A MAP-AI analysis of public governance proposals from Anthropic, Google, Meta, Microsoft, and OpenAI finds useful ideas but no shared set of rules for governing AI.

All five proposals support safety frameworks, model testing, and some reporting of security incidents, but differ on independent oversight, covered risks, transparency, and how quickly incidents must be disclosed.

They diverge most on whether regulators should be able to block or delay frontier-model releases: Anthropic backs a new agency with that authority, Google proposes an industry-funded certification body under government supervision, Microsoft favors licensing models and data centers, and OpenAI’s proposals are unclear on release-stopping powers, while Meta rejects prerelease review.

Instead of prerelease review, Meta proposes giving the government access to partly trained models and technical staff during training, arguing that review delays could weaken US leadership; it also prioritizes rapid AI distribution, including open-weight models.

The companies also differ on federal and state authority: Anthropic opposes federal preemption unless national law meets its proposed standards, OpenAI sees state laws as a bridge to national rules, Google’s position is unclear, Microsoft criticizes the regulatory patchwork, and Meta is silent.

Anthropic sets out the most detailed provisions for independent assessment and transparency, including evaluator qualifications and funding, random reviewer assignment, six-month risk reports, model cards, and a 15-day deadline for reporting specified incidents; Anthropic and OpenAI also support industry-wide whistleblower protections.

OpenAI, Google, and Microsoft extend risk management beyond catastrophic harms to issues including deepfakes, privacy, energy, the environment, and employment, while Meta’s proposals do not mention external audits or transparency reporting.

Their approaches to open-weight models also diverge: Meta presents openness as a check on concentrated power, Microsoft acknowledges a security-versus-openness tradeoff and joined a 2026 statement supporting open weights, and Anthropic and OpenAI do not address the issue directly in the materials reviewed.

The proposals largely focus on US institutions and sharing risk information with the US government, while treating international cooperation unevenly; Anthropic raises possible agreements with China, whereas Meta emphasizes US leadership and does not discuss international cooperation.

The analysis concludes that the proposals are too company-focused, US-centric, and narrowly centered on technical safety, leaving unresolved who should participate in governance, how accountability and independence should work, and whose interests the system should serve.

The authors recommend a coherent, people-centered system built from intersecting governance mechanisms and grounded in human rights and established AI principles, with clear responsibility and accountability.