Join Federal AI Model Review Program - Techstrong.ai — reactive:us-ai-policy-regulation
  57. [57] Trump Administration Presses Meta To Submit AI Models For Federal Review — reactive:us-ai-policy-regulation
  58. [58] US Government Reportedly Urging Meta To Share Its AI Models — reactive:us-ai-policy-regulation
  59. [59] President Trump Signs Executive Order to Block State AI Laws — reactive:us-ai-policy-regulation
  60. [60] WATCH: Trump explains why he postponed signing AI executive order — reactive:us-ai-policy-regulation
  61. [61] A Pro-AI Super PAC’s Secret Meme Sockpuppets — reactive:us-ai-policy-regulation
  62. [62] A Dark-Money Campaign Is Paying Influencers to Frame Chinese AI as a Threat | WIRED — reactive:us-ai-policy-regulation