Originally published on DropThe.org.
Photo by Nataliya Vaitkevich on Pexels
Developers using AI tools took 19% longer to finish tasks. They believed they were 20% faster. That 39-point gap between perception and reality is the story of AI in the workplace right now.
Three studies published in the last eight months examined what happens when workers actually use AI tools. Not demos. Not marketing decks. Real people in real jobs. The findings line up. AI does not save time. It shifts time into new obligations that feel productive but measure as burnout.
Three Studies, One Direction
| Study | Source | Sample | Finding |
|---|---|---|---|
| AI Work Intensification | UC Berkeley / Harvard Business Review (Feb 2026) | 200-person tech company, 8 months, 40+ interviews | Burnout, expanded workloads, blurred work-life boundaries |
| Developer Productivity Trial | METR (Jul 2025) | 16 experienced open-source devs, 246 real issues | 19% slower with AI. Developers believed they were 20% faster. |
| LLM Labor Market Impact | NBER (2025) | Thousands of workplaces, all occupations | 3% time savings. No impact on earnings or hours. |
The UC Berkeley researchers spent eight months inside a company where nobody was forced to use AI. No new targets were set. Employees adopted the tools voluntarily. Then to-do lists expanded. Work bled into lunch breaks. Evenings disappeared.
The researchers described the mechanism: when AI makes tasks faster, organizations do not reduce hours. They increase expectations. The freed time becomes new tasks. Workers absorb the load until they break.
The 39-Point Perception Gap
The METR developer study controlled for this precisely. Sixteen experienced developers completed 246 real issues from their own open-source repositories. Each issue was randomly assigned: AI tools allowed, or no AI. The developers used Cursor Pro with Claude 3.5 Sonnet. Frontier models on familiar code.
Results: developers using AI took 19% longer. When asked to estimate their speed, they reported being 20% faster. That gap held even after the study ended and developers could see their own timing data.
This matters because every corporate AI adoption decision relies on user self-reports. “Our team loves Copilot.” “ChatGPT saves me hours.” If users systematically misjudge by 39 percentage points, the feedback loop driving AI procurement is broken.
What the Broad Data Shows
This is an excerpt. Read the full analysis with charts and data on DropThe.org
About DropThe
DropThe is a data platform tracking 1.83 million entities across movies, games, companies, people, and crypto — connected by 2.18 million knowledge graph links. We don't guess. We count.
Top comments (0)