Seventy-seven percent of employees using AI tools say the technology has increased their workload. Not decreased it. Increased it. That statistic, from Upwork’s 2024 workforce study, should stop every CTO mid-pitch. The promise of AI was fewer hours, less drudgery, more time for creative thinking. The reality, documented by researchers at UC Berkeley, MIT, and Carnegie Mellon, is the opposite: AI makes you faster at producing things nobody asked for until your boss realizes you can produce them, and then everyone asks for them. Welcome to the AI productivity trap.
UC Berkeley professors Aruna Ranganathan and Hatim Rahman published the definitive framing of this problem in Harvard Business Review. Their finding: AI tools create a workplace version of the Jevons Paradox, where efficiency gains don’t reduce workload but instead expand it through two reinforcing mechanisms.
The Jevons Paradox, Applied to Your Office
In the 1860s, economist William Stanley Jevons noticed something counterintuitive about steam engines. As they became more fuel-efficient, coal consumption didn’t drop. It skyrocketed. Cheaper steam power opened up new applications that hadn’t been economical before, and total demand outpaced the efficiency gains.
Ranganathan and Rahman identified two mechanisms that create the same dynamic in AI-augmented workplaces.
Mechanism 1: The Expectation Ratchet
When AI lets you draft a report in two hours instead of eight, your manager doesn’t give you six hours of free time. Your manager assigns four more reports. A study on data scientists found that generative AI tools initially improved efficiency, but managers responded by assigning more tasks with faster deadlines. The time savings evaporated.
The same pattern appeared on Upwork: AI reduced the cost and time of content production, so clients demanded more content, tighter deadlines, and lower prices. The freelancers using AI were working just as hard, just faster, with worse margins.
“The time savings get absorbed by rising expectations,” Ranganathan explained in a Haas podcast interview. “Managers see the work getting done faster and think, ‘Great, now do more.’ The treadmill speeds up.”
Mechanism 2: The Quality Escalator
When every lawyer has AI research tools, a brief that cites five precedents is no longer competitive. You need twenty. When every designer has Midjourney, a mood board with three concepts is table stakes. You need twelve.
AI raises the baseline of what counts as professional-quality output. The National Science Foundation documented that science and engineering publications grew over 50% between 2010 and 2022, a trend accelerating with AI writing tools. More papers get published, but no individual researcher’s life gets easier.
In law, AI-powered research tools sped up brief preparation, but the profession responded by redefining what a “well-researched” argument looks like. Lawyers now spend more time refining and expanding documents because their AI-equipped peers are doing the same. The bar moved up for everyone.
What the Productivity Numbers Actually Show
The corporate narrative says AI is supercharging productivity. The research says something more nuanced.
The MIT Reality Check
MIT researchers Daniel Rock and Neil Thompson ran one of the most rigorous AI productivity studies to date. Their finding: the time saved by generative AI tools was worth $0.50 per worker per day, roughly 0.5% of total task time. That’s the cost of a gumball, not a productivity revolution.
The study controlled for task type, worker skill, and tool access. The tiny gains weren’t because workers didn’t use AI. They did. But the gains from AI-assisted tasks were offset by the overhead of managing AI outputs: checking for errors, reformatting, and correcting hallucinations.
Google’s 6% Finding
Google ran an internal study across 10,000+ developers over six months. The result: a 6% increase in code output. Developers accepted about 25% of AI suggestions. Newer developers benefited more than experienced ones, which makes sense: AI coding assistants essentially give juniors access to pattern libraries that seniors already have in their heads.
The GitHub Copilot study that claimed 55% faster task completion has been widely criticized for testing simple HTTP server tasks in unfamiliar codebases. GitHub’s own 2024 survey of 2,000 developers found that 88% felt more productive, but only 10% reported “significant” gains. Stack Overflow’s data is bleaker: 97% of developers used AI tools, but only 3.4% saw productivity gains above 25%.
The Burnout Signal
Microsoft’s Work Trend Index found that 68% of workers say they don’t have enough focus time. Gallup’s 2025 data shows global engagement at its lowest point in a decade. The Pew Research Center found that only 36% of workers who use AI say it has made them more productive.
Marc Zao-Sanders put it well in his HBR piece on AI burnout: “It’s like giving someone a car that occasionally drives into walls and expecting them to relax behind the wheel.” AI tools don’t replace other tools. They add to the stack. Workers now juggle AI chatbots alongside email, Slack, project management platforms, and video conferencing. Each new AI feature generates notifications, outputs to review, and decisions to make.
