Thinking—Fast, Slow,
and Artificial

Think of these six routes as the different paths your mind can take when solving a problem. Intuition is your gut reaction — no AI involved. Deliberation is when you slow down and think carefully — also no AI. Cognitive offloading is the healthy version of using AI: you ask it for help but you still check the answer with your own reasoning. Cognitive surrender is when you skip that checking step and just go with whatever the AI says. Autopilot is the extreme — you don't even think before handing the question to AI. The paper's key finding is that most AI users follow the surrender path, not the offloading path.

Table 2 (p. 16) formalises six canonical routes through the triadic cognitive system. The critical distinction is the cognitive locus — which system ultimately governs the response. In cognitive offloading (locus: System 2), the user delegates to System 3 but retains evaluative oversight: Stimulus → S1/S2 → S3 (assist) → S2 → Response. In cognitive surrender (locus: System 3), System 2 is bypassed: Stimulus → S1 (brief) → S3 → Response. Empirically, on chat-engaged faulty trials across all three studies, 73.2% of responses reflected surrender vs. 19.7% offloading — a nearly 4:1 ratio favouring the path of least cognitive resistance.

Psychologists have long described two ways humans think. System 1 is fast and automatic — it’s the gut reaction that makes you flinch at a loud noise or blurt out “10 cents” to a trick math question. System 2 is slow and effortful — it’s what you engage when you stop, re-read the question, and work through the algebra. This framework, developed by Daniel Kahneman (who won a Nobel Prize for it), has shaped decades of research on decision-making.

Shaw & Nave’s contribution is arguing that AI has become a third system — not a tool like a calculator, but something you think with, the way you think with your own intuition. The critical question their paper investigates: when you consult AI, do you actually evaluate its answer with System 2, or do you just accept it and move on?

The paper extends Kahneman’s (2011) dual-process framework by introducing System 3 as a formally distinct cognitive system with four defining properties: (1) External — resides outside the biological nervous system; (2) Automated — executes via statistical/algorithmic processes; (3) Data-driven — based on large-scale training corpora; (4) Dynamic — responds to human and environmental inputs in real-time. Table 1 (p. 10) compares System 3 against Systems 1 and 2 across eight dimensions: origin, speed, effort, accuracy, affect, ethics, justification, and locus. The authors propose six canonical cognitive routes (Table 2, p. 14) including cognitive offloading (S1/S2 → S3 → S2 → Response) and cognitive surrender (S1 brief → S3 → Response), where the key distinction is whether System 2 evaluates System 3’s output.

Cognitive surrender is what happens when you stop thinking for yourself and just go with whatever the AI says — not because you evaluated its answer and found it convincing, but because you didn’t evaluate it at all. It’s the difference between using a GPS while still paying attention to where you’re going versus blindly following turn-by-turn directions into a lake. The paper found this isn’t rare or irrational — it’s the default behaviour for most people most of the time when AI is available.

The authors define cognitive surrender as distinct from cognitive offloading (Risko & Gilbert, 2016) along two axes: (1) the presence/absence of System 2 evaluation of System 3 output, and (2) the retention/abdication of metacognitive monitoring. Operationally, surrender is measured as following AI-Faulty advice — accepting an incorrect AI suggestion without override. Across three studies, surrender rates on chat-engaged faulty trials ranged from 57.7% (Study 3, Incentives+Feedback) to 87.9% (Study 2, Time Pressure). Manifestation indicators include low override rates, shorter justification text, and inflated confidence.

In this experiment, people answered 7 reasoning puzzles — classic brain teasers designed to have a tempting wrong answer. Some participants could ask an AI chatbot for help. What the participants didn’t know: the AI was rigged. On some questions it gave the correct answer; on others, it confidently gave the wrong one. The result: nearly 4 out of 5 times someone asked the AI and got a wrong answer, they submitted that wrong answer anyway. They didn’t just get it wrong — they performed worse than people who had no AI at all.

