Do heavy AI users actually encounter more hallucinations?

A survey of 1,038 US adults who use AI, fielded by Centiment for Rev between July 25–28, 2025 (margin of error ≈±2% at 95% confidence), finds self-reported experience with AI correlates with more friction, not less. "Heavy AI users are nearly 3x more likely to experience frequent hallucinations" than casual users; while "77% get an answer they like in under two minutes" overall, that share drops to 50% among people who use AI "over six hours a week." Those power users are "10x more likely than casual users" to spend more than 11 minutes revising before they're satisfied (21% vs. 2%), and when a session runs past 20 minutes, "almost 9 out of 10 (88%) say they 'very often' have to revise for hallucinations." Daily users are also "14x more likely than casual users to double-check AI's work."

Rev's own reading is explicitly not that the tools get worse with heavier use: "This isn't necessarily tied to inefficiency — complexity is the likely culprit." Heavy users "are trying to do much harder things, like getting the AI to perform an in-depth analysis or perfectly mimic a specific writing style," while light users "stick to simple requests that lead to quick victories"; heavy users also "have higher standards." Asked why more use brings more reported flaws, Rev hedges between two rival explanations rather than choosing one: "experienced users are either asking tougher questions or have just gotten much better at noticing when the AI gets something wrong."

This runs parallel to Do AI coding tools actually speed up experienced developers? and Does AI assistance help less experienced workers most?: across three populations — developers, support agents, general AI users — the people with the most AI experience or skill are the ones for whom the tool's promised speed and reliability fail to show up, while novices and light users capture the easy wins. It also echoes Does chat delegation actually save time on task completion?'s split between effort and time savings, and complicates the more optimistic Do AI coding features actually speed up engineer productivity? by suggesting an aggregate time-saved figure can hide large variance by user type.

Every figure here is self-reported — felt hallucination frequency, felt session length, felt double-checking — none measured against ground truth or a clock, and the 1,038-respondent, US-only, unweighted sample doesn't generalize past US AI users in July 2025. It's also commissioned content on a vendor's blog: Rev, an AI transcription company, publishing survey results that rank competing chat tools (ChatGPT, Gemini, Perplexity), not neutral research. Correlation between hours of use and reported hallucination frequency is not evidence that AI tools degrade with use or that heavy users are less skilled — Rev's own account points the opposite way, toward harder tasks and sharper scrutiny. The defensible implication: self-reported hallucination and satisfaction-time rates are confounded by task difficulty and user calibration, so raw survey "hallucination frequency" figures shouldn't be read as product-quality scores without controlling for what respondents were actually attempting.