The Historical Pattern Nobody Wants to Hear
This has happened before. Every time.
Email was supposed to streamline communication. The average office worker now processes 120+ emails per day. Nobody’s inbox is “streamlined.”
Spreadsheets were supposed to reduce the time accountants spent on calculations. Instead, they enabled increasingly complex models and reporting requirements. The total hours spent on financial analysis went up, not down.
Smartphones promised flexibility: work from anywhere, on your schedule. The result was the erasure of the boundary between work and personal time.
Each technology delivered on its core promise (faster communication, faster calculations, portable computing) while simultaneously expanding the scope of what was expected. The efficiency gain was real. The workload reduction was not. AI is following the same script, just faster.
The difference with AI is speed of adoption. Previous technologies took years to saturate a workforce. 79% of organizations already run AI agents in production, according to Mercer. The Jevons Paradox cycle that took email a decade to fully develop is playing out with AI in months.
Breaking the Treadmill: What Actually Works
Recognizing the trap is the first step. Here’s what the research suggests actually helps.
Subtract Before You Add
Kelly Monahan, Managing Director of Upwork’s Research Institute, argues that the core problem is additive: companies layer AI on top of existing workloads without removing anything. “Adding more technology to already overloaded plates isn’t going to solve productivity issues.”
Before deploying AI tools, identify which tasks should disappear entirely, not just get faster. A report that shouldn’t exist at all doesn’t need AI to write it faster.
Set Output Caps, Not Speed Targets
The expectation ratchet kicks in when organizations measure AI success by output volume. “We produced 4x more reports” is not a productivity win if nobody reads the extra three.
Organizations that avoid the trap set output caps: the same number of deliverables, at higher quality, with the time savings returned to workers for strategic thinking, skill development, or (radical thought) leaving on time. ActivTrak’s Gabriela Mauch put it directly: “When you layer new technology on top of existing workloads without removing anything, you’re not boosting productivity. You’re boosting burnout.”
Give Workers Control Over Integration
A systematic review in the Journal of Occupational Health found that AI-related technostress is a significant predictor of burnout, particularly when employees lack autonomy over how AI tools are integrated into their work. The fix is not “more training.” It’s letting workers decide which parts of their workflow benefit from AI and which don’t.
Calix CEO Michael Weening took this approach: he let employees build their own AI agents, resulting in over 700 employee-generated agents. When workers control the automation, they see themselves as the agent’s supervisor, not its replacement.
Measure What Matters
Stop measuring AI ROI by tokens processed, tasks completed, or time saved on individual activities. Start measuring employee satisfaction, focus time, work hours, and output quality. If your AI deployment saves two hours per person per day but everyone is working an extra hour because the expectations expanded, you have a net gain of one hour and a culture problem that will cost you in turnover.
Josh Bersin, the HR technology analyst, calls it the burnout paradox: organizations optimize for efficiency metrics while ignoring the wellbeing metrics that actually predict retention and long-term performance. The companies that will benefit from AI are the ones willing to let efficiency gains translate into less work, not just different work.
Frequently Asked Questions
What is the AI productivity trap?
The AI productivity trap is a phenomenon identified by UC Berkeley researchers Aruna Ranganathan and Hatim Rahman where AI tools, rather than reducing workloads, increase them through rising employer expectations and escalating professional standards. It is a workplace version of the Jevons Paradox, where greater efficiency leads to greater total demand.
Does AI actually increase workloads?
Yes, according to multiple studies. Upwork found that 77% of employees say AI has increased their workload. MIT researchers found AI saves only $0.50 per worker per day. Only 36% of AI-using workers told Pew Research that AI made them more productive. The gains from faster individual tasks are absorbed by expanded expectations and new work.
What is the Jevons Paradox and how does it apply to AI?
The Jevons Paradox, first observed in the 1860s, states that greater efficiency in resource use increases total consumption rather than reducing it. Applied to AI, it means that faster task completion leads to more tasks being assigned, higher quality standards, and expanded scope of work, rather than reduced working hours.
How can companies avoid AI-driven burnout?
Research suggests four strategies: remove tasks entirely before adding AI (subtract before you add), set output caps instead of speed targets, give employees control over how AI integrates into their workflow, and measure wellbeing metrics like focus time and satisfaction alongside efficiency metrics.
How productive does AI actually make workers?
Less than you would expect. MIT found AI saves $0.50 per worker per day. Google’s internal study of 10,000 developers showed a 6% output increase. Stack Overflow data shows only 3.4% of developers saw productivity gains above 25%. The gap between perceived and measured productivity gains is significant.