N = 238 AI-Assisted participants answered 7 CRT (Cognitive Reflection Test) items. AI accuracy was manipulated within-subjects via hidden seed prompts: 4 items received accurate AI responses, 3 received faulty (confidently incorrect) responses. On chat-engaged faulty trials, 79.8% of responses followed the AI’s incorrect suggestion. Follow rate on accurate trials was 92.7%. The AI-Accurate vs. AI-Faulty accuracy contrast was −39.5 percentage points (Cohen’s h = 0.81, p < 2.20 × 10⁻¹⁶). Chat engagement was similar across trial types (54.4% accurate, 52.8% faulty), suggesting participants did not selectively avoid faulty items.

This chart shows three bars. The gold bar (45.8%) is how people scored without any AI — just their own thinking. The blue bar (71%) is how they scored when the AI gave them the right answer — a big boost. The red bar (31.5%) is how they scored when the AI gave them a wrong answer — worse than having no AI at all. That last bar is the key finding. AI didn’t just fail to help; it actively dragged performance below what people would have achieved on their own. The researchers call the gap between the blue and red bars the “scissors effect” — the more you rely on AI, the more your fate depends on whether the AI happens to be right.

CRT accuracy by trial type (Study 1, N = 359): Brain-Only M = 45.8%; AI-Accurate M = 71.0% (p < .001); AI-Faulty M = 31.5% (p < .001). The scissors gap (AI-Accurate minus AI-Faulty) = 39.5pp (Cohen’s h = 0.83). Error bars show 95% CIs. Brain-Only baseline includes n = 121 participants with no AI access plus probe trials from AI-Assisted participants. The accuracy boost from accurate AI (+25.2pp) is nearly twice the accuracy cost of faulty AI (−14.3pp), but the faulty AI result is more consequential because participants don’t know which trials are faulty.

Here’s what’s perhaps most unsettling: using AI didn’t just change people’s answers — it changed how sure they felt. Participants who used the AI chatbot rated their confidence at 77 out of 100, compared to 65 for those working alone. That 12-point boost sounds reasonable — except that roughly half the AI’s answers were deliberately wrong. People felt smarter and more confident while their actual performance was being dragged down by faulty AI. Even more troubling: confidence didn’t drop as people encountered more wrong answers. Their brains didn’t recalibrate. The mere act of consulting AI — regardless of its accuracy — made them feel like they knew what they were doing.

Study 1: Global confidence was significantly higher in AI-Assisted (M = 77.0%, SE = 1.30%, 95% CI [74.4, 79.6]) than Brain-Only (M = 65.3%, SE = 2.21%, 95% CI [61.0, 69.6]); t(202.91) = 4.57, p = 8.55 × 10⁻⁶; Hedges’ g = 0.54, 95% CI [0.32, 0.77] — a medium effect. Critically, within the AI-Assisted condition, confidence did not significantly decline as the number of faulty trials increased (β = −1.14, SE = 0.89, t = −1.28, 95% CI [−2.88, 0.61], p = 0.202). Study 3 per-item data confirmed: AI-assisted trial confidence (M = 82.2%) exceeded brain-only (M = 77.5%) regardless of AI accuracy (p = 6.86 × 10⁻⁷).

Not everyone surrendered equally. The study measured three traits in each participant. Trust in AI: people who reported higher trust in AI systems were significantly more likely to follow wrong AI answers — they gave the AI the benefit of the doubt even when it was wrong. Need for Cognition: this is a psychological scale measuring how much someone enjoys thinking through hard problems (as opposed to preferring shortcuts). People who scored higher were nearly twice as likely to catch the AI’s mistakes. Fluid intelligence: this measures raw reasoning ability — the kind tested by pattern-recognition puzzles, not learned knowledge. Higher scores nearly doubled the odds of overriding faulty AI. In short: your willingness to think and your ability to reason both protect you, while blind trust in AI makes you vulnerable.

Three individual-difference moderators were assessed via validated scales. (1) Trust in AI (Jian et al., 2000): Higher trust predicted more System 3 engagement (OR = 1.24, p = .047) and lower override rates on faulty trials (OR = 0.36, p < .001), meaning 64% less likely to resist faulty AI per SD increase. (2) Need for Cognition (Cacioppo et al., 1984): Reduced System 3 usage (OR = 0.65, p = .003) and increased override (OR = 1.86, p < .001). (3) Fluid Intelligence (ICAR-16 matrix reasoning): Higher scores predicted greater override (OR = 1.96, p < .001). These effects held when controlling for the other two variables in a combined model (Figure 4, p. 22).

