The Paradox at the Center of the AI Boom
Inside the companies racing hardest to build artificial intelligence, workers are telling a story that quietly undercuts the marketing. Employees at OpenAI, Anthropic, Meta, and Google told the BBC that the AI push has brought longer hours, heavier workloads, and nights and weekends swallowed by urgent projects. That is not what the industry promised.
Tech companies have spent years saying AI will help people accomplish more in less time. The awkward finding inside the leaders of that race is that more productive tools have not produced shorter working days. If anything, the opposite is happening, and the reasons are more structural than any single boss or deadline.
Productivity is real. The time it saves is not. The twist is that employers increasingly have evidence that AI genuinely accelerates work. OpenAI says Codex has become the primary AI work tool across every department, including Legal, Finance, and Recruiting. As its capabilities improved, employees handed it longer and more complex jobs, and used it to take on tasks outside their usual roles.
Anthropic's internal numbers point in the same direction. More than 80 percent of code merged into its codebase was authored by Claude as of May, and the typical engineer merged eight times as much code per day during the second quarter of 2026 as in 2024.
Anthropic cautions that code volume overstates the true productivity gain, but says the increase still indicates substantial acceleration. The pattern is consistent across the industry: the tools work, the output climbs, and the workload climbs with it.
90-Hour Sprints and Weekend Catch-Up
The human side of that acceleration is punishing. A former OpenAI technical employee told the BBC they worked roughly 70 hours a week, sometimes using weekends to catch up. Workers at OpenAI and Anthropic said intense release sprints could exceed 90 hours in a seven-day period.
At Meta, one current and one former employee described workers being abruptly moved onto urgent AI projects, with some assignments stretching into nights and weekends. Former Google engineer Amin Shali said AI priorities affected him more indirectly: as resources were redirected toward AI projects, internal engineering problems piled up, and overnight work followed. He said his sleep and overall health improved after he left Google.
Here is what the BBC's reporting found at each company:
- OpenAI: roughly 70-hour weeks, weekend catch-up sessions, and release sprints past 90 hours.
- Anthropic: engineer output up eightfold per day since 2024, with most merged code now written by Claude.
- Meta: abrupt reassignments to urgent AI work, with shifts bleeding into nights and weekends.
- Google: overnight work tied to internal engineering problems as staff moved to AI priorities.
Who Keeps the Hours AI Saves?
In-progress UC Berkeley research offers one explanation. During an eight-month study at a 200-person tech company, researchers found employees using generative AI worked faster and expanded what they attempted. Work also crept into periods that previously acted as breaks, often without managers explicitly asking employees to work longer.
OpenAI itself has floated the opposite outcome for those efficiency gains. Its Industrial Policy for the Intelligence Age suggests incentivizing 32-hour or four-day workweek trials when AI reduces routine workloads and operating costs, with reclaimed hours potentially becoming permanently shorter weeks or paid time off.
That may be the more important AI productivity question. If these tools really can give companies hours of human labor back, who gets to keep those hours? For the moment, the evidence inside AI's biggest employers suggests the answer is often nobody. The saved time gets reinvested in the next sprint, the next model, the next release, and the workers building it stay on the clock.
Comments