Whose voice is this? This is a researched Digital Detox explainer. Michael's first-person 90/30 method is identified in its own section. The workplace findings come from the linked sources and their limitations are stated beside them.
A tool can reduce the time needed for a task without reducing the length of the working day.
The difference is easy to miss. A draft that once took two hours may take forty minutes. That does not mean the remaining eighty minutes automatically becomes rest. It may become another draft, another project, more checking or a new kind of work that was previously handed to somebody else.
Using AI well therefore requires two decisions. The first is how the tool will help with the task. The second is what happens to the time and attention it saves. If the second decision is left open, faster work can simply become more work.
A faster task and a shorter day are not the same result
In 2023, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of an AI assistant among 5,179 customer-support agents. Access to the tool increased productivity, measured as issues resolved per hour, by about 14 per cent on average. The largest gains went to newer and less-skilled workers.
This is strong evidence that generative AI can improve performance in a bounded workplace task. It does not show that the agents worked fewer hours. The study measured work completed per hour, not time returned to the worker.
The concession matters
AI really can save time on some tasks. Denying that would make this page ideological. The unresolved question is who receives the saved time and what new work is allowed to fill it.
AI can remove the natural stopping points from work
In an eight-month study at a 200-person American technology company, UC Berkeley researchers Xingqi Maggie Ye and Aruna Ranganathan observed how generative AI changed everyday work. Their in-progress research found that employees worked faster, took on a wider range of tasks and extended work into more hours of the day.
AI moved into pauses that once separated one task from the next: another prompt over lunch, a quick iteration before a meeting, or several AI-assisted threads running at the same time. A person no longer had to wait for a colleague or specialist before continuing, but that removed some of the moments when the work would naturally stop.
The limits are important. This was one company, not a representative sample of every workplace. The research is still in progress and does not prove that AI inevitably lengthens every working day. It does describe a mechanism that workers and teams can look for in their own work.
AI can quietly turn one job into several jobs
OpenAI's July 2026 analysis of more than 800,000 work-related ChatGPT messages found that 43.5 per cent of non-generic, occupation-specific messages concerned tasks associated with another occupation. A salesperson might analyse a dataset. A small-business owner might draft marketing copy, review a contract or troubleshoot a website.
This can be useful. It can also make a job expand sideways. A person who once handed a task to a specialist may now be expected to attempt it themselves while keeping all their existing responsibilities.
The research comes from the company that sells ChatGPT and does not measure hours, wellbeing or work quality. It shows task crossover inside one service, not the complete effect of AI on employment. Its practical value is narrower: when adopting AI, write down which tasks are being added and which tasks are actually being removed.
The longest weeks inside AI companies show that culture can swallow productivity
BBC technology reporter Kali Hays spoke with current and former workers at several companies developing AI. Some described seventy-hour weeks and product sprints reaching ninety hours. The report also described weekend work, urgent transfers to AI projects and workers feeling on call outside normal hours.
These accounts are reporting from a small number of workers, many of them anonymous. They do not establish how common a ninety-hour week is or prove that AI caused those hours. They do show that a company can possess advanced automation while choosing a culture that consumes more human time.
The tool does not decide whether a productivity gain becomes rest, more output, lower staffing or a larger workload. People and organisations make that decision.
Michael's 90/30 rhythm protects the human part of the work
Michael writes his own starting idea before asking AI to help. He then uses the tool for roughly sixty to ninety minutes and takes twenty to thirty minutes away from the screen. The exact timing is flexible. The break is the point.
Walking, meditation, exercise, a shower or quiet reflection gives the material time to settle. He returns as the editor: keeping what still feels true and useful after the speed and novelty of generation have passed.
This is a personal working method, not a clinical or productivity study. It is useful because it creates a finish for the AI session and protects time in which Michael is not responding to another machine-generated possibility. Read the complete first-person method in 7 Human Skills AI Can't Replace.
Decide what will finish before opening the AI tool
“Use AI for the proposal” has no end. “Find three objections to this proposal, check each one and then close the tool” does. A clear finish makes it possible to judge whether the tool helped and prevents the session from becoming an endless search for a better version.
Name one result
Write the document, decision, analysis or problem that must be completed. Do not open AI simply because work feels difficult.
Do the first thinking yourself
Record what you already know, the point you want to make and the standards the answer must meet. This gives the tool a direction and makes its mistakes easier to see.
Choose what AI is allowed to do
Ask for a bounded contribution: alternatives, objections, a calculation, a summary of supplied material or help organising a draft. Do not quietly expand the task because the next possibility appears interesting.
Run one thread at a time
Several agents or chats running together can create more material than anyone can check. Finish or stop one line of work before starting another unless parallel work is genuinely necessary.
Budget time for verification
A fast answer still needs checking. Decide how much evidence, testing or human review the result requires before treating it as finished.
Write down where the saved time goes
If a two-hour task takes forty minutes, decide in advance whether the remaining time belongs to another named task, a shorter working day or something outside work.
Close the work
Stop the AI session, close work accounts and leave a short return note for tomorrow. A tool with endless suggestions needs an external end.
Individual boundaries cannot repair an unlimited job
A personal finish line helps only when a person has some control over the work. It cannot fix an employer that adds tasks faster than anyone can complete them or expects replies at all hours.
A 2025 study in Nature Human Behaviour followed 2,896 employees in 141 organisations during reduced-working-time trials. Wellbeing improved relative to twelve comparison organisations. The companies were not randomly selected, and the model will not transfer unchanged to every workplace. The study nevertheless shows that time can be returned when organisations deliberately reduce it.
The health boundary is serious. WHO and ILO reviews compared people working at least 55 hours a week with people working 35 to 40 hours. They estimated a 17 per cent higher risk of ischaemic heart disease and a 35 per cent higher risk of stroke in the longer-hours group. These are population estimates over long exposure, not a prediction about one difficult week.
If the workload itself is excessive, the answer cannot be another personal-efficiency system. The job, staffing, expectations and right to disconnect need to be part of the conversation.
Sources and limitations
- Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161, revised November 2023. Measures task productivity among 5,179 customer-support agents; not shorter working time.
- UC Berkeley Haas, AI promised to free up workers' time. Researchers found the opposite, 18 February 2026. In-progress ethnography at one 200-person technology company.
- OpenAI, How AI is expanding what people do at work, 27 July 2026. First-party analysis of more than 800,000 messages from a company with a commercial interest.
- Kali Hays, BBC, Tech leaders say AI means less work — their staff say they work up to 90 hours a week, 10 August 2026. Original reporting based substantially on worker accounts; not a prevalence study.
- Fan, Schor, Kelly and Gu, Work time reduction via a 4-day workweek finds improvements in workers' well-being, Nature Human Behaviour, 2025. Multi-country intervention with comparison organisations; participating companies were not randomly selected.
- WHO and ILO evidence on long working hours, heart disease and stroke, 17 May 2021. Population estimates based on systematic reviews and meta-analyses; not an individual diagnosis.
Evidence checked 11 August 2026. This page does not claim that AI always lengthens work or that one personal method can repair an excessive workload.
Michael has spent much of his working life online—and much of the rest of it trying to stop that technology taking over. He has thrown a phone into a lake, locked devices in boxes and experimented with long periods offline. He writes about what helped, what failed and how to keep the useful parts of digital life.