What AI Efficiency Cannot Deliver — and What Workers' Hours Reveal
Tech leaders promise AI will reduce work hours. Workers report logging up to 90 hours a week. Research shows why efficiency gains become more work, not less.
For years, executives at companies investing hundreds of billions of dollars in artificial intelligence have insisted the technology will reduce how much time people spend working. An engineering director at Google said in 2021 that AI would deliver a four-day work week by 2025. Earlier in 2026, OpenAI formally urged companies to test a four-day work week with no change in pay, claiming AI would soon speed up so much human labor that the corporate world should prepare.
The workers actually using these tools tell a different story.
A former OpenAI employee who left the company in 2025 told the BBC they worked at least 70 hours a week — far more than in previous tech jobs. The person now works at an AI start-up and described their hours as closer to 50 to 60 per week, “outside of sprints.” At OpenAI and Anthropic, those sprints can stretch for weeks and exceed 90 hours in a seven-day period, according to multiple tech workers the BBC interviewed.
The irony is not lost on observers: OpenAI never trialled the four-day work week it suggested others should adopt. Instead, the former employee described a culture of frequent “crisis meetings,” weekend work, and “super cut-throat” performance reviews.
What the research shows
The gap between promise and experience is not limited to Silicon Valley’s most visible companies. A study from UC Berkeley followed hundreds of workers at a US tech company over eight months as they used AI tools. The employees worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day.
A longitudinal study by the UC Berkeley Labor Center found that 67 percent of AI adopters worked more hours by year-end. The pattern held across roles: workers who initially saved time through AI tools saw those gains absorbed by expanded workloads.
The mechanism is straightforward. A 2025 study from METR found that experienced developers were 19 percent slower when using AI tools, despite believing they were 20 percent faster. The discrepancy came from the overhead of managing AI output — checking for errors, correcting mistakes, and integrating results — which added cognitive load on top of the original task.
Where the time savings go
Neil Thompson, an innovation scholar at MIT, told the BBC that even at companies leading AI development, workers are rarely told to deploy efficiency tools and move on with fewer hours. “It leads to a situation where even if there were real time savings, it would be sucked up by the changes, and implementing them, and making sure they worked,” Thompson said.
The UC Berkeley research found that workloads expanded partly because employees needed to constantly verify AI output. In situations where workers had smoothed out processes for AI tools, any time saved tended to fill with new work — either voluntarily or because proving value to an employer required taking on more.
“People assume that 20 percent less work means four-day weeks,” Thompson said. “But new work emerges.”
The NBER published a survey in March 2026 of 6,000 executives that found approximately 90 percent of firms reported zero impact on employment or productivity from AI adoption. The tools were being used. The promised efficiency had not materialized at the organizational level.
What companies are doing
At Meta, workers described being abruptly pushed onto teams handling urgent AI projects. They called it being “drafted,” according to current and former employees, because people were not given a choice. “They just move you over,” one former employee said. “You can’t say no — or if you do, you have to quit.”
The hours on these teams are often long, with staff working into the night and on weekends, feeling “on call” during off-hours. Meta’s current AI projects include building tools for software engineering and infrastructure that measures how well AI models replicate human tasks. “You’re literally working in teams of people trying to replicate humans doing jobs,” the former employee said. That work, they added, seems “endless.”
Even workers not directly involved in AI development face longer hours. A former Google employee, Amin Shali, left the company in May 2026 due to the ways AI negatively impacted his job — including working through the night because internal engineering functions had failed as Google shifted crucial resources to AI projects. Since leaving, Shali told the BBC his sleep and health improved. He realized that heavy reliance on AI tools “creates a bad culture with excess pressure on engineers.”
The broader picture
The ManpowerGroup reported in 2026 that regular AI use among workers rose 13 percent, but worker confidence in the utility of those tools dropped 18 percent. The experience of using AI did not match the promise.
A BCG study from 2026 found that 42 percent of frontline employees saved eight hours weekly through AI — but 66 percent received no corresponding reduction in workload. The time savings existed. The work simply replaced it.
In the US, there are no legal limits on how many hours a person over 16 can work. In the UK and Europe, laws cap the work week at 48 hours, including overtime. OECD data shows US employees work roughly 1,800 hours annually — more than any other G7 nation. Italy comes next at approximately 1,715 hours, while Germany averages 1,332.
The structural context matters. When efficiency gains arrive in a market with no cap on hours, there is no mechanism that converts saved time into reduced work. The savings become a signal for more tasks, not fewer.
What the tension reveals
The promise of AI reducing work hours rests on a specific assumption: that saved time will be returned to workers as leisure. But the evidence from companies building the tools, workers using them, and academic studies tracking outcomes all point to a different result.
Efficiency gains are being captured as expanded scope. The tools make individual tasks faster. The organization responds by assigning more tasks. The worker ends the week having accomplished more — and worked longer.
As Shali put it: “AI is supposed to be doing so much for us now, so many more people should at least have better health and better sleep.”
Instead, the workers closest to the technology report the opposite.