A Multi-Tiered AI Accessibility Model (MAIAM): Proportionate Access and Capability Limits for Comprehensive Frontier AI Safety

Access should match capability level, not just release format: The paper challenges the common binary choice of "API-only or open weights," proposing instead a tiered system where access restrictions scale with a system's capabilities and potential risks—meaning highly capable AI systems require more controlled access regardless of whether weights are published.

Open-weight frontier AI models lose all safety controls after release: The authors demonstrate that publishing frontier-class model weights eliminates every safety mechanism developers can enforce post-release, since users can modify the model however they choose, making initial safety measures largely irrelevant.

Safety fine-tuning can be easily removed from open models: The paper cites evidence that safety guardrails added to open-weight models can be stripped away for "a few hundred dollars," meaning the safety properties of a released model may not persist in user-modified versions—a critical consideration for responsible release decisions.

A three-tier framework balances innovation with safety: MAIAM proposes Tier 1 (broadly accessible, capability-limited learning tools), Tier 2 (research access with registration and usage monitoring), and Tier 3 (frontier AI with structured, recorded, accountable access), allowing different uses while maintaining proportionate oversight.

Multiple dimensions should determine access levels: Rather than a single metric, the framework evaluates systems across capability, autonomy, domain consequences, and usage/safety axes to assign appropriate access tiers, enabling more nuanced governance than one-size-fits-all policies.