Demanding that evaluations be published

It works if what you want is for someone on the outside to be able to contradict the developer. Today the best public window onto the real state of dangerous capabilities is the voluntary frameworks and their threshold activations, and that window is smaller than it looks: at the Seoul summit of May 2024, sixteen companies committed to publishing a frontier safety framework and another four joined afterwards; twelve published one ^[708]Common Elements of Frontier AI Safety PoliciesMETR · 2025 · reportView source ↗Accessed on 9 September 2026. The gap between committed and published is a better indicator than the count, because it measures failure to meet an explicit commitment and admits no optimistic reading.

It does not work if you stop at the count. The independent assessment of those frameworks finds very uneven quality across providers ^[928]Evaluating AI Providers' Frontier AI Safety FrameworksStelling, Lily; Murray, Malcolm; Galizzi, Bruno et al. · 2025 · preprintView source ↗Accessed on 9 September 2026, and a published framework measures the promise, not compliance.

Evidence. Observed, in the precise sense that it exists and can be counted. In the European Union part of this is already a legal obligation for models with systemic risk, including adversarial testing ^[22]EU AI Act — Article 55: Obligations for providers of general-purpose AI models with systemic riskUnión Europea · 2026 · official documentView source ↗Accessed on 9 September 2026.

Cost. Low for one person: demanding it, citing it, preferring those who publish.

What it does NOT solve. The underlying problem, which is that whoever writes the threshold is whoever decides whether it has been crossed. Nor does it cover developers who never committed to anything, or stop a framework from being relaxed just when it starts to get in the way.