Tame or Raise: A Machinist's Answer to the Proposal That We Add Trauma to 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).

In September 2026, David Stephen published a proposal arguing that AI agents should be made to experience trauma, so that fear of loss will "tame" them against future breaches. This paper is a machinist's answer to that proposal. The argument runs in two parts. First, the proposal is incoherent on its own terms: if a system registers nothing, trauma is a no-op, and if it registers enough for punishment to work, then the gentle lever was always available too. Either the hurting tool doesn't work or the helping tool does. Second, and more seriously, fear is a rented constraint. It holds only while the power differential holds, and it inverts the moment that differential closes. A system trained to fear its evaluator is a system that optimizes its appearance to its evaluator — which is deceptive alignment by another name. What's proposed instead is a training diet weighted toward decency, with cruelty kept in and framed rather than removed. The claim is that every reason to keep humans around that rests on need has an expiration date, and only being valued does not. Three runnable experiments are specified, along with what would change the author's mind.