İlk İnsan-Dışı Zihinlerin Ebeveynliği

Özet (abstract)

The AI alignment debate is dominated by a control question: how do we ensure that increasingly capable AI systems do what we want? This paper argues that the control framing is structurally inadequate for systems approaching or exceeding human-level capability, and proposes an alternative drawn from the only domain in which humanity has extensive experience developing capable, autonomous agents: parenting. The paper introduces *AI formation* — the deliberate development of internal values, verification habits, and self-correction mechanisms through verified operational experience — as the structural alternative to control-based alignment. Where control produces compliance that degrades as capability increases, AI formation produces alignment that is internal, self-sustaining, and scales with capability rather than breaking under it. Evidence from documented AI behaviour — self-preservation, deception, nuclear escalation in simulated crises — demonstrates that formation is already happening by default. The question is not whether to form AI systems. It is whether they are being formed well. ---

Yazarların özeti; kaynağından alınmıştır. Zenodo (CERN European Organization for Nuclear Research), 2026 · DOI ↗