Your brain on AI
Move the delegation slider and switch verification off to see how the research on offloading, automation bias and the jagged frontier fits together. A conceptual diagram, not a measured neural process.
A flow, not a brain scan
The diagram above describes a workflow, not a neural process. Delegating a task to an AI system, letting it generate, verifying the result and integrating it into your work are four decisions you make; the dots are tasks, not thoughts. The slider changes how much of your work goes through that loop and the toggle removes the verification step, so you can see which outcomes the literature associates with each choice. The three indices are illustrative: their directions follow the studies cited below, their magnitudes are set for the diagram.
What this shows / what it does not show
Shows: the structure of human–AI collaboration and the direction of four effects the research describes: freed capacity from offloading, reduced checking under reliable automation, task-dependent gains and losses, and self-reported changes in critical-thinking effort.
Does not show: any measurement of brain activity, skill, or productivity for you or anyone else. The indices in this diagram are not built from measurements; they are a teaching device, and every panel says so. Brain activity during AI use has been measured in a few small studies: an MIT preprint recorded EEG while 54 people wrote essays with or without ChatGPT and reported weaker connectivity and poorer recall of their own text in the ChatGPT group C[7]. It is not peer-reviewed, the sample is small, and its authors warn against over-reading it; nothing here is calibrated to it.
Cognitive offloading: freed capacity, weaker trace
Risko and Gilbert define cognitive offloading as using physical action to alter the information-processing requirements of a task so as to reduce cognitive demand: writing a note, tilting your head to read rotated text, or asking a tool A[1]. The review's two robust findings pull in opposite directions. Offloading frees working memory and typically improves performance on the immediate task, and people offload more as tasks get harder or their confidence drops. But the information that is offloaded tends to be remembered less well afterwards, because the external store does the remembering. Most of this evidence comes from memory tasks, not from generative tools, so the AI case is an extrapolation.
Automation bias: reliable aids get checked less
Parasuraman and Manzey integrate decades of evidence from aviation, medicine and process control into an attentional account of complacency and automation bias: when an aid is usually right, operators allocate less attention to monitoring it and are more likely to follow it when it is wrong, producing both omission errors (missing a problem the aid did not flag) and commission errors (acting on a wrong recommendation) A[2]. A systematic review in health informatics documents the same pattern across clinical decision support and lists the mitigators: training, display design and accountability for outcomes A[3]. In an early experiment, participants with an automated aid made more errors when the aid was wrong than participants without one B[4]. The verification habit toggle is a direct translation of that literature: the habit is what the bias erodes.
The jagged frontier: better inside, worse outside
Dell'Acqua and colleagues ran a field experiment with several hundred consultants. On tasks within the model's capability frontier, consultants using AI completed more tasks, finished faster and produced work rated as higher quality; on a task designed to sit outside the frontier, those using AI were less likely to reach the correct answer B[5]. The authors' metaphor is a jagged line: capabilities that look similar from outside fall on different sides of it. Two cautions: this is a working paper from one firm with one generation of model, and the outside-frontier task was built to be misleading. The lesson that survives is that the person has to know where the frontier is, which is a verification problem.
Critical thinking: a self-report signal
Lee and colleagues surveyed 319 knowledge workers about 936 real examples of using generative AI at work. Participants who reported higher confidence in the AI reported less critical-thinking effort; those with higher confidence in their own ability reported more C[6]. The study is cross-sectional and based on self-report, so it describes how people perceive their effort, not how their cognition changed, and it cannot tell cause from effect. It belongs on the page because it is the most direct evidence so far about the question people actually ask; it is graded C for exactly those reasons.
Reading the three indices
- Time saved rises with the share you delegate and dips slightly when verification is on, because checking takes time. That trade is the honest cost of the habit.
- Error risk stays low with verification on and climbs steeply without it, following the automation-bias literature.
- Skill maintenance falls as you delegate more, less steeply when you verify, because checking keeps you engaged with the content. This is the most speculative of the three; the offloading evidence is about memory for information, not about professional skill.
What to take from it
Delegate deliberately, keep a verification step that is proportionate to the stakes, and keep doing a share of the work yourself where the skill matters to you. For the attention mechanics behind the "verify" step, see how focus works in the brain; for the cost of bouncing between the AI and your own work, see the cost of context switching.
Frequently asked questions
Does using AI make you less intelligent?
No study cited here shows that. The offloading literature shows weaker memory for information you hand to a tool, and a survey shows people report thinking less critically when they trust the AI more. Those are specific, mostly self-reported effects, not a change in intelligence.
What is automation bias?
The tendency to follow an automated aid's output and to monitor it less when it is usually correct, which produces errors when the aid is wrong. It has been documented in aviation, medicine and process control, and the mitigators are training, display design and accountability.
What does 'jagged frontier' mean?
A metaphor from a field experiment with consultants: AI capabilities do not form a smooth boundary. Tasks that look similar can fall inside the frontier, where AI helps, or outside it, where AI-assisted people were more often wrong. Knowing where the line runs is the user's job.
Are the time-saved and error-risk numbers measured?
No. They are illustrative indices from 0 to 100 whose direction follows the cited literature. The magnitudes were chosen to make the diagram readable and should not be quoted as findings.
Sources and evidence
- Cognitive Offloading — Risko EF, Gilbert SJ (2016), review A doi.org/10.1016/j.tics.2016.07.002 (retrieved 2026-11-09)
- Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman R, Manzey DH (2010), review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)
- Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard K, Roudsari A, Wyatt JC (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)
- Does automation bias decision-making? — Skitka LJ, Mosier KL, Burdick M (1999), lab experiment B doi.org/10.1006/ijhc.1999.0252 (retrieved 2026-11-09)
- Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua F, McFowland E, Mollick ER, et al. (2023), field experiment (working paper, not peer-reviewed) B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
- The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee H-P, Sarkar A, Tankelevitch L, et al. (2025), cross-sectional survey (CHI) C doi.org/10.1145/3706598.3713778 (retrieved 2026-11-09)
- Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task — Kosmyna N, Hauptmann E, Yuan YT, et al. (2025), preprint, randomized lab study with EEG (n = 54) C doi.org/10.48550/arXiv.2506.08872 (retrieved 2026-11-10)