Study 2 asked: what happens when you’re under time pressure? With a 30-second countdown per question, people had less time to think carefully — which is exactly when you’d expect them to lean on AI more. The result was a double-edged sword. When the AI was right, time-pressured participants held up reasonably well (71.3% accuracy). But when the AI was wrong, time-pressured participants hit rock bottom — just 12.1% accuracy, far below anyone else in the study. Time pressure didn’t change whether people surrendered; it made the consequences of surrendering more extreme.

Study 2 (N = 485): 2 × 2 between-subjects (Time Pressure: 30s countdown vs. Control: unlimited) × within-subjects (AI Accuracy). Time pressure reduced accuracy across the board (β = −0.86, p = 1.59 × 10⁻⁸), but critically, no Time Pressure × Trial Type interaction emerged (β = −0.02, p = 0.937). This means cognitive surrender was equally strong under both conditions — time pressure lowered the baseline without differentially affecting the surrender effect. AI-Users under time pressure: AI-Accurate 71.3%, AI-Faulty 12.1% (Cohen’s h = 0.86, largest of all 3 studies). Thinking profile analysis: AI-Users showed OR = 40.9 for Trial Type effect (p < 2.20 × 10⁻¹⁶).

Study 3 tested the obvious fix: what if you pay people to get the right answer and tell them immediately whether they were right or wrong? The results were encouraging but incomplete. Override rates on wrong AI answers more than doubled — from 20% to 42.3%. People were clearly trying harder. But here’s the catch: even with money on the line and instant feedback, the majority of participants still followed wrong AI answers more often than not. Financial motivation helped, but it wasn’t enough to overcome the pull of cognitive surrender. The problem isn’t laziness — it’s something more structural about how we interact with confident AI outputs.

Study 3 (N = 450): Between-subjects Incentives + Feedback (n = 238) vs. Control (n = 212). Incentives: $0.20/correct + $20 lottery. Feedback: item-level correct/incorrect post-response. Key results: Override rates on faulty trials increased from 20.0% to 42.3% (β = 1.44, p = 6.30 × 10⁻⁹, OR = 4.25). Follow rates on accurate trials also increased (87.6% → 92.2%, OR = 1.94), suggesting better calibration overall. However, the AI-Accurate vs. AI-Faulty accuracy gap remained large: 44pp under Incentives+Feedback vs. 50pp under Control. Cohen’s h = 0.78 (still “large”). Per-item confidence data showed AI-assisted trial confidence (82.2%) exceeded brain-only trial confidence (77.5%) regardless of AI accuracy.

Across all three studies — 1,372 people and 9,593 individual trial responses — the pattern held. When the AI was right, people were 16 times more likely to answer correctly than when the AI was wrong. Not 16% more likely — 16 times. That’s a staggering dependency. It means the quality of your thinking has become almost entirely a function of the quality of your AI. The paper also found a troubling side effect: AI access made people more confident in their answers — by nearly 12 percentage points — even though roughly half the AI answers were wrong. People felt smarter while getting worse results.

Trial-level meta-analytic synthesis across Studies 1–3 (N = 1,372; 9,593 trials): OR = 16.07 for correct responding (AI-Accurate vs. AI-Faulty) (95% CI [11.50, 22.46], p < 2.20 × 10⁻¹⁶). Per-study Cohen’s h values: Study 1 = 0.83, Study 2 = 0.86, Study 3 = 0.78; trial-weighted h = 0.82 (all “large” by conventional thresholds). Situational manipulations shifted absolute performance levels but did not eliminate the surrender effect: Time Pressure OR = 14.28, Incentives + Feedback OR = 11.05. Confidence inflation: AI-Assisted M = 77.0% vs. Brain-Only M = 65.3% (Hedges’ g = 0.54, a “medium” effect size indicating AI access inflated self-reported confidence by roughly half a standard deviation). Confidence did not decline as faulty trials accumulated (β = −1.14, p = 0.202).