A response to Theo Leggett’s “In the AI gold rush, tech firms are embracing 72-hour weeks” (BBC News, February 2026) https://www.bbc.co.uk/news/articles/cvgn2k285ypo
Leggett's BBC article warns that the AI boom is normalising gruelling 72-hour work weeks, framing this as a risky gamble on productivity. Long hours might generate innovation, the article suggests, but they carry high costs in burnout and health damage. On that narrow point, the evidence is clear. The World Health Organization links working more than 55 hours a week to roughly 745,000 annual deaths from stroke and heart disease. UK four-day-week trials show that reduced hours do not lower output. The trend is real, and the health risks are undeniable.
But the entire narrative rests on a model of productivity that no longer fits the world it claims to describe. The article treats work as if it still operates according to assumptions that collapsed decades ago. No one involved appears to notice.
Mistake One: Treating AI as a Generic Input
The BBC treats AI as if it were a simple technological input. Add more hours of it, get more output. This framing assumes that tasks are fixed, productivity is measurable, and working longer with AI means working better. None of these assumptions hold.
Large language models are not like earlier workplace tools. A spreadsheet does one thing. A database does one thing. An LLM can be a drafting assistant, a research engine, a coding aid, a conversational partner, or something genuinely unexpected, depending entirely on how it is used.
The same model, with the same interface, can produce radically different outcomes. Someone using AI to process emails faster is operating in a completely different mode from someone using it to interrogate a half-formed idea and discovering that the project itself needs to be rethought. Treating both as interchangeable “AI-assisted workers” whose productivity can be inferred from hours logged is analytically incoherent.
LLMs certainly speed up well-defined tasks. But their real value lies elsewhere. They change what the task is. That kind of outcome cannot be scheduled, standardised, or forced by longer hours. It requires attention, curiosity, and cognitive flexibility. Exhaustion reliably destroys all three.
This open-endedness is the most important feature of LLMs and the first thing organisations eliminate when they try to standardise their use. The BBC article does not even register the problem.
What Is Productivity, Actually?
The article concedes, in passing, that hours worked are only a proxy for productivity. Everyone knows this. But the concession does no work, because productivity itself is never defined.
What is productivity as a phenomenon rather than a metric? Where does it come from? What conditions generate it, and what conditions undermine it?
The article never asks. Productivity is treated as self-evident. Innovation is treated as a vague acceleration of output. This is not carelessness. It is the effect of an inherited framework that prevents these questions from being asked at all.
The article relies on what might be called a laboratory model of work. Isolate a variable, hold everything else constant, measure output. This model worked for assembly lines. It worked for tightly scripted service work. It works wherever labour has been deliberately engineered so that productivity can be measured.
This logic originates with Adam Smith’s pin factory. Break work into discrete steps, specialise labour, count outputs, and observe gains. The argument is elegant, but it assumes the world behaves like a controlled experiment. Industrial systems were later organised to make that assumption true. Assembly lines are not just technologies. They are epistemological machines that strip work of everything that resists measurement.
Most contemporary work does not behave this way, especially not AI-assisted knowledge work. Tasks change during execution. The most valuable outcome may be abandoning the original problem altogether. When studies show that ChatGPT enables people to complete writing tasks 40 percent faster, they are measuring optimisation of a fixed task. They cannot measure what happens when working with an LLM changes what the task even is.
This is not a minor oversight. It is a structural blind spot.
Productivity Is Not Innovation
The article collapses productivity and innovation into a single register. This is a serious error.
Productivity gains are real. When tasks are stable and goals are clear, output per hour can increase dramatically. Automation and AI excel in such settings. That matters.
But it is not innovation. Innovation begins when something fails to work as expected. When a task resists completion. When a goal becomes unclear. When existing approaches produce discomfort or doubt. That felt misalignment is not inefficiency. It is the signal that something new may be required.
From the outside, productivity and innovation both look like doing more with less. Internally, they are opposites. Productivity presupposes that the task is already known. Innovation presupposes that it is not yet knowable. The BBC article slides between the two as if the distinction were trivial. It is not.
Even process optimisation can be genuinely innovative, but only when it emerges from lived practice rather than abstract efficiency targets. Skilled practitioners constantly make micro-adjustments that improve coordination. These innovations typically disappear directly into routine. They leave no symbolic trace. They are invisible to metrics and therefore absent from studies. The article’s framework cannot see them at all.
How Organisations Destroy Value
The laboratory model is not just analytically flawed. It is performative.
When organisations adopt it, they reshape work to fit it. They install KPIs, time tracking, and output metrics. They reward what is visible and penalise what is not. In doing so, they suppress the very forms of value creation that depend on open-ended exploration.
This is not merely a failure to recognise innovation. It is active destruction. When measurement is imposed on domains where coordination depends on remaining unsymbolised, it does not simply misrepresent value. It dismantles the conditions under which that value can exist.
AI could support genuinely new forms of thinking. Instead, it is often deployed as a speed-enhancing tool inside frameworks that can only recognise it as a faster typewriter.
What the Evidence Actually Shows
The BBC cites studies showing that output barely increases beyond 50 hours per week. The health evidence is devastating. The UK four-day-week trial found no loss of productivity. More striking still is the most rigorous real-world study of AI and productivity to date. Humlum and Vestergaard’s 2025 analysis of 25,000 Danish workers found no significant effects on earnings or working hours. Average time savings were just 2.8 percent.
This gap between laboratory demonstrations and real-world outcomes mirrors the productivity paradox observed by Robert Solow in 1987. You can see the computer age everywhere except in the productivity statistics.
The BBC’s appeal to “studies show” reinforces the deeper problem. These studies measure what can be measured and correlate proxies with proxies. They do not ask whether the thing being measured corresponds to the phenomenon being explained.
What Would Actually Help
For organisations using AI, 72-hour weeks are not just unhealthy. They are counterproductive in the strict sense.
LLMs can help sustain exploration, allow rapid iteration without commitment, and enable people to hold questions open longer. None of this survives exhaustion. None of it survives surveillance metrics. None of it survives cultures that treat hours logged as a proxy for commitment.
The organisations that benefit most from AI will not be those that work the longest. They will be those that protect slack, trust judgment, and allow problems to be followed wherever they lead.
Innovation cannot be mandated. It cannot be reliably measured. And it certainly cannot be produced by treating an open-ended technology as if it were a faster assembly line.
The ease with which we slide between hours, output, productivity, and innovation is itself diagnostic. If these distinctions were genuinely installed in how we think about work, the BBC article would read as irresponsible rather than unremarkable. That it does not tells us something important. Our dominant myths about work refuse to die, even as the tools we use demand that we rethink them.