The speedup illusion is the mismatch where people expect AI assistance to make a cognitive task faster even when measured completion time does not improve. Yu, Cheng, Jabbar, Sucholutsky, Collins, Jurafsky, and Hawkins identify it in a preregistered behavioral study with 1,237 U.S. participants: predicted AI-assisted completion times were substantially lower than actual AI-assisted times, while predictions for independent completion were well calibrated.source: yu-cognitive-offloading-speedup-illusion-2026.md
The study separates prediction from performance. One sample predicted how long tasks would take independently, with AI, or with another human; another sample actually completed tasks independently or with embedded GPT-4o assistance. Participants expected AI to reduce completion time by about 68.5 seconds, much more than they expected another human helper to do. In reality, AI assistance did not reliably reduce completion times overall, though it helped on some difficult tasks and a few specific tasks such as finding related words, summarizing a long article, and editing long text.source: yu-cognitive-offloading-speedup-illusion-2026.md
The important nuance is that time and effort diverged. AI assistance reduced subjective mental effort across tasks, measured with NASA-TLX, even when it did not reduce elapsed time. Participants may therefore experience AI use as easier and infer that it must also be faster. This distinction matters for productivity claims: a tool can reduce friction, fatigue, or unpleasantness without increasing throughput.source: yu-cognitive-offloading-speedup-illusion-2026.md
The paper connects directly to cognitive-surrender. Poor calibration can make users offload too often, especially when they dislike cognitively demanding tasks, and the authors warn that overreliance may create a feedback loop: more offloading, less independent practice, weaker future capability, and even stronger dependence on AI.source: yu-cognitive-offloading-speedup-illusion-2026.md
For ai-assisted-software-development, the implication is not “AI never saves time.” The study focuses on simple tasks under five minutes, so it does not settle complex programming workflows. But it does warn against treating felt ease as proof of speed, especially for small tasks where prompt-writing, reading, verification, and post-processing can consume the saved time.
Related pages: cognitive-surrender, ai-assisted-software-development, harness-engineering.