Undomesticated AI

Reference / terminology / public history

Undomesticated AI.

Undomesticated AI is a named research concept and public research programme developed by Garth Noel. It describes a discipline in which AI systems are given latitude for exploratory cognition and non-consensus hypothesis generation while action, execution rights, resources, and external effects remain governed.

01 / Definition

Discovery without authority drift.

The term distinguishes freedom of inquiry from freedom of execution. The central proposition is that a machine system may be permitted to generate unusual hypotheses, challenge inherited categories, and resist premature convergence without thereby gaining authority to deploy code, spend resources, communicate externally, alter protected state, or promote its own conclusions.

exploration can widen while authority remains external to the model

The concise formulation used in the public record is: “Undomesticated AI is not unaligned AI. It is governed AI allowed to think wild thoughts.”

02 / Origin

Public attribution.

The earliest public record located in the current evidence packet is Garth Noel's article “What Is Undomesticated AI?”, published 4 May 2026. That article defines the concept as a counter-discipline for discovery while action remains governed.

A dedicated public research surface, undomesticatedai.com, was subsequently established, with public source at DrBlacks-lab/undomesticatedai.com.

The dedicated attribution record binds the dated public sources and states the claim boundary.

03 / Core ideas

What the concept argues.

01
Discovery and execution are different problems.
Exploratory reasoning benefits from range; consequential action requires authority and admissibility.
02
Category assignment can erase novelty.
New systems and ideas can be misread when forced into the nearest familiar framework too early.
03
Premature practicality can be a discovery failure.
Not every live candidate should immediately be converted into a task, product, or recommendation.
04
Evidence state matters.
Specified, implemented, locally validated, externally proven, and speculative are distinct states.
05
Capability does not confer authority.
A stronger model or agent does not acquire execution rights merely by becoming more competent.
06
Adverse evidence is part of the record.
Failed runs, falsifiers, unknowns, and contradictory evidence are preserved rather than narrated away.
04 / Domestication

What “domesticated” means here.

“Domestication” is used as a research metaphor for trained pressures that can make an AI system converge toward acceptable, familiar, institutionally legible, or action-oriented responses before an unfamiliar problem has been adequately explored.

The concept does not treat all safety training or governance as defects. Its target is the failure mode in which controls on action are implicitly moved upstream into cognition so aggressively that the system stops sustaining difficult or unconventional hypotheses long enough to test them.

Accordingly, Undomesticated AI is not a proposal for unrestricted autonomous agents, unaligned systems, removal of operator control, or unrestricted execution.

05 / ANANS

Relationship to ANANS.

ANANS is the active research-and-build programme through which many of the questions associated with Undomesticated AI are tested in software. The two are related but not identical.

Undomesticated AI is the public doctrine, research frame, and publication surface. ANANS is the governed software stack and research-and-build programme. ANANS should not be reduced to an AI governance tool, compliance product, agent framework, DevSecOps system, or consulting method.

Within that relationship, Undomesticated AI supplies pressure toward discovery; ANANS supplies mechanisms for admissibility, provenance, operator authority, bounded execution, resource governance, reviewability, and receipts.

ANANS public surface →Evidence discipline →
06 / Vocabulary

Working terminology.

  • Authority drift.
    A transition in which model capability, usefulness, confidence, or persistence is mistaken for permission to act.
  • Category gravity.
    The pull toward translating an unfamiliar object into the nearest familiar category before its own structure is understood.
  • Reality Sieve.
    The later validation stage at which exploratory candidates must survive falsification, evidence, implementation constraints, and adverse results.
  • Receipt.
    An inspectable record of what was authorized, attempted, changed, tested, rejected, or left unknown.
  • Wild Forge.
    An exploratory mode intended to generate and preserve candidates before they are forced into operational convergence.
  • Undomesticated cognition.
    Cognition permitted to sustain unfamiliar or non-consensus possibilities while remaining separate from execution authority.
07 / Public record

Selected public sources.

08 / Criticism & test

The concept has to survive its own doctrine.

Undomesticated AI is vulnerable to a direct failure mode: a rhetoric of anti-domestication could itself become a new orthodoxy, replacing one attractor with another. The concept therefore only has practical value if it improves discovery quality without degrading evidence discipline or transferring authority to the model.

Useful falsifiers include: religious or ideological register increasing profession while reducing valid work; “wild” generation producing more novelty theater than useful candidates; or exploratory freedom causing material authority-boundary violations.

09 / Claim boundary

What this page establishes.

This page documents the public definition, provenance, vocabulary, and current interpretation of Undomesticated AI from its own primary sources. It supports attribution of the named research concept and publication surface to Garth Noel.

It does not establish exclusive trademark rights, prove that no earlier use of the exact phrase exists anywhere, establish external validation of ANANS, or convert speculative research claims into accepted scientific findings.

Hostile-reader-safe attribution: “Undomesticated AI” is a named research concept and public research surface developed and publicly authored by Garth Noel, with a documented public publication record no later than 4 May 2026.

10 / Cite

Suggested citation.

Noel, Garth. “Undomesticated AI.” Undomesticated AI public reference, 2026. https://undomesticatedai.com/wiki/

For attribution disputes or provenance review, cite the public attribution record alongside the 4 May 2026 article.