A map of frontier AI governance, organized around one question: how much power does each actor or mechanism really have? Five tiers, from a genuine off-switch down to hopeful communiqués. Click anything for details and connections.
Compiled during BlueDot Impact's Frontier AI Governance course; facts verified against primary sources July 2026, Tier 2 revisions re-verified August 2026. Dashed borders mean proposed, not enacted. The Tier 2 chokepoint framing draws on Sastry, Heim, Belfield et al., Computing Power and the Governance of Artificial Intelligence (2024).
Actors arranged by power tier, most powerful on the left. Threads show who funds, supplies, regulates, evaluates, or pressures whom. Hover a thread for the relationship; click a node for its full record. Dashed threads are proposed mechanisms.
The core ranking. Tier labels describe the strongest lever each actor holds, not how reliably it gets used: most of Tier 1 is self-enforced.
Every entry bucketed by governance layer (columns) and actor type (rows). Hatched cells are empty, and some gaps are load-bearing: no evaluator on this map has binding authority.
Claimed power crossed with enforcement. The uncomfortable read: nearly everything that can stop a deployment is voluntary and self-certified, and nearly everything binding sits well below stop-power.
The governance gap. The one binding, currently exercised pre-deployment gate on this map belongs to the Cyberspace Administration of China, and it screens for political control rather than catastrophic risk. In the US, stop-power is either voluntary (lab safety frameworks) or latent (Congress holds the constitutional authority but has never passed a binding rule on frontier development or deployment). And the far column is the escape hatch: once weights are published, nothing in any other column applies. Closing the highlighted cell, before the escape hatch swallows it, is the central project of frontier AI governance.
When each mechanism bites. "Current" means operating today; the further right, the more the mechanism is a bet on future institution-building.
The full dataset. Click a column header to sort, a row for details.