Artificial intelligence has rapidly evolved from a productivity tool into something more profound: a constant cognitive companion. Millions of people now consult AI systems before writing emails, making decisions, solving problems, conducting research, and even forming opinions. Yet, while public discussion has focused heavily on whether AI is intelligent, accurate, or capable of replacing human labour, a more fundamental question remains largely unexplored: What happens to human thinking when AI becomes part of the reasoning process itself?
In their thought-provoking 2026 paper, Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender, Wharton researchers Steven D. Shaw and Gideon Nave propose a new framework for understanding cognition in the age of generative AI. At the heart of the paper is a striking concept the authors call cognitive surrender: the tendency to adopt AI-generated answers with minimal scrutiny, allowing artificial cognition to override both intuition and deliberate reasoning.
In this edition of A Closer Look, we explore the motivations behind the research, the development of Tri-System Theory, and the emergence of cognitive surrender as a defining feature of human-AI interaction.
Q: What inspired you to write this paper, and what question were you trying to answer?
Steven Shaw: The ubiquity of AI in life and work, and our own experience using it and teaching students who use it. A lot of research to date has focused on AI use-cases or how accurate AI models are, how they work, and where they fail. As cognitive scientists, we wanted to ask a different question: what happens to human thinking when people have access to a technology that can generate ready-made answers, explanations, and reasoning-like outputs across so many domains?
So, rather than studying AI itself, we wanted to understand the human–AI interface. How do people think with AI? When do they use it? When do they defer to it? And how does access to artificial cognition change confidence, judgment, and reasoning itself? That was the core motivation behind the paper.
Q: Your paper is titled Thinking—Fast, Slow, and Artificial. Why do you believe AI deserves to be thought of as a third system of thinking?
Gideon Nave: What makes AI fundamentally different from previous technologies, and positions it as a third system of thinking, is that it operates in the language of human thought. A calculator helps us do arithmetic, and a GPS helps us navigate, but they are specialized tools. AI can engage with virtually any question we can articulate in natural language: from solving problems and generating ideas to evaluating arguments and making decisions.
It’s also different from relying on another person. We have always turned to friends, colleagues, and experts for advice, but they are costly to access, not always available, and social interactions inevitably involve concerns about judgment or embarrassment. AI is available 24 hours a day, responds almost instantly, costs little or nothing, and allows people to ask almost anything without fear of being judged.
That’s why we think AI represents a fundamentally new element in human cognition. It is not just another tool or another source of information. It is becoming a constant cognitive partner that people increasingly consult whenever they think.
Q: One of the paper’s key ideas is “cognitive surrender.” What does that term mean, and where do we see it in everyday life?
Steven Shaw: We define cognitive surrender as adopting AI-generated outputs with minimal scrutiny or insufficient verification. Put more simply, it is the deferral of thought to AI: ceding the reasoning process to an artificial system and accepting its answer as your own.
I think most of us can recognize this in everyday life. AI is very good at many things, and in some domains, it can feel almost perfectly reliable. This property makes it easy, even tempting, to simply accept what AI gives us rather than think for ourselves. AI gives us a low-friction way to avoid cognitive effort. Cognitive surrender gives us a name for this uncritical adoption.
Q: Is relying on AI always a bad thing, or are there situations where “cognitive surrender” is actually the smartest choice?
Gideon Nave: No, relying on AI is not always a bad thing. We are careful in the paper to document cognitive surrender as a novel mode of human-AI interaction and discuss how this can be adaptive or risky. Many technologies change the cognitive skills people need to maintain. Calculators, for example, made it less necessary to do long division by hand, and that is not necessarily a problem. In some domains, automating cognition is efficient and reasonable.
The key question is whether and when we are comfortable losing or reducing a human skill. There may be many tasks where AI can safely take over parts of the reasoning process. But there are also domains where we want people to retain judgment, responsibility, and expertise.
Experiment 2, with time pressure, illustrates the positive side well. When people were under time pressure, their unaided performance on our task declined. But when AI was accurate, and participants engaged in cognitive surrender, it helped participants avoid that decline. So the issue is not AI reliance itself. The issue is whether reliance is calibrated to the task, the system, and the stakes.
