To Control AI, Use Humanity's Oldest Tool: Collective Intelligence — A Response to the UN Independent Scientific Panel on AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

On 1 July 2026, the UN's Independent International Scientific Panel on AI, co-chaired by Yoshua Bengio and Maria Ressa, stated in its preliminary report that "reliable methods for retaining control over highly autonomous AI systems are lacking." We argue that this claim is accurate as a description of currently deployed systems but incomplete as an assessment of what is technically buildable. We propose and analyze a control architecture grounded in collective intelligence: a committee of AI models drawn from independent, ideally rival, laboratories that monitors a target system's consequential actions, governed by a "super agent" that selects — rather than dictates — the committee's decision procedure from a constitutionally pre-authorized menu. We ground the proposal in recent empirical work on AI-monitors-AI schemes, in Arrow's impossibility theorem and mechanism design theory, and in a thirty-eight-year line of the authors' own multi-agent-systems research (1988–present). We identify the architecture's central vulnerability — the procedure-selecting agent itself — and describe how constitutional constraints, enforced outside the model rather than merely trained into it, can bound its discretion. We close by stating plainly what is not yet demonstrated: independence across monitors erodes as models improve, existing evidence comes from narrow domains, and inter-laboratory cooperation is a governance problem as much as a technical one. Governments facing a self-imposed end-of-2026 deadline for AI "red lines" need not wait for a single technical breakthrough; the components exist, and what is missing is assembly, stress-testing, and institutional will.