AI Giant Employees Reveal: AI Tools Make Work Harder Not Easier, 90-Hour Weeks Become Normal

This work intensity forms a sharp contrast with AI leaders' public promises. Google predicted several years ago that AI could make a four-day workweek possible. Earlier this year, OpenAI encouraged companies to experiment with four-day schedules without cutting pay, saying AI would soon accelerate much of the work people do. However, the former OpenAI employee said the company did not try a four-day workweek while they were there. Instead, they described regular 'crisis meetings,' weekend work, and 'super cut-throat' performance reviews that could lead to sudden layoffs. OpenAI and Anthropic did not respond to BBC's requests for comment. This 'saying one thing but doing another' phenomenon raises a profound question: if the creators of AI tools cannot themselves enjoy the work conveniences AI brings, how can ordinary workers expect AI will improve their work lives?

The workload is partly a result of the pace of the AI race. Companies are trying to improve models, build more computing infrastructure, and add AI features to products at the same time. That creates pressure on engineers, researchers, and product teams to move quickly and fix problems as they arise. Anthropic has said its Claude tool can work independently for seven hours on some tasks. Meta CEO Mark Zuckerberg has said AI will allow smaller teams to do more work. Yet employees say these systems are not reducing the amount of work expected from them. In many cases, they are simply being asked to produce more. Meta workers described being moved onto AI teams with little warning. A current employee and a former employee told BBC workers called the process being 'drafted.' The former employee said people often had little meaningful choice about joining an AI project. 'They just move you over,' the former employee said. 'You can't say no — or if you do, you have to quit.'

Those teams are working on AI systems for software engineering and other tasks, along with infrastructure that measures how well models can perform work typically done by people. The former employee said the work had no clear endpoint. 'You're literally working in teams of people trying to replicate humans doing jobs,' the former employee said. This paradox of 'using humans to replicate humans' reveals a fundamental contradiction in current AI development: AI system training and evaluation still heavily relies on human labor. From data annotation, model evaluation to safety testing, every step requires substantial human intervention. Ironically, to create 'automation' tools, humans need to expend more labor. This 'automation paradox' is not unique to the AI field — every technological revolution in history has experienced similar transition periods. But AI's uniqueness lies in that it replaces not just physical labor but cognitive capabilities, making the transition process more complex and difficult.

This phenomenon has sparked discussion of 'Jevons Paradox.' 19th-century economist William Stanley Jevons discovered that technologies improving coal use efficiency actually led to increased total coal consumption — because efficiency gains lowered costs, stimulating larger-scale use. Similarly, AI tools may follow similar logic: when AI makes certain tasks faster, companies won't reduce workload but will increase output expectations. Employees don't work less but are asked to do more. This 'efficiency trap' is particularly evident in the tech industry — email, instant messaging, and project management tools were all considered 'productivity-enhancing' tools, but ultimately they just let people process more information and tasks in the same timeframe. AI may be repeating this pattern: it's not liberating humans but creating new, higher-intensity work forms.