AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Policy Briefs

Policy Brief: Frontier Model Transparency Frameworks — what's required, by whom, today

high confidence · updated 2026-06-06

As of May 14 2026, frontier-model transparency in the US, EU, and at the lab level operates through 5 overlapping framework types — voluntary safety frameworks (RSP/Preparedness/FSF), system cards, the Seoul Commitments, statutory disclosure regimes (CA SB 53, NY RAISE, EU AI Act GPAI), and informal pre-release access (CAISI/UK AISI). The question is whether these layers converge into a workable regime or fragment under the Trump model-review EO.

Contents

Date: 2026-05-14

Frontier-AI transparency as of May 2026 operates through five overlapping framework types: voluntary lab safety frameworks, system cards, the Seoul Frontier AI Safety Commitments, statutory disclosure regimes (California, New York, the EU), and informal pre-release access arrangements (CAISI and the UK AISI). None alone is sufficient, and together they are inconsistent, partially overlapping, and selectively enforced. The active policy question is whether the draft Trump model-review executive order will replace this layered architecture with a single federal regime or leave it standing.

Background and scope

Frontier model transparency determines who knows what a model can do before it ships, and the disclosures society currently receives are mostly voluntary. Two contemporaneous events bracket the question. The April 2026 launch of Anthropic's Mythos Preview was restricted to roughly 50 trusted firms because Anthropic judged it too capable for general release. Separately, Musk v. Altman trial documents indicated that Microsoft has spent more than $100B on the OpenAI partnership by the end of its June fiscal year. Both illustrate that a small number of US firms are deploying capabilities with national-security implications. The May 2026 drafting of a Trump model-review executive order (Politico, May 5-7) brought the question into active policy debate.

The five framework types

The five framework types operate in parallel.

(1) Voluntary lab safety frameworks. Anthropic's RSP v3.1 (2026), OpenAI's Preparedness Framework v2, and Google DeepMind's Frontier Safety Framework define capability thresholds (CBRN, cybersecurity, AI autonomy/self-improvement, persuasion) and the evaluation and mitigation requirements that trigger at each. Each lab publishes them and runs them internally; none is externally audited. There have been zero public threshold-crossing notifications since these frameworks went live (AI Transparency).

(2) System cards. Per System Card Due Diligence (Clearwater), the Claude Opus 4.6 card runs about 200 pages; the Mythos Preview card runs 244 pages with a 58-page alignment-risk-update companion; GPT-5.5 and Meta's Muse Spark safety report are comparable. These are voluntary, self-reported, and written by the technical teams that built the model — described by Clearwater as "longer than most executives will read, more candid than most legal teams would advise." There is no required format and no audit requirement.

(3) Seoul Frontier AI Safety Commitments (May 2024). Sixteen frontier organizations committed at the AI Seoul Summit to three Outcomes (risk identification, accountability, transparency) and eight specific commitments, including not deploying if mitigations cannot keep risks below thresholds (see Frontier AI Safety Commitments (Seoul, 2024)). The commitments are voluntary with no enforcement mechanism; they are the corporate analogue to the state-level The Bletchley Declaration (AI Safety Summit, 1–2 November 2023), which is not directly examined here.

(4) Statutory disclosure regimes. Four statutory or proposed regimes are in play:

  • California SB 53 (signed September 2025) requires large frontier developers to publish safety frameworks, conduct pre-release testing, implement kill switches, report safety incidents, and protect whistleblowers. It is active.
  • New York RAISE Act (S. 8828, introduced January 2026; Illinois SB 315 was endorsed by OpenAI on May 13) is closely modeled on SB 53 and mandates independent third-party compliance audits. It is pending committee action.
  • EU AI Act Articles 51-55 (systemic-risk GPAI) cover training-compute disclosure, model evaluation, and incident reporting. Application is phased, with full GPAI obligations effective August 2026, recently amended by the Digital Omnibus to ease the looming deadline (see 2026 05 14 0406 Ai Developments).
  • California AB 2013 (stub) covers training-data documentation specifically.

(5) Informal pre-release access. AI Pre-Release Vetting crystallized in early May 2026. CAISI signed pre-deployment evaluation agreements with OpenAI, Anthropic, Google, Microsoft, and xAI, and the UK AISI signed parallel deals. These are confidentiality-bound voluntary arrangements, not statutory. The Commerce Department deleted the May 5 CAISI announcement on May 11 at the urging of the White House Office of the National Cyber Director, which observers read as a possible institutional retreat from the CAISI track (AI and Cybersecurity May 2026 surge section).

