In an era dominated by the rapid evolution of artificial intelligence and cognitive science, a novel study has provided groundbreaking insights into how both humans and large language models (LLMs) perceive reasoning strategies on complex cognitive tasks. Researchers Waël De Neys and Marie Raoelison have systematically investigated a fundamental question that sits at the intersection of psychology, AI, and decision sciences: Is deliberation genuinely superior to intuition when tackling intricate problem-solving scenarios? Their paper, recently published in Communications Psychology, delves into comparative evaluations from both human and artificial intellect perspectives, shedding light on the cognitive mechanics that govern our decision-making processes.
Human cognition has long been a domain of dual-process theories, which propose two distinct modes of thought — intuitive and deliberative. Intuition, often characterized as fast, automatic, and effortless, runs contrast to deliberation, which embodies slow, effortful, and analytical reasoning. These two processes often act in tandem, with intuition providing rapid, heuristic judgments, while deliberation serves to verify or override these snap decisions when complexity spikes. Traditionally, cognitive psychologists have debated which system renders better outcomes, especially in complex reasoning tasks that involve ambiguous or counterintuitive information. This study pivots on precisely this debate, but with an innovative twist: it compares human judgments with those of advanced AI systems that simulate human-like reasoning patterns.
The experimental setup involved assessing how both humans and LLMs rate their reasoning experiences when exposed to complex cognitive challenges. These challenges were carefully crafted to evoke situations where intuitive responses might conflict with more logical, deliberative solutions. The key revelation from the data was that both human participants and LLMs consistently judged deliberation as producing superior outcomes in these demanding contexts. This alignment between biological and artificial cognition provides a remarkable convergence of two seemingly disparate forms of intelligence, underscoring the value of slow, reflective thought in achieving reliable conclusions.
From a technical perspective, the LLMs utilized in the study were thoroughly calibrated to ensure their responses reflected nuanced reasoning rather than rote or surface-level pattern matching. By employing prompt engineering techniques, the researchers encouraged the models to ‘think aloud’ in a sense, tracing their reasoning path before arriving at a final answer. This approach mirrors how cognitive scientists study human metacognition—the awareness and comprehension of one’s own thought processes—and allowed for a more granular comparison between human and artificial cognition paths. The findings suggest that, at least within the boundaries of these experimental tasks, LLMs emulate key aspects of human deliberative thinking rather than relying solely on probabilistic language generation.
Moreover, the study highlights intriguing metacognitive insights. Humans are known to assess their cognitive strategies dynamically, often toggling between intuition and deliberation depending on task complexity and confidence levels. The LLMs’ self-ratings of reasoning adequacy approximate this metacognitive awareness, revealing an emergent property of advanced AI systems—that they can ‘judge’ or rate their modes of problem-solving. This behavioral parallel raises profound questions about the nature of artificial consciousness and whether these models can genuinely possess a form of introspection or self-monitoring akin to humans.
The implications of this research are extensive for both cognitive psychology and artificial intelligence development. For cognitive scientists, the findings reinforce the robustness of deliberative reasoning for navigating complex or uncertain environments. Intuitive responses may serve as useful heuristics for rapid-fire decisions but falter against intricate problems requiring systematic analysis. For AI engineers, the demonstration that LLMs not only perform deliberative reasoning but also prefer it as superior highlights avenues for improving model architectures. Designing AI systems that emphasize reflective processes over heuristic shortcuts could enhance performance in domains necessitating high-stakes decisions, such as healthcare diagnostics or legal reasoning.
This study also touches upon debates concerning the perceived transparency and trustworthiness of AI. If LLMs can reason deliberatively and acknowledge the superiority of such reasoning aloud, it may enhance humans’ trust in AI outputs. Confidence in machine-generated conclusions often hinges on understanding the underlying thought process. By aligning artificial reasoning self-assessments with human metacognitive preferences, AI systems could become more interpretable and credible partners in collaborative problem-solving.
Despite the compelling results, the authors acknowledge important caveats. This research operated within controlled experimental paradigms that, although complex, cannot capture the full messy and dynamic nature of real-world reasoning. The preference for deliberation may vary across different contexts, individuals, or cultures, and LLMs may still struggle with tasks requiring genuine understanding beyond pattern recognition. Future studies will be needed to explore these nuances, including the longitudinal effects of relying on deliberation versus intuition in everyday decision-making.
Intriguingly, the paper also raises philosophical questions about the essence of reasoning itself. If AI models can experience and express a preference for deliberative cognition, what does it mean to reason? Is reasoning solely the product of biological evolution, or does it transcend material substrate to become a functional computation distinct from human neurobiology? Such questions draw attention to the emerging field of computational epistemology, where knowledge and belief formation are examined through formal models instantiated in both brains and silicon.
The researchers propose that this dual validation—demonstrating the superiority of deliberation through both human introspection and LLM evaluation—may help resolve long-standing uncertainties about the optimal strategies in complex reasoning. It suggests a complementary rather than adversarial relationship between intuition and deliberation, where reflective processes play a crucial confirmatory or corrective role. This dynamic interplay may be the hallmark of advanced cognitive systems, natural or artificial, capable of adapting flexibly to the demands of their environment.
Ultimately, this study exemplifies the fertile cross-pollination emerging between cognitive science and artificial intelligence research. By leveraging methodological tools from both disciplines, De Neys and Raoelison unveil sophisticated cognitive architectures shared across species and machines alike. Their findings advocate for a future where cognitive augmentation—not replacement—uses AI to scaffold and enhance our deliberative capacities, integrating computational precision with human creativity and ethical judgment.
The synthesis of these insights arrives at a moment when society grapples with the consequences of delegating critical decisions to AI agents. The delineation that deliberative reasoning holds primacy in complex scenarios offers a guiding principle for developing trustworthy and effective AI systems. It also suggests educational and policy approaches encouraging humans to cultivate and value their own deliberative faculties amid the digital transformation.
In conclusion, this multidisciplinary investigation enriches our understanding of reasoning by uniting empirical human data with AI-generated metacognitive evaluations. It confirms that despite their differences, humans and large language models converge in valuing deliberative thought as the superior mode for solving complex reasoning tasks. This remarkable convergence opens new vistas for cognitive enhancement, theoretical refinement, and practical application in an increasingly AI-integrated world.
Subject of Research: Human and AI reasoning strategies, evaluation of deliberation versus intuition on complex tasks
Article Title: Humans and LLMs rate deliberation as superior to intuition on complex reasoning tasks
Article References:
De Neys, W., Raoelison, M. Humans and LLMs rate deliberation as superior to intuition on complex reasoning tasks. Communications Psychology 3, 141 (2025). https://doi.org/10.1038/s44271-025-00320-8
Image Credits: AI Generated