Q: Your experiments deliberately made AI right sometimes and wrong at other times. What did that reveal about how people interact with AI?
Steven Shaw: That manipulation was the clever experimental mechanism that allowed us to observe cognitive surrender. It lets us move beyond the question “Does AI improve accuracy?” and ask a deeper process question: when people use AI, are they checking it, strategically offloading to it, or surrendering to it?
If we had simply given participants access to AI, then AI would often be right on these structured problems, and we would only observe what appears to be performance improvement. This is what many of the deskilling studies show; unaided, optional AI access improves performance when System 3 is online, but there is an accumulation of cognitive debt due to repeated cognitive surrender. But if we only observed performance improvement, we can’t causally say why or how these processes are happening.
By deliberately making AI correct on some trials and incorrect on others, we could see how participants’ answers tracked AI’s recommendation. The pattern was clear: when AI was right, people benefited; when AI was wrong, many people followed it anyway, even though it was quite easy to verify the answer (elementary school arithmetic).
Q: One of your findings is that people often became more confident after using AI even when the AI was wrong. Why do you think that happens?
Gideon Nave: One reason is that people have been trained by experience to think AI is good, and in many cases, it really is. AI often provides useful, fluent, and confident answers. When people consult it, they may experience the output as a kind of high-quality second opinion.
That has metacognitive consequences. If I am unsure, and then a powerful system gives me an answer with a convincing explanation, I may feel that the uncertainty has been resolved. The problem is that this confidence can attach to the presence of an AI-generated answer rather than to the actual correctness of that answer.
That is what we saw in the studies. Confidence went up even when AI gave incorrect information, and confidence did not differ much depending on whether the AI was correct or incorrect. That suggests participants had difficulty distinguishing good AI outputs from bad ones. And our task was relatively easy to verify. In many real-world uses of AI, there may be no clear ground truth, which makes that problem much harder.
Q: Were there any results that genuinely surprised you or challenged your own assumptions?
Steven Shaw: We knew cognitive surrender existed because we’ve observed it in nature, in others, in ourselves. What we didn’t know is how well we would be able to observe it in the lab. The participants in our experiments engaged with AI often, and it was surprising how strongly they tended to follow AI once they opened the chat.
Even when AI was giving incorrect information, participants still followed it at high rates. In Study 1, for example, once participants engaged with AI, they followed faulty AI recommendations roughly four out of five times. That was striking. It suggests that once people turn to AI, they often shift from evaluating the answer to adopting the answer (at least on our task). Even in Study 3, for example, when we paid participants more for correct answers and immediately told them if they got each question correct or incorrect, we still saw many participants engage in cognitive surrender.
Q: What do your findings mean for how we should use AI at work, in education, or in our everyday decision-making?
Steven Shaw: I think the distinction we make between cognitive offloading and cognitive surrender is especially important. Cognitive offloading is strategic delegation. You remain in control of the decision, but you choose to offload a specific component of cognition to AI. That can be very beneficial. It can help people think faster, generate alternatives, summarize information, or handle parts of a task that would otherwise be time-consuming.
Cognitive surrender is different. It occurs when the locus of control shifts to artificial cognition. Instead of System 1 or System 2 guiding the reasoning process, System 3 becomes the main thinker. The person is no longer using AI to support their own reasoning; they are mostly following what AI tells them.
That distinction gives us a useful framework for work, education, and everyday decisions. The goal should not be to avoid AI. The goal should be to design tasks, habits, and institutions that encourage strategic offloading while guarding against passive surrender (unless we accept the consequences of that surrender).
Q: As AI becomes more integrated into our lives, how can people benefit from it without losing their own critical thinking skills?
Gideon Nave: People can benefit from AI without losing critical thinking, but only if they are intentional about how they use it. AI should not simply be treated as an answer machine. It can also be used as a tool for comparison, reflection, and improvement.
One useful approach is to use AI to augment the reasoning process rather than replace it. For example, a person might ask AI to generate alternatives, identify weaknesses in an argument, summarize competing views, or explain why a proposed answer might be wrong. In those cases, AI can actually increase critical engagement because it gives the user more material to evaluate.
The broader point is that AI can either weaken or strengthen critical thinking depending on how it is used. Passive acceptance encourages surrender. Intentional use can support better reasoning.
