Cognitive offloading

Cognitive offloading is the use of physical actions, tools, or other external resources to alter the information-processing requirements of a task, so as to reduce the demand on internal cognition.^[1:1] Tilting the head to read a rotated label, jotting an appointment in a diary, programming a phone alarm, or asking a search engine for a half-remembered fact are all examples: in each case, the person delegates some part of a mental operation to the world. The term was consolidated in cognitive psychology by Evan Risko and Sam Gilbert in 2016, building on a longer tradition in philosophy and cognitive science that treated such use of the environment as part of cognition rather than merely a context for it.^[1][2]

Empirical research has been concerned with two main questions: when people decide to offload rather than rely on internal processes, and how offloading reshapes the cognitive abilities it supports.^[1] On the first, evidence points to metacognition — the person's belief about their own ability to perform the task — as a central trigger, alongside a more general preference to avoid cognitive effort.^[3] On the second, results are mixed: offloading reliably boosts performance on the offloaded task, can free internal resources for new information, but can also weaken encoding and recall of the very material that was offloaded.^[4][5][6]

The phenomenon has become a focus of debate again with the spread of generative artificial intelligence, where individuals can delegate not only memory and retrieval but also synthesis, drafting, and decision-making to a system. Recent studies link frequent use of such tools to reduced self-reported critical thinking and reduced cognitive engagement, though the field is young and methodologically contested.^{[7][8][9][10]}

Concept and origins

Long before the term cognitive offloading was in common use, philosophers and cognitive scientists were arguing that thinking does not stop at the boundary of the skull. In their 1998 paper The Extended Mind, the philosophers Andy Clark and David Chalmers proposed an "active externalism" in which an external resource, when reliably coupled to a person and used in the same way an internal process would be, counts as part of that person's cognitive system.^[2:1] Their thought experiment compared Inga, who recalls from biological memory that a museum is on a particular street, with Otto, who has Alzheimer's disease and instead consults a notebook he always carries; Clark and Chalmers argued that, in the relevant respects, the notebook plays the same role for Otto that memory plays for Inga, and so its contents qualify as his beliefs.^[2:2]

A parallel line of work in cognitive psychology examined how groups and couples develop transactive memory systems, in which what each member knows is supplemented by what they know other members know — and how this same structure extends to external aids.^[4] By the 2010s, experimental psychology had accumulated many separate findings — on gestures, head movements, calculators, global positioning system devices, photographs, search engines, and external reminders — without a single unifying label.^[1] Risko and Gilbert's 2016 Trends in Cognitive Sciences review proposed cognitive offloading as that label, defining it as the use of physical action to reduce the cognitive demand of a task, and offered a metacognitive framework to organise the empirical literature.^[1]

Forms of cognitive offloading

Risko and Gilbert distinguish two broad classes of offloading. The first uses the body itself: tilting the head to align a stimulus with an internal representation, pointing or counting on fingers to keep track of items, or gesturing while thinking through a problem all reduce the load carried by internal processes.^[1] The second uses objects and information structures in the world: lists, calendars, sticky notes, calculators, GPS devices, smartphones, search engines, and increasingly artificial intelligence assistants.^[1][3] In both cases the person is doing some of the work through manipulation of the environment, rather than wholly inside the head.

Different forms target different cognitive functions. Notebooks, photographs, and digital files most clearly target memory: information that would otherwise have to be encoded internally is stored in an external medium for later retrieval.^[5][4] Calendars, alarms, and strategically placed objects target prospective memory — the memory for delayed intentions — by creating an external cue that will trigger the intended action at the right moment.^[3] Calculators, slide rules, and spreadsheets target computation. Generative AI tools, recommender systems, and decision-support software target higher-level cognitive activities such as drafting, synthesis, and evaluation.^[8] A given task often blends several of these: writing an essay with a word processor and a search engine combines spelling support, external storage, retrieval of facts, and, increasingly, generative assistance.^[8]

