You started using AI tools expecting to free up your afternoon. Instead, somehow you're working later than before. Sound familiar? UC Berkeley researchers spent eight months observing 200 employees at a U.S. tech company and found that AI didn't reduce work — it consistently intensified it.
What's Going On?
Back in 1865, economist William Jevons noticed something odd: as steam engines got more efficient, coal consumption didn't drop — it surged. More efficient engines made coal-powered production cheaper, so more industries adopted it. This became known as "Jevons Paradox." The exact same thing is happening with AI and knowledge work right now.
Aruna Ranganathan and Xingqi Maggie Ye from UC Berkeley published their findings in HBR. From April to December 2025, they observed a 200-person tech company. Nobody mandated AI use. Yet employees voluntarily worked faster, took on broader tasks, and extended their hours — all on their own initiative. As one engineer put it: "You thought maybe you could work less. But you don't work less. You just work the same amount or even more."
Hamilton Mann at California Management Review calls this the "AI Productivity Blind Spot." He identified five structural forces that neutralize or reverse expected AI productivity gains. The core insight: automation doesn't reduce cognitive load — it relocates it. Someone still has to review, filter, and judge every AI output.
The Numbers Tell the Story
Upwork 2024 survey: 96% of executives expected AI to boost productivity, but 77% of employees said it increased their workload. 71% experienced burnout, and 39% spent more time reviewing and correcting AI outputs.
What's Actually Changing?
The study identified three mechanisms of AI work intensification:
1. Task Expansion — PMs start writing code. Designers do engineering work. Researchers take on development tasks. AI fills knowledge gaps, so "let me just try this" experiments pile up and quietly expand everyone's job scope. Engineers end up spending more time reviewing and fixing AI-assisted code from colleagues who are "vibe-coding."
2. Blurred Boundaries — AI reduces the friction of starting any task to nearly zero. People send prompts during lunch, run AI during meetings, fire off "one last prompt" before leaving. Natural breaks in the workday quietly disappear. One worker said: "Once prompting during breaks became habitual, downtime no longer felt like recovery."
3. Multitasking Overload — Workers write code manually while AI generates alternatives in parallel, run multiple agents simultaneously, and revive long-deferred tasks. It feels like having a partner, but in reality it's constant attention-switching and output-checking that compounds cognitive load.
| What We Expected | What Actually Happened | |
|---|---|---|
| Workload | AI handles tasks, less for humans | Job scope inflates, net workload increases |
| Hours | Save time, leave early | Fill saved time with extra tasks, prompts after hours |
| Cognitive Load | Less repetitive work, clearer head | AI output review + multitasking = more cognitive fatigue |
| Breaks | More free time | Zero-friction prompting erodes rest periods |
| Long-term | Sustainable productivity gains | Burnout, weaker decisions, quality decline, higher turnover |
Here's the thing — this is Jevons Paradox playing out in real time. When AI lowers the cost of cognitive work, organizations don't use the savings for rest. They fill it with more work. Just like ATMs were supposed to eliminate bank tellers but actually led to more branches and more tellers, AI efficiency creates more total work, not less.
AI makes it easier to do more — but harder to stop.
— UC Berkeley Research Team
How to Start: Building an AI Practice
The researchers propose what they call an "AI Practice" — intentional norms that structure how AI is used, when to stop, and how work should expand.
- Design Intentional Pauses
Before major decisions, add structured checkpoints: one counterargument and one link to organizational goals. A 30-second pause prevents speed-driven judgment errors. - Create Sequencing Rules
Don't react to every AI output immediately. Batch non-urgent notifications, protect focus windows, and advance work in coherent phases rather than constant context-switching. - Run Regular Scope Audits
Ask quarterly: "Is this actually my job, or did AI quietly add it to my plate?" Check whether AI-expanded responsibilities are formally recognized and compensated. - Protect Human Grounding Time
Create deliberate spaces for tool-free human interaction: brief check-ins, shared reflections, structured dialogue. This breaks the solo-plus-AI isolation cycle. - Document Team AI Guidelines
Agree at the team level not just on when to use AI, but when to stop. Without this agreement, individual overwork becomes the new team baseline.
Key Distinction
Efficiency is local and functional. Productivity is systemic and purpose-driven. AI boosting efficiency doesn't automatically mean productivity went up. Confusing speed with direction and execution with value is exactly how you fall into the Jevons Paradox trap.
