Cognitive Surrender and Tri-System Theory
Steven D. Shaw and Gideon Nave extend the familiar distinction between fast intuition and slow deliberation with a third system: external artificial cognition. Their model asks where the reasoning happened and whether the person retained control over the conclusion.12
Tri-System Theory remains a working hypothesis. Shaw and Nave tested how people use AI on one class of reasoning problems, which supports part of the model without establishing a new cognitive architecture.
Systems
| System | Mode | Role | Weakness |
|---|---|---|---|
| System 1 | Human, fast, automatic, and associative | Produces intuitions and quick responses | Relies on shortcuts and can confidently produce a tempting wrong answer |
| System 2 | Human, slow, effortful, and reflective | Checks intuitions and applies rules | Is costly to engage and is easily avoided under pressure |
| System 3 | External, automated, data-driven, and interactive | Recognizes patterns, predicts, summarizes, and synthesizes | Can be fluent and confident even when its answer is wrong |
System 3 changes whether Systems 1 and 2 engage at all. It may replace an intuitive response, prompt deliberation by surfacing ambiguity, help with part of a problem, or supply the final judgment.
The authors describe several possible routes:
| Route | What happens | Who evaluates the answer |
|---|---|---|
| Intuition | System 1 responds | System 1 |
| Deliberation | System 1 flags a conflict and System 2 checks it | System 2 |
| Cognitive offloading | AI assists and the person checks its work | The person |
| Cognitive surrender | AI answers after a brief intuitive reaction | System 3 |
| Autopilot | AI receives the prompt and returns the response | System 3 |
In practice, a person may question an answer, ask the model again, verify it, override it, or rationalize it after deciding to accept it.
Cognitive Surrender
Cognitive surrender is the uncritical acceptance of an AI-generated answer in place of one's own reasoning. People are more likely to surrender when the model sounds confident, produces fluent explanations, and makes acceptance easy. Time pressure, limited subject knowledge, and prior trust in AI also increase the risk.
Cognitive offloading delegates part of a task while the person checks and integrates the result. Surrender transfers that evaluative role to the model. Consulting AI or following its answer does not by itself distinguish the two. The person must independently assess the answer to retain control.
Deference can be rational when a system reliably outperforms people in a structured domain. The danger lies in an uncalibrated or unnoticed transfer of judgment. The model's confidence substitutes for evidence, the person stops looking for conflicts, and responsibility becomes hard to locate.
Findings
Across three preregistered experiments, 1,372 participants completed 9,593 trials using seven Cognitive Reflection Test problems. Participants either reasoned alone or could consult GPT-4o. When available, the model was secretly instructed to give either the correct deliberative answer or a confident version of the problem's tempting intuitive answer.2
-
In the first study, accuracy was 45.8 percent without AI, 71.0 percent when the available AI was accurate, and 31.5 percent when it was faulty. On trials where participants consulted it, they adopted 92.7 percent of its correct advice and 79.8 percent of its faulty advice.
-
Across all three studies, participants surrendered to faulty advice on 73.2 percent of the trials in which they engaged with the chatbot. They successfully overrode it on 19.7 percent.
-
Access to AI raised confidence in the first study even when the model was wrong about half the time. Participants who already trusted AI accepted more bad advice. Those with stronger fluid reasoning or a higher need for cognition resisted it more often.
-
Accuracy incentives and immediate feedback helped. They increased rejection of faulty advice from 20.0 to 42.3 percent, though most faulty answers were still accepted.
The experiments define surrender as accepting deliberately faulty AI advice. They show uncritical adoption on these tasks. They do not measure a participant's motivation, sense of responsibility, or long-term dependence.
Judgement
Accuracy incentives, immediate feedback, visible uncertainty, and prompts to reflect can make System 2 more likely to engage. Before using AI, decide which claims require independent verification and name the person responsible for each consequential decision.
For important work:
- Write an initial view or evaluation criteria before asking the model.
- Ask for uncertainty, alternatives, and evidence instead of only a polished answer.
- Check consequential claims against an independent source.
- Reward accuracy and sound reasoning. Speed and volume are poor proxies for either.
- Preserve unaided practice for skills that must remain available without the tool.
The last point is a precaution rather than a finding from these studies. They measured immediate decisions in controlled tasks, not whether repeated AI use causes skills to atrophy.
Limitations
The evidence comes from a working paper that used one family of trick questions in short, controlled experiments. It cannot establish how often people surrender judgment in medicine, law, education, or ordinary knowledge work. The experiments also did not test long-term changes in trust, learning, or skill.
“System 3” names the role AI plays in a distributed decision but not a third biological system in the brain.