Frontier AI: Who wants which rules for the most powerful models?

European Commission · Mandatory safety evaluations

Article 55 requires providers of general-purpose AI models with systemic risk to perform model evaluations using state-of-the-art protocols, including documented adversarial testing. These are binding duties, not merely a voluntary code.

perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art

European Commission · Mandatory independent audits

Article 92 lets the AI Office conduct evaluations and allows the Commission to appoint independent experts; providers must supply requested access. The mechanism is mandatory when invoked, but the AI Act does not impose a general recurring independent third-party audit on every systemic-risk model.

The Commission may decide to appoint independent experts to carry out evaluations on its behalf

European Commission · Voluntary safeguards as primary regime

The Commission describes the GPAI Code of Practice as a voluntary tool for demonstrating compliance, but the underlying AI Act duties for systemic-risk models are mandatory and fully enforceable from August 2026. Voluntary safeguards therefore are not the primary regime.

U.S. White House · Mandatory safety evaluations

Executive Order 14409 orders a classified benchmarking process but directs agencies to design a voluntary framework through which developers may provide covered frontier models to the government before release. It does not create a mandatory developer safety-evaluation duty.

U.S. White House · Government release approval

Section 3(c) says the frontier-model provisions cannot be read to authorize mandatory government licensing, preclearance or permitting for developing or releasing new AI models, including frontier models.

Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement

U.S. White House · Voluntary safeguards as primary regime

Executive Order 14409 directs federal agencies to build a voluntary framework with developers for frontier-model designation, secure pre-release access and trusted-partner collaboration rather than a binding release-approval regime.

UK Government · Mandatory safety evaluations

The government reports that AISI has tested 30 frontier models and funds frontier-model testing as a core safety function. AISI separately describes its current work with leading developers as voluntary, while the government has also announced an intention to legislate on risks from the most powerful models.

UK Government · Mandatory independent audits

AISI is institutionally separate from developers and performs frontier-model evaluations, yet the reviewed source describes these collaborations as voluntary. The government’s announced legislation has not, in the reviewed material, specified a recurring mandatory third-party audit requirement.

ongoing voluntary collaborations with Anthropic and OpenAI

UK Government · Voluntary safeguards as primary regime

The present AISI collaboration model is voluntary, but the government has said it will legislate to address critical risks from the next generation of the most powerful models. That makes the UK a mixed case rather than a purely voluntary regime.

State of Illinois · Mandatory safety evaluations

SB 315 requires large frontier developers to implement and comply with a frontier AI framework covering catastrophic-risk assessment and mitigations, and to publish summaries of catastrophic-risk assessments for covered deployments. Violations can trigger civil penalties.

State of Illinois · Mandatory independent audits

Beginning in 2028, large frontier developers must annually retain an independent third party to audit compliance with the Act’s frontier-AI framework requirements, with independence and access requirements for the auditor.

shall annually retain a third party to perform an independent audit

State of Illinois · Voluntary safeguards as primary regime

SB 315 imposes mandatory frontier-AI frameworks, risk disclosures, incident reporting and third-party audits backed by civil penalties. Voluntary industry practices can supplement the law but do not replace its core duties.

Anthropic · Mandatory safety evaluations

Anthropic’s Advanced AI Framework proposes obligations for the most capable model developers to test models for defined catastrophic risks, publish findings and answer to a government agency with authority to act.

Anthropic · Mandatory independent audits

The framework says covered frontier developers should submit their risk work to qualified independent evaluators and calls for independent evaluation to become a standard part of the governance system rather than a purely internal assessment.

Anthropic · Government release approval

Anthropic proposes legal authority to block or deter deployments that pose significant catastrophic risk and a mechanism for doing so. The reviewed framework does not clearly require every covered model to obtain a government licence or pre-release approval before deployment.

government should have the legal authority to block or deter its deployment

Anthropic · Voluntary safeguards as primary regime

Anthropic explicitly calls for government action, binding developer obligations and enforcement authority. Voluntary company safeguards remain relevant, but the proposed primary regime is legal regulation with independent scrutiny.

OpenAI · Mandatory safety evaluations

OpenAI’s federal blueprint says that once CAISI has sufficient capacity, policymakers should require the most capable frontier models to undergo a CAISI evaluation before public release.

OpenAI · Mandatory independent audits

The blueprint calls for large frontier developers to retain an independent third party annually to audit compliance with frontier-safety requirements and for certified independent technical assessment of covered models.

OpenAI · Government release approval

Its proposed CAISI would evaluate frontier models and recommend mitigations, but developers would remain responsible for deployment decisions; the blueprint says CAISI should not approve or block deployments.

not to approve or block deployments

OpenAI · Voluntary safeguards as primary regime

The blueprint argues for a binding national framework, mandatory evaluations, independent audits and enforceable consequences, treating voluntary company policies as insufficient on their own.

Google · Mandatory independent audits

Google’s 2026 U.S. governance proposal says a federally overseen, industry-backed body could set safety standards and verify independent voluntary audits for the most advanced frontier models. That directly falls short of a mandatory audit duty.

independent, voluntary audits

Google · Voluntary safeguards as primary regime

For frontier AI, Google proposes an independent, federally overseen, industry-backed organization that sets safety standards and verifies voluntary audits, rather than a binding licensing or mandatory-audit regime in the proposal reviewed.

Yoshua Bengio · Mandatory safety evaluations

In his September 2026 UN Security Council speech, Bengio argued that developers must prove the safety of systems before training and deployment, framing the requirement as a collective governance rule rather than a voluntary practice.

Yoshua Bengio · Mandatory independent audits

His proposal explicitly assigns the safety demonstration to independent experts, making outside scrutiny part of the required gate for training and deployment.

Yoshua Bengio · Government release approval

Bengio proposes licensing frontier AI, comparing it with regulated domains such as medicine, aviation and nuclear energy. This is the clearest endorsement in the matrix of a formal approval regime.

Frontier AI should be licensed like other critical technologies

Yoshua Bengio · Voluntary safeguards as primary regime

His proposal uses mandatory language for independent safety demonstration, incident reporting and licensing. Voluntary developer commitments therefore do not constitute his preferred primary governance model.