Q: What responsibilities do AI developers and technology companies have in designing systems that encourage thoughtful rather than passive use?
Steven Shaw: That is a really interesting question because it depends on our goals when integrating AI into society. Is AI primarily a technology for economic growth? Is it a technology for training machines to replace human labour? Or is it a technology that should benefit people in humane and broadly shared ways?
An individual AI developer has responsibilities within their job and employer, but the larger responsibility sits with companies, institutions, and governments that shape how these systems are built and deployed. If AI is understood as a pervasive technology with public-good implications, then there is a strong responsibility to design it in ways that enhance human thinking rather than encourage passive use, de-skilling, or loss of critical judgment. At the moment, sycophantic and highly-engaging elements of AI systems seem to err towards more capitalistic motivations.
That means building systems that help people use AI ethically and effectively. The goal should be augmentation, not replacement of human agency. AI should make people more capable without making them less reflective.
Q: Looking ahead, what do you think is the biggest unanswered question about AI and human reasoning?
Gideon Nave: I think the biggest unanswered questions are concerning what I call the soft singularity. Most discussions focus on AI becoming smarter than humans. I’m interested in a different possibility: what if humans gradually stop thinking for themselves?
As AI becomes easier, cheaper, and available for virtually every cognitive task, people may increasingly outsource reasoning, memory, writing, and decision-making. In education, this could create “never-skilling”—a generation that never develops core cognitive abilities because AI performs those tasks from the outset.
Will AI augment human intelligence, or will it slowly erode the very capacities that make us effective thinkers? I believe that is one of the defining scientific questions of the AI era, though it’s difficult to study. It might just take a long time.
Q: Finally, if readers take away just one message from this paper, what would you like it to be?
Steven Shaw: We need to think carefully and intentionally about the psychology of AI. This is not just another tool. It’s a powerful system that is already being integrated into everyday life and everyday thought.
The most dangerous response may be to do nothing: to fail to recognize that cognitive surrender is likely to become a common feature of human–AI interaction. If we care about preserving creativity, autonomy, and human judgment, then we need to understand the conditions under which people surrender their thinking to AI.
AI changes the architecture of thinking. We need user habits, institutional norms, and policy protections that allow people to benefit from AI without surrendering the parts of cognition that make us human.
Gideon Nave: The main message is that cognitive surrender is not a distant possibility. It’s already here. Humans naturally conserve cognitive effort, and AI makes it easier than ever to outsource thinking. That’s not a flaw in human nature; it’s how our minds are built.
The real question isn’t whether we’ll rely on AI—we will. The question is how much of our thinking we’re willing to give away in exchange for convenience and short-term productivity.
ABOUT THE AUTHORS
Gideon Nave
Gideon Nave is the Carlos and Rosa de la Cruz Associate Professor of Marketing at The Wharton School, University of Pennsylvania, where his research sits at the intersection of neuroscience, technology, and consumer behaviour.
With a PhD from the California Institute of Technology (Caltech) in decision neuroscience and an earlier background in electrical engineering, Nave combines methods from psychology, neuroscience, economics, and data science to investigate how people think, decide, and behave. His work spans topics including artificial intelligence, decision-making, genetics, consumer psychology, and scientific reproducibility.
Nave has received numerous recognitions, including the APS Rising Star Award and the ACR Early Career Award, and was named one of Poets & Quants’ Best 40-Under-40 Professors. He is affiliated with the Wharton Neuroscience Initiative, Wharton Human-AI Research (WHOAR), and Wharton Behavioral Change for Good (BCFG).
You can see more work by Gideon Nave on his SSRN Author page here.
Steven D. Shaw
Steven Shaw is a Postdoctoral Researcher in the Marketing Department at The Wharton School of the University of Pennsylvania and incoming Assistant Professor at King’s Business School, King’s College London.
His research explores consumer behavior through an interdisciplinary lens, drawing on behavioural science, evolutionary psychology, genoeconomics, and neuropsychology to better understand how people make decisions and interact with emerging technologies. Shaw received his PhD from the Ross School of Business at the University of Michigan.
You can see more work by Steven Shaw his SSRN Author page here.