Source: Noy & Zhang (2023) — “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science
I like papers that test AI on work that feels painfully normal.
Not chess. Not benchmark puzzles. Not “can the model solve a problem invented by someone with too much coffee?”
Just writing. Reports, emails, analysis plans, press releases. The sort of work that quietly eats the calendar and then pretends it was only a small task.
This experiment gave 453 college-educated professionals realistic writing tasks and randomly gave half of them access to ChatGPT. The results were not subtle. Time fell by about 40%. Output quality rose by about 18%.
That is already a big deal.
But the part I find more interesting is the inequality result. ChatGPT helped lower-performing workers more, which narrowed the gap between stronger and weaker writers. In other words, AI did not just make everyone faster. It acted a little like a writing exoskeleton for people who needed more support.
That has a very practical implication for organisations.
If a task depends heavily on turning messy thoughts into acceptable prose, AI can raise the floor. The first draft becomes less terrifying. The blank page becomes less blank. People who might have spent twenty minutes politely wrestling with one paragraph can get to something usable much faster.
Of course, this does not mean ChatGPT makes everyone brilliant. It means the boring middle layer of writing becomes cheaper.
And that changes what we should value.
If everyone can produce a decent draft faster, the advantage shifts toward knowing what to say, what not to say, and whether the draft is actually true.
My takeaway: AI productivity is not just about speed. It changes where the human bottleneck sits. The writing gets easier. The judgment becomes more visible. Slightly inconvenient, but probably healthy.