A new study from arXiv (paper #2605.22687) drops a uncomfortable truth bomb on the AI productivity narrative: we're all using these tools way more than we think, and most of us are getting zero meaningful benefit from it. Researcher Sunny Yu conducted three pre-registered studies involving 2,691 participants to dig into whether people's reliance on AI for simple tasks is actually well-calibrated. The short answer? Absolutely not.

The Two Flavors of Self-Deception

Yu identified two distinct miscalibration patterns that are equally troubling. First, there's "self-estimate miscalibration"—participants systematically believed they were using AI less than they actually were. This isn't just a minor awareness gap; it's a fundamental disconnect between perception and reality. Second, and perhaps more damning for the efficiency-obsessed crowd, is what Yu calls "efficiency-gain illusions." Users consistently overestimated how much time and effort they were saving by delegating simple cognitive tasks like arithmetic, spell-checking, or answering basic questions to AI assistants.

The Feedback Loop Nobody's Talking About

Here's where it gets really interesting—and frankly, a little dystopian. Yu discovered a "session-level carryover effect" that suggests prior AI use makes users more likely to reach for AI again in subsequent tasks. This creates a feedback loop where initial reliance breeds further dependence, which then entrenches the miscalibration about time savings even deeper. Think of it as compounding interest on cognitive outsourcing, except your bank account is your ability to do basic math without help.

The Inefficiency Paradox

Perhaps most striking is the finding that people frequently chose to use AI even when doing so provided "no meaningful time or effort savings." Let that sink in. We're talking about tasks like checking spelling or answering simple factual questions—things most humans could handle faster by just... doing them. Yet participants kept reaching for AI anyway, while simultaneously believing they weren't over-relying on it.

Key Takeaways

  • Self-estimate miscalibration means most users have no idea how dependent they've become on AI tools
  • The efficiency gains from AI use are systematically overstated by users across multiple study conditions
  • Prior AI usage creates a compounding effect that increases future reliance—this isn't linear growth, it's exponential
  • Simple cognitive tasks (arithmetic, spell-check) show the most egregious examples of unnecessary AI delegation

The Bottom Line

This research should be required reading for every "AI will make us more productive" think piece out there. We're not just overusing these tools—we've lost the ability to accurately perceive that we're overusing them. That's not a productivity win; that's a dependency trap wearing the costume of efficiency.