The three systems
- System 1
- Fast, automatic, pattern-matched judgment. Always available, metabolically cheap, and right often enough that most people stop checking it. It is not the failure point in this framework.
- System 2
- Deliberate, effortful reasoning. The capacity to lay out options, argue against your own position, and change your mind. It is metabolically expensive, and it is the first capacity lost under acute physiological load.
- System 3
- System 3 is artificial cognition that operates outside the human brain, as an external layer of decision-making. It extends Daniel Kahneman's dual-process theory of human thought: System 1, fast intuition, and System 2, slow deliberation.
The distinguishing claim
The term System 3 is in circulation with more than one meaning, and the differences are not cosmetic.
One definition, put forward by Wethos AI, describes System 3 as expert judgment moving at the speed of intuition: decades of accumulated experience compressed into a single moment of clarity, and the capacity they argue AI cannot replicate. That names something real. It describes what an expert brings into the room before anything else happens.
Version 2.0 means something else. System 3 here is artificial cognition operating outside the human brain, as an external layer of decision-making. It extends Kahneman's distinction between fast intuition and slow deliberation by naming the machine's contribution, not an expert's accumulated judgement.
The practical difference is sharp. On the Wethos definition, the response to AI is to protect and trust hard-won expertise, because the machine cannot match it. In v2.0, the question is how Systems 1 and 2 engage with external cognition. Expertise can still be exploited under load: a fluent, well-sourced output pattern-matches to something an experienced person recognises as correct, and recognition is the fast System 1 process that pressure leaves running. Experience is not the protection here. Without a trained pause that keeps System 2 engaged, it is part of the exposure.
The investment follows the definition. Wethos's emphasis on accumulated expertise points towards hiring and keeping experienced people. The v2.0 definition also requires decision architecture for the machine and state regulation and deliberate practice for the person. Those human practices keep Systems 1 and 2 engaged when System 3 is doing the thinking; they are not System 3 itself.
This also lines up with research outside Behaviorist's own work. Wharton researchers Steven D. Shaw and Gideon Nave propose what they call Tri-System Theory: Kahneman's System 1 and System 2, plus a third system they define as artificial cognition operating outside the brain. Their paper, Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender, measured what they call cognitive surrender directly: across three preregistered experiments, human accuracy rose 25 points when an AI's answer was right and fell 15 points when it was wrong. Their third system names the machine, the same thing v2.0 names. It does not itself specify what keeps a person's own two systems engaged when System 3 supplies an answer; that is the question this framework and Behaviorist's checkpoints take up.
The gate
Which human systems a person can actually reach is governed by physiological state, not by intelligence or seniority. Under acute load, deliberate reasoning narrows first. What remains is pattern and instinct, the mode least equipped to catch its own errors.
An AI output arriving during high load is therefore processed by the fastest available system, and a fluent, well-sourced output is an unusually good fast answer. The tool arrives exactly when scrutiny is least available. This is the mechanism behind what current AI/psychology commentary calls cognitive surrender, see Behaviorist's field note for the concrete version engineered into agentic systems.
Markers
Two observable markers show whether Systems 1 and 2 stay engaged when System 3 supplies an answer, and both sit outside the moment of asking.
- Before: was the question asked from a regulated state or a keyed-up one. The prompt written at minute one of a crisis is not the same instrument as the same prompt written after a breath count.
- After: did the output get one more option pulled out of it before anyone acted, or did the first confident answer become the plan. System 2, when it is actually online, asks a model for a spread. System 1 asks a model for permission.
Evidence status
- Established: acute stress narrows the scope of deliberate reasoning. This is well replicated and not controversial.
- Established: paced breathing at roughly six breaths per minute measurably shifts autonomic state, and has been tested in controlled work.
- Contested: the anatomical account given by polyvagal theory is disputed. This framework uses the breathing protocols and the measurements, and does not rest on that theory's name or its anatomy.
- Not established: the specific claim that a person under load using an AI system converges on confirming a prior they already held, rather than widening the decision. That is this author's own claim. It is currently untested, and it is stated here as a hypothesis rather than a finding. It is listed as an open question on the Behaviorist Lab research agenda, and the study that would test it has not yet been run.
Scope
The framework describes individual human decision making, and extends to group decisions only with the additional mechanism set out in the Research essays. It is a framework for decision quality, not a clinical model, not a diagnostic instrument, and not a wellbeing intervention.
Version history
v1.0, 17 September 2026. First published reference definition.
v2.0, 14 October 2026. System 3 defined as artificial cognition outside the human brain.
What changed from v1.0
Version 1.0 (17 September 2026) defined System 3 as a trained human capacity: keeping System 2 online when an AI output arrives. Version 2.0 aligns with Shaw and Nave's Tri-System Theory (Wharton, SSRN 6097646), which independently defines a third system as artificial cognition operating outside the human brain. The human skill v1.0 described still matters; it is now framed as how Systems 1 and 2 stay engaged when System 3 is doing the thinking.
In practice, train people to check their state and evaluate the machine's output, while setting checkpoints for what the machine can decide. A fluent answer from System 3 does not mean Systems 1 and 2 have done their part.