Cognitive offloading and the speedup illusion in human-AI interaction

Synthesis: This preregistered large-scale study (N = 1,237) investigates whether people are well-calibrated in estimating the time savings from AI assistance on simple cognitive tasks. The key finding is a speedup illusion: participants accurately predict how long they'll take independently but significantly underestimate how long they'll take with AI assistance — despite actual completion times being equivalent between independent and AI-assisted conditions. Notably, this bias is AI-specific; the same miscalibration does not appear when participants imagine help from another human. A critical dissociation emerges between time and effort: participants report lower subjective effort with AI even when completion times are identical, suggesting subjective experience drives AI adoption more than objective efficiency.

The findings complement the companion paper on the The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks (same authors, arXiv:2605.22687) which found people underestimate their rate of AI usage and overestimate benefits on simple tasks. Together, these studies reveal a dual miscalibration that could entrench inefficient Over-Reliance patterns: users choose AI believing it saves time and effort, when for simple tasks it may do neither. This has direct implications for educational settings — students using AI for simple cognitive work (arithmetic, basic writing, spell-checking) may cognitively offload without actual efficiency gains, reducing learning transfer through diminished deliberate practice. Presented at CogSci 2026, the work bridges cognitive science and Metacognition research with practical design implications for AI tools in learning environments.

What this means for practice

• Learners. Time yourself on a short task both with and without AI before assuming the tool saves time: participants predicted AI would cut completion times by 68.5 seconds (β = 68.5, SE = 3.37, p < 0.001), but their actual AI-assisted times were significantly longer than their predictions while independent estimates were accurate.
• Judge a tool on two separate questions — "was it faster?" and "did it feel easier?" — because AI assistance reduced NASA-TLX effort by 0.61 points on a 7-point scale while speeding up only three of the 24 tasks.
• Budget time for reading and checking model output: on the logic problem, post-response processing took 113.9 seconds longer than prompting, and AI-assisted completion of that task was slower.
• Write your own prompts rather than pasting the assignment text; 18.5% of prompts were copied directly from the instructions, and copying did not significantly reduce completion time.
• Compare AI against a competent human peer before delegating: participants expected AI to be 51.4 seconds faster than assistance from another participant.

Limitations

• The study used a between-subjects design with N = 1,237 (401 in the prediction sample and 836 in the completion sample), so calibration was never measured within an individual.
• All 24 tasks were completable in under 5 minutes; the authors identify the complexity level at which AI genuinely saves time and effort as an open question.
• The prediction and completion samples were not completely disjoint — some participants completed both on Prolific.
• AI use was not standardized: participants used the model in very different ways, and the experiment did not control for participant motivation or incentives, with 6.3% of independent-condition and 4.0% of AI-condition responses excluded as low-quality.

Citation

Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins (2026). Cognitive offloading and the speedup illusion in human-AI interaction. Proceedings of the 48th Annual Meeting of the Cognitive Science Society (CogSci 2026).