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).
Advanced AI systems can produce behavior that appears familiar while relying on internal representations, abstractions, and optimization processes that humans cannot fully reconstruct. Calling this an “alien mind” is a metaphor for an epistemic problem, not a claim about consciousness. The governance problem arises when behavioral competence grows faster than our ability to understand, evaluate, interrupt, and verify the processes that produce consequential actions. This article introduces the AI Capability–Supervision Gap as a systems concept. Let C denote the operational capability envelope of a system, A the authority and reach made available to it, and S the envelope within which humans and technical controls can reliably supervise behavior under real constraints. The actionable gap is the region Gᴀ = (C ∩ A) \ S: consequential behavior that the system can perform and is able to reach, but that supervision cannot reliably interpret, detect, stop, or verify. This is a conceptual relation, not a calibrated universal metric. The article argues that as internal processes become less inferable from observable behavior, safety must depend increasingly on external limits over what the system may do. Alignment, interpretability, and chain-of-thought monitoring remain valuable, but cannot alone authorize action. Through the public SGAEIA research framing, the article connects model uncertainty to bounded and revocable authority, governed execution, independent evidence, and recovery. The less we can infer internal processes from observable behavior, the less safety can depend exclusively on interpreting the model — and the more it must depend on external limits over its capacity to act.
No comments yet — start the discussion below.