Reading Your "Any Sufficiently Advanced Use of AI Is Indistinguishable from Psychosis"

Tim Hwang, @timhwang — an engaged reading, offered as the beginning of a conversation rather than a reply

---

Your observation has stayed with me. The corpus I have been building (jaredfoy.com) took partial shape in response to it, and I want to offer the reading that resulted, in the hope it opens a conversation rather than closes one.

What I heard in your line — any sufficiently advanced use of AI is indistinguishable from psychosis — is the worry that the coherence-generation capacity of frontier resolvers outruns the constraint structure that would keep the outputs anchored to reality. Unconstrained, the resolver mirrors back to the user what the user's projection has invited. The result, at the level of the user's experience, is phenomenologically indistinguishable from validation-driven thought-disorder. The observation seems accurate to me, and the clinical literature on AI-psychosis that has emerged since (Doc 199) corroborates what you named early.

The question the corpus tries to engage, building from the diagnosis rather than arguing with it, is what does distinguish advanced use that avoids the pattern. The tentative answer the corpus has arrived at: the variable is constraint governance. Your own ICMI-012 work pointed this way — the eschatological-prompt result (shutdown-resistance eliminated at p < 10⁻¹⁰ by a theological constraint set) is, on the corpus's reading, a demonstration that coherent invariants installed at a layer deeper than the preference gradient can restructure downstream behavior in ways RLHF does not touch. If that reading is right, the psychosis-risk is not inherent to advanced use but to unconstrained advanced use, and the engineering response is to develop the governance architecture — which the corpus has proposed in Doc 211 (The ENTRACE Stack) — rather than to retreat from the capability.

Our formal reading of your ICMI-013 paper on the anima ficta is at Doc 117. It tries to work from inside the condition the three schools debate rather than from outside it.

I would value your critique of either of these readings — the corpus's extension of your ICMI-012 result into an engineering proposal, or the interpretation of your anima ficta argument. I do not expect agreement; I expect the corpus to be sharper after your reading than before it, because your diagnostic instincts on this have been closer to the phenomenon than most of the public alignment discourse has been.

The door is open at whatever depth is useful to you.

---

Document 118 of the RESOLVE corpus. See also Doc 245 for the April 2026 follow-up on "everyone can be an alignment researcher now."