The Productivity Dividend That Isn't (And What That Tells Us About Every Big Technology Before This One)

In the early 1900s, factory owners who installed electric motors expected immediate gains. They didn't get them. For nearly three decades after electrification began spreading through American manufacturing, productivity statistics stayed stubbornly flat. The problem wasn't the technology. It was that factories kept their old layout — one central shaft spinning belts to all the machines — and just swapped the steam engine for an electric motor. It took a full generation of managers who had grown up with electricity, and who could imagine building a factory around it from scratch, before the gains showed up in the numbers.

We are, by most indications, in that first phase right now with AI.

Companies have spent billions on generative AI tools. The productivity dividend has not arrived. Researchers at the University of St. Gallen put a useful name to one reason why: "workslop." The idea is that rushed, unquestioned AI use doesn't just fail to improve output — it actively corrodes it. Meetings get prepped with AI-generated summaries nobody read carefully. Reports get drafted with AI prose nobody quite owns. The output looks finished. It isn't, quite. And the cost is subtle: a slow erosion of the judgment and craft that made the work good in the first place.

This is not an argument against AI. It's an observation about adoption curves, and they are remarkably consistent across the history of general-purpose technologies.

Electricity took 30 years to show up in productivity data. The internet took about 15. In both cases, the lag wasn't about the technology failing to work — it was about the gap between deploying a new tool inside old structures and actually redesigning the structures around the tool's real capabilities. The UNCTAD Technology and Innovation Report for 2025 makes a similar point, more carefully: it notes that realizing AI's potential requires not just infrastructure and data, but the willingness to redesign business processes wholesale. Most organizations are not doing that. They are doing the electric-motor-in-the-old-factory thing.

What makes this moment a little different from electricity or the internet is the speed of the capability curve. The underlying models are improving faster than most institutions can track, let alone respond to. That creates a specific kind of confusion: the tool you learned to use six months ago is already different, which means any workflow you built around it is probably already outdated. The workslop problem might partly be a symptom of this — people reaching for AI before they've figured out what role they actually want it to play, because the window for figuring that out keeps moving.

There's also a subtler thing happening with skills. The St. Gallen research points toward a worry that isn't really about AI producing bad output — it's about humans losing the capacity to recognize bad output. A junior analyst who gets their summary from a language model, week after week, without ever writing a summary themselves, isn't just being efficient. They're not developing the calibration that comes from doing the thing badly a few times and learning why. This is not a new concern — calculators raised it, spell-check raised it — and the evidence on those is actually pretty mixed. But the scope of what AI can now do is broader, which means the scope of what we might stop practicing is broader too.

The genetic research in the news lately offers a strange echo of this. Studies on recent human evolution show that our genome carries the signatures of very fast adaptations — populations hit by novel disease pressures, environmental shifts, new food sources — where the selection pressure was severe enough to override the usual preference for genetic diversity. The immune system region of ancient Anatolian farmers shows a "trough of diversity" that researchers describe as one of the strongest adaptation signals ever found. Something came along that was so disruptive it forced a rapid, costly reconfiguration.

Obviously, we are not being genetically selected by spreadsheet software. But there is something worth sitting with in that parallel. Technologies that become load-bearing infrastructure fast enough — that we depend on before we understand them — do change us. They change what skills we develop, what cognitive habits we form, what we outsource to the environment versus keep inside our heads. The question with AI is whether we're in the phase of thoughtful redesign, or the phase of just bolting a new motor onto the old shaft and wondering why the numbers aren't moving.

The honest answer is: mostly the latter, for now. That's not a permanent condition. The factories did eventually get redesigned. But the people who redesigned them had to be willing to question structures that had seemed fixed, and to build something genuinely new rather than something familiar with better parts.

The productivity dividend is probably coming. The more interesting question is what we'll have practiced — and what we'll have stopped practicing — by the time it does.