Government Proposes Tiered Access Model for AI Development

A newly proposed framework from government researchers advocates for a tiered system of access to artificial intelligence, linking the level of access directly to the capability, safety profile, and associated risks of the AI in question. The Multi-Tiered AI Accessibility Model (MAIAM), detailed in a recently released paper, moves beyond a binary approach of open release versus restricted API access, instead proposing granular controls based on intended use and potential harm.

The MAIAM model establishes three tiers. Tier 1, designated “AI for Learning,” would be broadly accessible but capped in its capabilities. Tier 2, “AI for Exploration,” would serve the research and development community, but under a system of registration, identity verification, and usage controls that intensify alongside increasing capability. The most restricted tier, Tier 3, “Full-Scale Frontier AI,” would be available only through a highly structured process emphasizing recording and accountability. The core principle underlying this structure is that access to AI systems should be proportionate to their capabilities, safety characteristics, and potential risks.

The framework explicitly addresses the challenges of open-weight models, arguing that publishing frontier-class AI weights relinquishes crucial safety controls. Researchers cite evidence indicating that safety fine-tuning can be readily stripped from openly available models for a minimal cost, rendering evaluations of the base model insufficient in assessing the safety of its derivatives. The proposal outlines criteria for assigning systems to each tier, based on factors including capability, autonomy, domain-consequence, and usage/safety considerations.

The MAIAM model is intended to be used in conjunction with a “SAFE AI Operating Point,” creating a two-part framework. This combination establishes both a “Capability Boundary” – defining the upper limits of safe AI performance – and an “Accessibility Boundary” – determining who may access capabilities below that limit. Critically, the document emphasizes that simply limiting capabilities is not enough to ensure AI safety; a tiered access regime administered by regulating bodies is also essential.

The implications of MAIAM extend to the entire AI development lifecycle, suggesting a future where access to powerful AI tools will be carefully managed and monitored. The model’s focus on proportionate access and accountability represents a significant step toward proactive AI regulation, acknowledging that responsible development requires not only technical safeguards but also robust governance mechanisms.