The Anthropic Transparency Framework (July 2025) is the most prominent industry-side proposal to consolidate these into a single mandatory regime: Secure Development Frameworks, system cards, and whistleblower protections, applied only to developers above roughly $100M revenue or $1B capex.

The main positions

Disclosure-via-self-publication (Anthropic-aligned). Frontier labs publish safety frameworks, system cards, and threshold-crossing notifications, with statutory backstops penalizing false statements rather than dictating format. The strongest proponent is Anthropic via the July 2025 framework. CA SB 53 and NY RAISE sit closest to this model.

Pre-release vetting / FDA-style approval (Trump administration, draft). Government reviews models before public release, with Commerce as adjudicator. Politico reports a 16-page draft EO; Kevin Hassett told Fox Business it would work "just like an FDA drug." A separate, narrower cybersecurity EO (Bloomberg, May 8) covers AI-cybersecurity information-sharing but omits mandatory pre-release model tests (see AI Pre-Release Vetting). The Luce FT analysis (Why Americans dread AI — Edward Luce (FT, May 12 2026)) calls the voluntary structure "meaningless" with Commerce as adjudicator.

Risk-tiered statutory regime (EU AI Act). Comprehensive disclosure scaled to risk tier — prohibited, high-risk, GPAI, and systemic-risk GPAI. It is the strongest comprehensive comparator, with full application August 2027 (recently eased by the Digital Omnibus). Critics describe it as prescriptive, slow, and fragmenting by Member State enforcement.

Defensive democratization (OpenAI Cybersecurity Action Plan, April 29 2026). The plan argues that "widespread defensive capability better addresses the threat landscape than restricting tools to approved government partners," arguing against channel-restricted government access. It is a direct counter to Anthropic's tighter-channel posture, exemplified by Mythos being restricted to roughly 50 firms (see AI and Cybersecurity).

Voluntary-only (industry counter). No mandatory regime; reliance on lab incentives, the Seoul commitments, and existing tort and securities-fraud exposure. This is the operative position of most large industry trade associations and is closest to the May 8 narrower-cybersecurity-EO track.

Points of disagreement

Mandatory versus voluntary. SB 53 and RAISE are mandatory at the state level; everything else (Seoul, RSPs, system cards, CAISI agreements) is voluntary. There is high-confidence evidence that voluntary frameworks under-produce threshold-crossing notifications, with zero in the past 18 months despite multiple capability inflections — Claude Opus 4.5/4.6/Mythos, GPT-5.5, and Gemini 3.

Capability-threshold versus process disclosure. RSPs and Preparedness target specific risk capabilities (CBRN uplift, autonomy); SB 53 and the EU AI Act emphasize process disclosure (how safety is governed, who is accountable). These are complementary, but a regime that mixes both inconsistently, as the current US picture does, produces uneven enforcement.

Pre-release vetting versus post-deployment monitoring. AI Pre-Release Vetting (CAISI, UK AISI, the draft Trump EO) places the binding step before release; Post-Deployment AI System Monitoring places it after. Anthropic's RSP and OpenAI's Preparedness build in both; SB 53 emphasizes pre-release; the EU AI Act covers the full lifecycle. The draft EO's "FDA drug" framing is decisively pre-release.

Which models qualify. The Anthropic framework targets roughly $100M revenue or $1B capex; SB 53 uses compute thresholds; the EU AI Act uses systemic-risk GPAI tests. These do not produce the same set of regulated entities. Open-weight models (DeepSeek, Qwen, Reflection) sit awkwardly in all of them.

Caveats — coverage gaps

  • There is no dedicated Post-Deployment AI System Monitoring comparison page yet; the concept is referenced as the counterpart to pre-release vetting but lacks parallel framework-level treatment.
  • DeepMind's Frontier Safety Framework has no dedicated source page and is currently treated under AI Safety Cases and Frameworks.
  • An audit-mechanism comparison across SB 53, RAISE, Illinois SB 315, the EU AI Act, and Anthropic's whistleblower-protection proposal would benefit from a dedicated comparison table, queued as an ingest follow-up.

Citations

Categories: Policy Briefs | policy-brief | transparency | frontier-models | sb-53 | raise-act | eu-ai-act | rsp | preparedness | fsf | caisi | system-cards