MAIAM Proposes Tiered Access and Capability Limits for Frontier AI Safety

Summary

The paper argues that debates over releasing capable AI systems incorrectly treat access as a binary choice between an API and public model weights. It separates two decisions: how capable a system should be allowed to become, and who should obtain a given capability under what conditions. Its proposed Multi-Tiered AI Accessibility Model (MAIAM) assigns access according to purpose, safety, and capability. Tier 1, AI for Learning, is broadly available but deliberately capability-bounded; Tier 2, AI for Exploration, supports research and development with registration, identity, and usage controls that become stricter as capability rises; and Tier 3, Full-Scale Frontier AI, permits only structured, recorded, and accountable access. The paper argues that publishing frontier-class weights removes the developer's post-release safety controls, and cites evidence that safety fine-tuning can be stripped from open weights for a few hundred dollars, making base-model evaluations poor indicators of how derivatives may behave. It proposes assigning tiers using capability, autonomy, domain-consequence, and usage/safety criteria. MAIAM is combined with the SAFE AI Operating Point into a framework with a Capability Boundary that sets limits not to cross and an Accessibility Boundary that governs who may obtain capabilities below those limits. The authors conclude that capability limits alone are insufficient and that regulating bodies are also needed to administer tiered access.