The Misery Index for AI Dependency Just Got a New Metric

Researchers Sunny Yu and colleagues have dropped what might be the most uncomfortable paper of the year for anyone who's ever told themselves they only use AI "sparingly." A trio of pre-registered studies encompassing 2,691 participants reveals a troubling pattern: people are choosing to use AI tools even when those tools provide zero meaningful time or effort savings. Worse still, they're largely unaware they're doing it. The paper, posted to arXiv in late May 2026 (arXiv:2605.22687), cuts through the hype with rigorous experimental methodology—and delivers findings that should make every developer and product team pause.

When AI Takes Longer Than Doing It Yourself

The core mechanism at play here isn't laziness—it's miscalibration operating on two distinct levels. First, there's self-estimate miscalibration: participants systematically believed they were using AI less frequently than they actually were. Second—and this is where it gets interesting from a developer perspective—there are efficiency-gain illusions. Users consistently overestimated how much time and cognitive effort they'd save by delegating tasks to AI, even for trivial operations like arithmetic, spell-checking, and answering simple factual questions that most humans could handle faster without switching context.

The Carryover Problem: Once You Start, You Can't Stop

Perhaps the most alarming finding is what the researchers call a "session-level carryover effect." When participants used AI earlier in a study session, they became more likely to reach for AI again for subsequent tasks—regardless of whether those tasks warranted AI assistance. This creates an overreliance feedback loop: prior adoption leads to increased future adoption, which further entrenches the user's miscalibration about actual time savings. In other words, once you start using AI habitually, you're not just stuck with it—you're convinced it's saving you more time than it actually is.

Why Developers Should Care About These Findings

From a software architecture and developer tooling perspective, this research has uncomfortable implications. If users can't accurately assess whether AI assistance is helping or hurting their efficiency, how do we design better tools? The answer isn't to push harder on adoption—it's to build better calibration mechanisms into AI-powered IDEs, code editors, and productivity suites. We need systems that help users understand when they're genuinely delegating value versus when they're just adding latency through unnecessary round-trips.

Key Takeaways

  • Three pre-registered studies (N=2,691) found people use AI for simple tasks even when it provides no efficiency benefit
  • Self-estimate miscalibration: users believe they rely on AI less than they actually do
  • Efficiency-gain illusions lead users to overestimate time savings from AI assistance across task types
  • Session-level carryover effects create feedback loops that entrench overreliance and calibration errors

The Bottom Line

This research is a wake-up call for the entire AI-assisted development ecosystem. We're building tools on the assumption that users can judge when AI helps versus hurts—but the data says otherwise. If we want AI to genuinely augment human developers rather than create a new class of cognitive outsourcing zombies, we need to start measuring and addressing this miscalibration head-on. The efficiency gains might be real; our ability to perceive them clearly is clearly not.