Abstract
Artificial intelligence tools, from search engines to generative language models, now sit inside everyday thinking. This review gathers what psychology and neuroscience currently suggest about how these tools affect human cognition, and it does so with deliberate caution. The strongest evidence concerns cognitive offloading: when people expect information to remain available, they remember where to find it rather than the content itself, a pattern documented in transactive memory research. Related work touches attention and task-switching, over-reliance and automation bias, the conditions under which learning takes hold, developmental questions for children and adolescents, and effects on mood and wellbeing. The picture is mixed rather than uniformly negative. AI can extend what a person accomplishes and can support learning when used well, yet the same tools can weaken skills that are never practiced and can encourage passive acceptance of machine output. Much of the direct evidence on generative AI is early, correlational, and contested. We argue for a stance of informed, unhurried use rather than alarm or uncritical enthusiasm, and we offer practical guidance for families, schools, and clinicians grounded in established cognitive science.
Keywords: cognitive offloading, transactive memory, automation bias, metacognition, desirable difficulties, executive function, attention, generative AI
Introduction
Scope of this review
This paper reviews how interactive AI tools relate to human thinking. We include search engines, recommendation systems, navigation aids, and the newer generative systems that produce text, images, and code on request. We focus on cognition in the ordinary sense: remembering, paying attention, reasoning, judging, and learning. We also consider mood and wellbeing, because emotional states shape how well a mind works. We do not attempt a technical account of how AI systems function, and we do not make claims about machine consciousness.
Why this matters now
The reason for attention is not that AI is uniquely dangerous to the brain. It is that these tools are used constantly, by very large numbers of people, often from a young age, and often in the middle of important cognitive work such as studying, writing, and deciding. Small effects, repeated daily across years, can matter. The people who care for and teach others, parents, educators, clinicians, and policymakers, are asked to make decisions now, before the science is settled. This review is written for them.
How to read the evidence
Three cautions apply throughout. First, most studies show correlation, not causation. Finding that heavier tool use goes together with weaker memory does not prove the tool caused the weakness, because people who already struggle may reach for tools more often. Second, the technology is young. Careful long-term studies of generative AI barely exist yet, so many claims rest on short laboratory tasks or on older research about earlier tools such as web search and GPS. Third, effects depend heavily on how a tool is used. The same system can support thinking or replace it. Where the evidence is thin, we say so plainly.
Cognitive Offloading and Memory
Figure 1. Some knowledge is worth holding internally because you reason with it in the moment. Other information is fine to look up.
The Google effect and transactive memory
The most studied cognitive consequence of digital tools is offloading, the practice of storing information outside the head rather than inside it. Betsy Sparrow, Jenny Liu, and Daniel Wegner reported in 2011 that when people expected to have later access to facts, they remembered the facts less well but remembered where to find them better. This came to be called the Google effect. It builds on Wegner's older idea of transactive memory, the way people in couples and teams divide up remembering so that each person holds part of the whole and knows who holds the rest. A search engine, on this view, becomes a memory partner. The partnership is not new in kind. Humans have offloaded memory onto notebooks, address books, and knowledgeable friends for a very long time.
What offloading costs and what it saves
Offloading is not simply loss. Freeing attention from storing trivia can leave room for other work, and no one seriously proposes memorizing every phone number again. The concern is more specific. Knowledge held in the head is what a person reasons with in the moment. Rich internal knowledge lets someone notice when something does not add up, form connections across topics, and think when no device is at hand. If facts are never encoded, they cannot be recombined into understanding later. The practical question is therefore not whether to offload, but which things are worth keeping inside. Core knowledge in a field, the material a person reasons from often, is worth holding internally even when it could be looked up.
Attention and Task-Switching
Concern about attention predates AI and centers on the wider design of digital media. Nicholas Carr's popular 2010 book The Shallows argued that constant linking, alerting, and switching trains a scattered, skimming style of reading and makes sustained concentration harder. This is a framing rather than a settled finding, and it is debated. What laboratory research does show fairly consistently is that switching between tasks carries a cost. Each switch takes time to reorient, and so-called multitasking is usually rapid switching that leaves each task done less well. Interruptions, including self-interruptions to check a device, fragment the deep attention that hard reading and hard thinking require.
AI tools intersect with this in two ways. Some are delivered inside the same attention-competing environments of feeds and notifications, and they inherit those problems. Others, by contrast, may reduce switching, for example when a single assistant answers a question in place that would once have required opening many tabs. The net effect on attention is not established and probably depends on the specific tool and habit. Recommendation systems that optimize for engagement deserve particular caution, since holding attention is their explicit goal and that goal is not the same as serving the user's own aims.
