Thinking with AI or Thinking Less? A Narrative Review of Generative AI Dependence, Cognitive Offloading, and Human Cognitive Functioning

Abstract

The new generation of digital assistance programs, including generative artificial intelligence (GenAI), is increasingly shifting from information retrieval to the generation of explanations, arguments, code, abstracts, and solutions. The extent to which GenAI support of human cognition can extend human capacity versus supplant critical cognitive processes underlying knowledge and skill formation has been a longstanding psychological question. This narrative review focuses on foundational research on cognitive offloading and on peer-reviewed studies of GenAI published through September 05, 2026. A structured narrative search for experimental, meta-analytic, and mixed-methods studies on memory, learning, critical thinking, creativity, perceived effort, trust, and dependence on GenAI was complemented by a search of crosssectional and qualitative studies that provided early characterization of emerging constructs of dependence on GenAI. Evidence does not support a simple positive or negative effect of GenAI on human cognition. GenAI-supported offloading of cognitive functions such as memory retrieval, processing, and learning can reduce the required cognitive capacity and support high performance across a variety of tasks, including memory-based and GenAI-supported academic tasks. However, the same cognitive offloading can, in some cases, reduce the quality of subsequent encoding, and recent studies on GenAI show a consistent performance-learning dissociation. While support that enables high performance when support is available can carry over to later unsupported performance, the gain is typically small, nonexistent, or even negative. Studies of support provided under instructional control indicate that retaining learner-generated nodes or restricting direct-answer substitutes for generators can support later independent performance better than unlimited support. It remains to be determined what the independent contribution of explanation, verification, revision, and fading support is. Nevertheless, findings from emerging dependence research correlate with indicators of academic stress, performance expectations, anxiety, and self-efficacy. However, currently available evidence does not support claims of irreversible cognitive decline. A key finding of this review is that whether GenAI is used at all is less critical than which functions are offloaded, which are maintained, and the extent to which users remain epistemically responsible for the resulting output. Implications are developed for education, knowledge work, tool design, and longitudinal research.