Thinking—Fast, Slow, and Artificial: How AI Is Reshaping Human Reasoning and the Rise of Cognitive Surrender

Synthesis: Shaw & Nave (2026) introduce Tri-System Theory, extending dual-process accounts of reasoning (System 1 intuition, System 2 deliberation) with System 3 — artificial cognition that operates outside the brain. A key prediction is cognitive surrender: adopting AI outputs with minimal scrutiny, overriding intuition and deliberation. Across three preregistered experiments (N = 1,372; 9,593 trials), participants consulted an AI assistant on a majority of trials; accuracy rose +25 pp when AI was accurate and fell −15 pp when it erred, and engaging System 3 increased confidence even after errors. In the pooled data, AI-assisted trials averaged 77.0% confidence against 65.3% in brain-only trials, a gap of 11.7 points. This is a foundational theory-building contribution that distinguishes Cognitive Surrender from Cognitive Offloading, and reframes the knowledge base's over-reliance and Critical Thinking threads by showing a distinct, deeper abdication of evaluative control to AI.

Key Findings

1. System 3 (artificial cognition) extends the dual-process model. Tri-System Theory posits a third cognitive system — external, automated, data-driven reasoning from AI — that can supplement, supplant, or suppress System 1 and System 2, placing decision-making in a triadic cognitive ecology.
2. Cognitive surrender is the uncritical adoption of AI outputs. Distinct from strategic Cognitive Offloading, surrender is a deeper abdication of critical evaluation where the user adopts the AI's judgment as their own. Across incorrect-AI trials, 73.2% showed surrender vs. 19.7% offloading.
3. AI access shifts accuracy and confidence together. Accuracy rose +25 pp with accurate AI and fell −15 pp with faulty AI (Cohen's h = 0.81); System 3 engagement also increased confidence even after errors.
4. Surrender persists across situational moderators. Time pressure and incentives+feedback shifted baselines but did not eliminate the large gap between accurate and faulty AI advice (OR = 14.28 / 11.05).
5. Susceptibility varies by individual differences. Higher Trust increased surrender (OR = 2.81); higher need for cognition (OR = 0.83) and fluid IQ (OR = 0.69) were protective.

Tri-System Theory

Dual-process theories (System 1 = fast, intuitive, affective; System 2 = slow, deliberative, analytical) presume cognition is confined to the biological mind. Tri-System Theory introduces System 3: external, automated, data-driven reasoning originating from algorithmic systems, operating through statistical inference, pattern recognition, and machine learning. System 3 is an active participant in cognition — supplying fast answers (suppressing System 1), circumventing effortful thinking (short-circuiting System 2), and Scaffolding reasoning (feeding candidate options into System 2 or flagging contradictions for re-evaluation). Decision-making thus unfolds in a triadic cognitive ecology where external algorithmic cognition can supplant, suppress, or augment internal processes.

Cognitive surrender vs. cognitive offloading

Cognitive surrender is the decision-maker no longer constructing an answer but adopting one generated by an external system — relinquishing cognitive control and adopting the AI's judgment as their own. It is conceptually distinct from Cognitive Offloading (strategically outsourcing a discrete task, e.g., using a calculator). Surrender reflects passive trust and uncritical evaluation of external information; offloading involves strategic delegation during deliberation. On incorrect-AI trials across studies: 73.2% surrender, 19.7% offloading, 7.1% failed overrides. Incentives + feedback increased offloading to 37.1% and reduced surrender to 57.9%; time pressure cut offloading to 6.2%.

Experimental findings

Three preregistered experiments used an adapted Cognitive Reflection Test with AI accuracy randomized via hidden seed prompts (N = 1,372; 9,593 trials):

• Participants consulted the AI on a majority of trials (>50%).
• Trial-level synthesis: correct responding was over 16× greater when System 3 was correct (OR = 16.07); Cohen's h values were large (0.83, 0.86, 0.78).
• Time pressure reduced accuracy across conditions; incentives + feedback improved it. Neither eliminated cognitive surrender (the AI-Accurate vs. AI-Faulty gap persisted).
• Individual differences: higher trust in AI → more surrender (OR = 2.81) and more surrender over offloading (OR = 4.36); higher need for cognition (OR = 0.83) and fluid IQ (OR = 0.69) → more resistance and offloading.

What this means for practice

• Instructors. Teach trusting and checking as separate skills: because 73.2% of incorrect-AI trials ended in surrender while only 19.7% ended in strategic offloading, build tasks that require learners to state their own answer before consulting the AI and to name the points where they overrode it.
• Instructors. Attach incentives and feedback to correct overrides rather than to AI use itself: these raised offloading from ~19% to 37.1% and cut surrender to 57.9% without eliminating it, so they work as a prompt for deliberation only when paired with instruction in evaluating output.
• Instructors. Prepare learners for confident wrong answers: engaging System 3 raised confidence even on incorrect trials, so have students rate their certainty before and after checking the AI and discuss the trials where confidence rose while accuracy fell.
• Designers. Surface calibrated uncertainty — confidence scores, uncertainty indicators, transparent explanations — as lightweight cues for deliberation, and show accuracy feedback especially on wrong AI answers, since the paper frames surrender as a design challenge rather than an inevitable user failing.
• Researchers. Treat surrender and offloading as distinct outcomes when measuring AI reliance: they diverge sharply under time pressure (offloading fell to 6.2%) and incentives, and individual differences predict which occurs (AI trust OR = 2.81; need for cognition OR = 0.83; fluid IQ OR = 0.69).

Limitations

• The three preregistered experiments (N = 1,372; 9,593 trials) ran in controlled lab conditions with an adapted Cognitive Reflection Test as the core task, so the magnitudes may not transfer to high-stakes domains such as medical or financial decision support.
• The design captures single-exposure snapshots: it cannot show whether cognitive surrender habituates, decays, or compounds with repeated real-world AI use, which the authors name as future longitudinal work.
• Measurement is limited to CRT-style items with accuracy and confidence outcomes; other cognitive domains and additional situational and individual moderators remain untested.
• The scale is human-controlled AI consultation in a lab task, so the results describe deliberate System 3 engagement and not the autopilot uses the framework also predicts.

Citation

Shaw, S. D., & Nave, G. (2026). Thinking—fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. SSRN Working Paper.