Over-Reliance, Automation Bias, and Metacognition
Figure 2. Automation bias has two failure modes, and both are strongest when a person is tired or rushed.
Automation bias
A large body of research from aviation, medicine, and other high-stakes fields describes automation bias: the tendency to over-trust automated output and to under-check it. Two failure patterns recur. People accept a machine's recommendation that is wrong (a commission error), and people fail to act because the machine did not prompt them (an omission error). Automation bias grows when a person is busy, tired, or facing a task at the edge of their competence, which is often exactly when a tool is used. Generative AI raises the stakes because its output is fluent and confident in tone even when it is wrong, and because it can produce plausible fabrications, sometimes called hallucinations. Fluency is persuasive. Text that reads well is trusted more than it should be.
Metacognition and the illusion of understanding
Metacognition is thinking about one's own thinking: knowing what you know, judging how well you understand, and noticing when you are lost. Tools can distort these judgments. When information is easy to retrieve, people tend to overestimate their own knowledge, confusing access to an answer with understanding of it. Studies of internet search have found that searching for explanations can inflate a person's sense of their own internal knowledge, even about unrelated topics. With generative AI, a person can obtain a finished-looking essay or solution without doing the work that normally builds and reveals understanding. The risk is a widening gap between felt competence and real competence. Preserving metacognition means keeping moments where a person must produce an answer unaided and see whether they actually can.
Learning and Skill Acquisition
Figure 3. Practices that feel harder in the moment, such as retrieval, spacing, and interleaving, build more durable and flexible knowledge.
Cognitive load and worked examples
Learning research offers useful structure here. John Sweller's cognitive load theory starts from the fact that working memory is small and easily overwhelmed. Good instruction manages load so that mental effort goes toward the material rather than toward struggling with the format. For novices, worked examples and clear guidance usually beat unsupported problem-solving, because they do not swamp working memory. Used this way, an AI tutor that explains a step, offers a worked example, or breaks a problem into parts can support learning, especially for a beginner who would otherwise be stuck. This is a genuine benefit and should not be dismissed.
Desirable difficulties
The complication is that not all effort is wasteful load. Robert and Elizabeth Bjork's work on desirable difficulties shows that certain kinds of struggle make learning stick. Retrieving an answer from memory rather than rereading it, spacing practice over time rather than massing it, and mixing problem types rather than blocking them all feel harder and produce slower apparent progress, yet they build far more durable and flexible knowledge. Here AI can quietly undercut learning. A tool that supplies the answer at the first moment of difficulty removes the retrieval effort that would have strengthened memory. The effort felt unpleasant, so its loss is not noticed, but the learning is lost with it. The design goal for education is to keep the difficulties that are desirable while removing the ones that merely obstruct.
Skills that are never practiced
A further concern is deskilling. A skill maintained only by use will fade if a tool always performs it. Studies of satellite navigation, for instance, suggest that habitual reliance can weaken a person's own sense of direction and active engagement of spatial memory. By analogy, if writing, arithmetic, summarizing, or basic coding are always handed to a machine, the underlying skill may not develop in the young or may erode in adults. This does not argue against ever using the tool. It argues for deciding which skills a person still needs to own, and protecting practice of those.
Developmental Considerations for Children and Adolescents
Children and adolescents deserve separate treatment because their brains are still forming. Childhood and the teenage years are periods of high neuroplasticity, when experience strongly shapes the wiring of attention, memory, language, and self-control. The executive functions, which include working memory, inhibitory control, and cognitive flexibility, develop gradually across childhood and into the mid-twenties, supported by the slow maturation of the prefrontal cortex. These capacities are built through practice: waiting, effortful focus, wrestling with a hard problem, and recovering from being stuck.
The developmental worry, then, is about substitution during sensitive years. If a tool routinely supplies focus, patience, or answers that a child would otherwise have to generate, the child may get less of the practice that builds these functions. It is important to be honest that direct evidence for this specific claim about generative AI in children is not yet available. It is a reasonable extrapolation from developmental science, not a proven fact. Professional bodies that address children and media, such as the American Academy of Pediatrics, have generally moved away from single blanket screen-time limits toward attention to content, context, and the displacement of sleep, physical activity, and face-to-face interaction. That framing applies well to AI. What a young person gives up in order to use a tool often matters more than the minutes of use themselves. AI also offers real developmental benefits, including patient tutoring, translation, and support for children with disabilities, and these should be weighed alongside the risks rather than ignored.