Mechanisms and triggers

A central question for the empirical literature has been why people choose to offload in some situations and not in others. Risko and Gilbert proposed that the decision is driven largely by metacognition: people are more likely to offload when they believe their own internal abilities are inadequate for the task.^[1] Studies of intention offloading — the use of external reminders for delayed intentions — have provided some of the most direct evidence for this account. In experimental paradigms developed by Sam Gilbert and colleagues, participants can choose between relying on internal memory or setting external reminders. They are more likely to set reminders when memory load is high or when ongoing tasks are interrupted, and their reminder-setting tracks their subjective confidence in their memory even when that confidence is uncorrelated with their actual accuracy.^[3]

Metacognitive beliefs do not, however, explain the whole pattern. When the costs and benefits of internal memory versus external reminders are quantified, participants tend to set reminders more often than would be optimal — a bias that persists even when participants are made overconfident through positive feedback and easy practice trials.^[3:1] This systematic bias is attributed in part to a preference to avoid cognitive effort, independent of beliefs about ability.^[3] Offloading decisions are also stable as an individual difference over time and shift with development: young children often fail to use reminders even when they know their memory is limited, while older adults sometimes underuse reminders because they overestimate their memory.^[3]

Metacognitive interventions can shift offloading without changing underlying memory ability. In one experiment, the difficulty of practice trials and the wording of feedback were manipulated independently of objective performance; participants who received easy practice and positive feedback became more confident in their memory and set fewer external reminders, with the change in reminder-setting mediated by the change in confidence.^[3:2] This result suggests that offloading behaviour can be influenced by interventions that target subjective metacognition, with implications for tool design and education.^[3]

A further trigger is the perceived reliability of the external resource. People offload more readily to tools they trust, including, by extension, to artificial intelligence systems whose outputs they treat as authoritative. In a survey of 319 knowledge workers, higher confidence in a generative AI tool was associated with reduced enaction of critical thinking when using it, while higher confidence in one's own ability to do the task was associated with more.^[8] This pattern is consistent with the broader metacognitive account: trust in the tool effectively substitutes for trust in oneself in the offloading decision.^[8][1]

Cognitive consequences

Offloading reliably improves performance on the task it supports: external reminders dramatically increase the rate at which delayed intentions are fulfilled, and offloaded calculations and lookups are typically faster and more accurate than internal ones.^[3][1] What is less obvious, and has been the focus of much research, is what offloading does to internal cognition itself — both during the offloading episode and over the longer term.

One robust finding is that information expected to remain externally available is encoded less deeply than information expected to be lost. In a series of four experiments published in Science in 2011, Betsy Sparrow, Jenny Liu, and Daniel Wegner examined this directly. A modified Stroop task showed that participants were slower to name the colour of computer-related words such as Google and Yahoo after failing to answer difficult trivia questions, suggesting that people are primed to think of computers when they encounter a gap in their knowledge.^[4] In a second experiment, participants who typed trivia statements into a computer believing that the statements would be saved later recalled fewer of the statements themselves than participants who believed the statements would be erased, and an explicit instruction to remember made no significant difference.^[4:1] A fourth experiment showed that participants were better at recalling which folder a statement had been saved in than the content of the statement itself.^[4:2] Sparrow and colleagues described this as evidence that the internet had become a primary form of external or transactive memory, with people increasingly remembering "where" rather than "what".^[4] The phenomenon is sometimes called the Google effect or digital amnesia.^[4]

A related effect appears with photography. In two studies of museum visitors published in 2014, Linda Henkel found that participants who photographed objects on a guided tour later remembered fewer objects, and fewer details about those objects, than participants who simply observed them.^[5] Henkel labelled this the photo-taking-impairment effect and interpreted it as cognitive offloading: relying on the camera to record the scene appeared to reduce the attention given to the scene itself.^[5] The effect was attenuated when participants zoomed in on specific parts of an object, suggesting that the additional attention required by selective photographing can offset the impairment.^[5:1]

Not all consequences of offloading are costs. Benjamin Storm and Sean Stone found in 2015 that saving a list of words to a computer file before studying a second list improved memory for the second list, an effect they called saving-enhanced memory.^[6:1] On their account, saving acts as a form of offloading that frees cognitive resources from the maintenance of the first list, allowing those resources to be allocated to the new information.^[6] Offloaded intentions show a similar profile: setting an external reminder reduces the need to keep the intention active in internal memory and can free attention for the ongoing task.^[3]

