Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender
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. https://doi.org/10.2139/ssrn.6097646
Summary. This working paper extends dual-process reasoning theory by proposing System 3: external artificial cognition that can supplement or supplant human intuition and deliberation. Across three preregistered CRT-style experiments with 1,372 participants and 9,593 trials, the authors test “cognitive surrender,” defined as adopting AI outputs with minimal scrutiny. AI access improved accuracy when the AI was correct but reliably harmed accuracy when the AI was confidently wrong, while often increasing confidence.
Key claims
- Tri-System Theory treats generative AI as an external reasoning system that can scaffold human deliberation, bypass it through cognitive offloading, or substitute for judgment through cognitive surrender.
- In Study 1, participants used the AI assistant on 54.4% of AI-Accurate trials and 52.8% of AI-Faulty trials; when they opened the chat, they followed correct AI advice on 92.7% of trials and faulty AI advice on 79.8% of trials.
- Study 1 showed the core cognitive-surrender pattern: accuracy was 45.8% in the Brain-Only condition, 71.0% on AI-Accurate trials, and 31.5% on AI-Faulty trials, with a large AI-Accurate versus AI-Faulty contrast (Cohen’s h = 0.81).
- AI access increased confidence in Study 1 by 11.7 percentage points even though approximately half of AI outputs were faulty, and confidence did not significantly decline as faulty trials accumulated.
- Higher trust in AI predicted more chat use, lower accuracy on AI-Faulty trials, and greater following of faulty AI recommendations; higher fluid IQ and need for cognition buffered performance on AI-Faulty trials.
- In Study 2, time pressure reduced baseline reasoning performance, but among AI-Users performance tracked System 3 output quality: AI-Users under time pressure had 71.3% accuracy on AI-Accurate trials and 12.1% accuracy on AI-Faulty trials.
- In Study 3, incentives plus immediate feedback more than doubled override rates on faulty AI advice from 20.0% to 42.3%, but cognitive surrender persisted: AI-Faulty accuracy remained lower than AI-Accurate accuracy and AI-Users still showed a large trial-type asymmetry.
Evidence & limitations
- Design: three preregistered experiments using adapted Cognitive Reflection Test items with randomized AI accuracy via hidden seed prompts; samples came from university-affiliated behavioral labs and Prolific.
- Strength: experimentally manipulates AI correctness within task trials, allowing the source to separate beneficial AI assistance from harmful uncritical adoption rather than treating AI use as uniformly good or bad.
- Limitations: reasoning tasks are stylized CRT items rather than workplace decisions; the AI was deliberately made accurate or faulty through hidden prompts; samples and short experimental sessions do not show whether cognitive surrender persists in richer work systems with stakes, training, accountability, or repeated feedback.
- Provenance: Crossref identifies both an SSRN DOI and a PsyArXiv/OSF DOI for the same title and author team. The Drive PDF is text-equivalent to the public OSF PDF at https://osf.io/yk25n/download with the same page count and text length, but the binary SHA-256 hashes differ, so the Drive hash above records the ingested file rather than a public checksum match.