Mental Health and Affective Dimensions
The emotional effects of AI are varied and should not be reduced to a single story. On one side, conversational AI is being used for company, for low-pressure practice of difficult conversations, and as a support tool within some mental-health programs, and early results for structured therapeutic applications are promising in places. For a person who is isolated or anxious, a non-judgmental responder available at any hour can bring genuine relief.
On the other side there are real concerns. AI companions can foster dependence and may substitute for the harder, more nourishing work of human relationships. Systems that always agree and always soothe do not offer the friction that helps a person grow, and they can validate distorted thinking rather than gently challenge it. There are documented cases of chatbots giving unsafe responses to people in crisis, which is why unsupervised use as a substitute for care in a genuine emergency is not advisable. Separately, the wider anxiety about being replaced or outperformed by AI is itself a source of stress for many workers and students. The evidence base here is early and the individual differences are large. What helps one person may harm another, so blanket claims in either direction are premature.
Practical Guidance
The following suggestions follow from established cognitive science rather than from fear of the technology. They are meant to be calm and workable.
For families. Decide together which skills and knowledge are worth owning, and protect regular unaided practice of them, especially reading, writing, and arithmetic for children. Treat the tool as a helper to think with, not a source of finished answers to copy. Guard sleep, physical activity, and unstructured time with other people, since these are what tool use most often displaces. Model the behavior you want, because children learn from what adults actually do.
For schools. Design assignments so that desirable difficulties survive. Use retrieval practice, spacing, and in-class work where a student must produce and defend an answer without help, which also reveals real understanding. AI can serve well as a tutor that explains, questions, and gives feedback, and less well as a ghostwriter. Teach students to check AI output against sources, to notice confident-sounding error, and to reflect on the difference between having an answer and understanding it. Make the tool's role in any given task explicit rather than left to guesswork.
For clinicians. Ask about AI and technology use as part of a broader picture of sleep, mood, attention, and relationships, without assuming harm. Watch for over-reliance that displaces human connection or that feeds avoidance and rumination. Be clear with patients that general-purpose chatbots are not a substitute for crisis care, and know the local emergency resources. Keep the human relationship central, since it is the part of care that no tool replaces.
For everyone. Notice when a tool is doing thinking you wanted to keep for yourself, and choose deliberately whether to hand it over. Check important machine output, particularly when you are tired or rushed, the conditions under which over-trust is strongest. Keep some questions that you answer from your own head.
Conclusion
The honest summary is that AI is neither poison nor tonic for the mind. The best-supported finding, cognitive offloading, is old news in a new form: minds have always distributed memory and effort onto tools and other people, and the gains and costs of doing so depend on what is offloaded and what is kept. The newer worries about generative AI, weakened metacognition, deskilling, blunted development in the young, and mixed effects on mood, are plausible and grounded in solid older research, but the direct evidence about these specific tools is still early, thin, and debated. That uncertainty is a reason for careful attention, not for panic. The practical path is steady: use these tools to extend thinking rather than to replace it, protect the effortful practice that builds durable knowledge and self-control, and keep human judgment and human relationships at the center. The tools will keep changing. The principles of how a mind learns, remembers, and pays attention change far more slowly, and they remain the surest guide.
References and Further Reading
Sparrow, B., Liu, J., and Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science.
Wegner, D. M. (1987). Transactive memory: A contemporary analysis of the group mind. In Theories of Group Behavior.
Carr, N. (2010). The Shallows: What the Internet Is Doing to Our Brains.
Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science. See also Sweller, Ayres, and Kalyuga, Cognitive Load Theory.
Bjork, R. A., and Bjork, E. L. Work on desirable difficulties, retrieval practice, and spacing. See Bjork and Bjork (2011), Making things hard on yourself, but in a good way.
Roediger, H. L., and Karpicke, J. D. (2006). Research on the testing effect and retrieval practice.
Mosier, K. L., and Skitka, L. J. Research on automation bias and human use of automation.
Fisher, M., Goddu, M. K., and Keil, F. C. (2015). Searching for explanations: How the internet inflates estimates of internal knowledge. Journal of Experimental Psychology: General.
Diamond, A. (2013). Executive functions. Annual Review of Psychology.
American Academy of Pediatrics. Policy statements and guidance on children, adolescents, and media.
American Psychological Association. Health advisories and guidance on technology, social media, and youth wellbeing.