Thinking with AI, reasoning without it: Cognitive offloading to generative AI weakens effort regulation

Synthesis: Ratajczyk and colleagues ask what most genAI-in-education research dodges: not whether ChatGPT helps while the window is open, but what is left of a learner's reasoning once it closes. Their preregistered experiment (final N = 193) ran a 15-entity relational network through chunk-based reasoning, a harder network-level task on the same material, and a memory test, with ChatGPT available only in Stage 1 and incentives crossed against it. Access raised Stage 1 accuracy (77.85% vs 72.72%) and lowered perceived difficulty, yet the unaided Stage 2 showed no aggregate gap; item-level modeling, however, linked prior genAI use to roughly 32% lower odds of a correct answer. The starker result was motivational: incentives lengthened time on difficult Stage 2 items for those who had worked alone but not for tool users — cognitive offloading weakening effort regulation, not reasoning capacity.

Key Findings

1. A five-condition preregistered lab experiment, not a classroom study. 2 (genAI access) × 2 (performance incentive) between subjects plus a control condition on an unrelated network; final sample 193, mean age 25.1 years.
2. GenAI helped immediately, and the incentives did not. Stage 1 accuracy was 77.85% with ChatGPT against 72.72% without it, and perceived difficulty fell from 4.71 to 4.17. Incentives added time but no accuracy.
3. Unassisted reasoning slipped, but only at the item level. Stage 2 accuracy was 61.6% with no aggregate effect of prior genAI access, yet the exploratory item-level model found roughly 32% lower odds of a correct answer after prior tool use (OR = 0.68).
4. Effort regulation is where the difference concentrated, and only on hard problems. Stage 2 time rose with incentives (874.6 s vs 751.5 s) and interacted with prior genAI access: without the tool, incentives stretched completion from 707.4 s to 941.2 s; with prior genAI there was no adjustment. The three-way interaction held only on hard items (126.4 s vs 84.7 s).
5. Memory showed nothing. Stage 3 accuracy was 57.5%; the one difference was meta-level, with incentivized tool users rating the memory test harder.
6. Offloading came in different shapes; neither working memory nor incentives predicted it. Of 78 codable conversations, 31 barely used the tool, 28 mostly copied questions and 16 used it exploratorily.

The three-stage task and what each stage isolates

The stimulus was a 15-entity network — creatures in the main version, celestial bodies in the control — described by 82 logical propositions and deliberately too large for working memory. Stage 1 (chunk-based reasoning) presented chunks of four to eight statements followed by three-option inference questions (yes, no, impossible to determine). Chunking follows cognitive load theory's element interactivity logic, aiming at a structured representation rather than bare answers. Stage 2 (network-level reasoning) supplied 20 information points containing 48 of those propositions and asked 13 integration questions, testing whether relational structure could be reconstructed and applied unaided. Stage 3 was a 10-item memory test with no access to the material.

The control condition is the clever part: a comparable Stage 1 on a different network, with the same demands but no transferable relational knowledge, separating offloading-induced harm to knowledge construction from fatigue or lost practice. Its Stage 2 accuracy was numerically below the no-genAI baseline, but it took significantly longer to finish: those who lacked prior network knowledge compensated with time, a benchmark that makes the genAI group's failure to add time look pointed.

Assisted performance, then unassisted reasoning: what the results support

Read against the preregistered hypotheses, the headline is negative: H1a and H1b, that prior genAI access would improve or impair Stage 2 accuracy, were unsupported at the aggregate level, as were H3, H6 and H7. What survived was an exploratory item-level effect in the direction of impairment and the preregistered incentive effect on time, which depended on prior genAI access. At the item level, prior tool users had roughly a third lower odds of answering an individual Stage 2 question correctly within the no-incentive condition, and their time on difficult questions did not rise when incentives made success valuable.

Two dissociations support reading that as regulation rather than ability: Stage 2 difficulty was perceived identically across conditions, so tool users were not overwhelmed, and memory was untouched, suggesting individual facts survived where their relational use did not. The control comparison adds a third: participants who genuinely lacked prior knowledge spent more time, and it did not rescue their accuracy.

Effort was measured three ways: item-level completion time on Stages 1 and 2; the Effort/Importance subscale of the Intrinsic Motivation Inventory, back-translated into Polish; and a seven-point perceived-difficulty rating per stage, with the AIAS-4 scale and self-reported usage. The incentives — a bonus for the top-performing third, a European comparison, individual scores on request — bought time, not accuracy, which is why time reads as effort.

The mechanism: effort withdrawal, calibration, or reliance habit

The paper's preferred reading is a change in effort regulation, with two mechanisms side by side: motivation intensity theory's potential motivation, in which prior genAI success lowered how much effort the task seemed to warrant, so hard questions hit that ceiling sooner; and expected value of control, in which Stage 1 taught an inaccurate contingency between effort and success, leaving participants unable to recalibrate once the tool vanished. The self-report pattern leans slightly toward the first, but no simple effect survived correction, so the authors decline to choose. They rule out one shallow alternative: participants did not forget the incentive instructions, since incentives lengthened Stage 1 time even with the tool present.

This is not the fatigue story (the control condition carried similar load with a different result) or knowledge loss (memory was unaffected). It is closest to Fan et al.'s metacognitive laziness, Liu et al.'s metacognitive decay, and Wu et al.'s motivation drop when genAI is withdrawn. Its distinctive contribution: offloading bites hardest on desirable difficulty, collapsing effort exactly where it is most productive. The usage patterns complicate any simple "offloading is bad" reading: sixteen participants used the tool exploratorily, spending more time on Stage 1 than the question-copiers. As the seung-basham and misiejuk pages show, access is not the same as offloading, and interaction style may matter more than the mere presence of a chatbot.

What this means for practice

• Educators. Test the unassisted residue: assisted accuracy is not evidence that structure was built, since prior tool use came with roughly 32% lower odds of a correct unaided answer.
• Educators. Focus on the hardest items: incentives lengthened time on difficult Stage 2 questions only for students who had worked without the tool (126.4 s vs 84.7 s), so plan the tool's withdrawal and Scaffolding rather than rewards after it.
• Educators. Keep generative AI out of the step the task teaches: the harm case was the tool doing the integration that was the learning objective, while exploratory use looked benign or better.
• Educators. Do not read a flat class average as safety: the preregistered aggregate test was null while item-level modeling found impairment concentrated on difficult questions.
• Researchers. Measure access, extent and type of use separately: nearly 40% of participants with tool access barely used ChatGPT, diluting any condition-level comparison.

Limitations

• One laboratory session with 193 adults in Poznań (105 with secondary and 88 with higher education), where Stage 2 followed Stage 1 immediately on the same network and the memory test spanned both stages.
• The impairment finding is exploratory and level-dependent: aggregate Stage 2 accuracy showed no prior-use effect and only the item-level model produced the association, so it needs replication.
• The manipulation was access rather than measured offloading, and almost 40% of the genAI condition used ChatGPT rarely or never, so condition-level differences understate any real effect; whether unassisted retrieval practice erases the effort signature is unanswered.
• Question difficulty was estimated from this sample's own accuracy, incentives moved time but never accuracy, and completion time is only an indirect measure of effort; the design tested extrinsic motivation only, so nothing follows about interest, autonomy or intrinsic motivation.

Citation

Ratajczyk, D., Dymarska, A., Matłoka, A., Tomczyk, M., & Wiącek, M. (2026). Thinking with AI, reasoning without it: Cognitive offloading to generative AI weakens effort regulation. PsyArXiv (preprint, not peer reviewed).