The literature thus does not support a simple narrative of decline. Where information is reliably available externally, people shift internal resources away from it; where this frees capacity for other tasks, the net effect can be beneficial. The costs become salient when the external resource is unavailable, when the offloaded information itself was important to encode, or when offloading substitutes for cognitive processes — such as deliberation or evaluation — that the person had reason to perform.^[1][3]

Cognitive offloading and artificial intelligence

The arrival of widely accessible generative AI tools in the early 2020s extended the scope of cognitive offloading from memory and retrieval to drafting, summarisation, synthesis, and evaluation. By 2025, this had become a distinct strand of the research literature, focused on whether such tools degrade critical thinking and what factors moderate any such effect.^[7][8]

In a mixed-methods study of 666 participants in the United Kingdom published in Societies in January 2025, Michael Gerlich reported a negative correlation between frequent AI tool usage and self-reported and assessed critical thinking, with cognitive offloading mediating the relationship.^[7] Younger participants showed higher AI use and lower critical thinking scores than older participants, and higher educational attainment was associated with higher critical thinking regardless of AI use.^[7:1] A correction was published in September 2025 noting that one table had been duplicated in the original article; the author stated that the substantive conclusions were unaffected.^[11:1]

A complementary picture emerges from a survey of 319 knowledge workers conducted by researchers at Carnegie Mellon University and Microsoft Research and presented at the CHI 2025 conference. Workers reported less critical-thinking effort on AI-assisted tasks than on the same tasks performed without AI, but the shift was not uniform: those with higher confidence in their own ability or in their ability to evaluate AI outputs reported more critical thinking, while those with higher confidence in the AI itself reported less.^[8:1] The authors describe a qualitative shift in cognitive effort from material production to verification, integration, and oversight of AI output, which they argue is not necessarily a reduction of critical thinking but a redirection of it.^[8:2]

A more dramatic claim was made by Nataliya Kosmyna and colleagues at the MIT Media Lab in a 2025 preprint, which used electroencephalography to compare 54 participants writing essays with a large language model, with a search engine, or without tools across three sessions, and reassigned 18 of them in a fourth session.^[9] The authors reported that participants in the LLM condition showed the weakest brain connectivity and the poorest recall of their own essays, and described an accumulation of "cognitive debt" with continued LLM use.^[9:1] The work attracted wide attention but also methodological criticism: a December 2025 commentary by Miloš Stanković and colleagues raised concerns about sample size, reproducibility of the EEG analyses, and inconsistencies in the reporting of results, and suggested that some findings should be interpreted more conservatively pending peer review.^[10:1] As of early 2026 the Kosmyna preprint had not been published in a peer-reviewed venue.^[9][10]

Analysis: cognitive offloading and human agency

The empirical record on cognitive offloading does not licence either of two simple conclusions about agency that are often drawn in public discussion. It does not show that offloading is generally harmful to the durable capacity to understand a situation, form aims, and act on them: external reminders are highly effective, saving information can improve memory for new information, and people often offload roughly in line with their own task demands.^[3][6] Equally, it does not show that offloading is neutral: information expected to remain externally available is encoded less deeply, photographed scenes are remembered less accurately, and frequent reliance on AI tools is associated, at least in self-report, with reduced cognitive engagement on the tasks for which they are used.^[4][5][7][8]

What the literature does suggest, and what bears most directly on agency, is that the value of offloading depends on whether the offloaded process is one the person had reason to perform internally — and on whether the external resource that replaces it remains under the person's control. An external reminder that the person sets, inspects, and can override preserves the agentic structure of forming and acting on an intention.^[3] An AI system whose outputs the person accepts on trust, without independently evaluating them, substitutes for the evaluative process rather than supporting it; the survey evidence is consistent with this asymmetry, in that confidence in the tool is associated with less critical thinking while confidence in oneself is associated with more.^[8]

This points to a practical implication that is widely endorsed across the surveyed sources, though stated in varying terms: tools that scaffold the user's own reasoning — by surfacing assumptions, supporting verification, and leaving final judgement to the user — appear better aligned with sustained cognitive capacity than tools that present finished answers for acceptance.^[8][7] The same surveys note that the responsibility is not solely on tool design: educational practices that train the explicit evaluation of AI outputs, and individual habits of checking, appear to moderate the effects observed.^[7][8]