The assumption almost everyone carries about a new tool: it pours rocket fuel on whoever is already at the front. Give the best people a sharper instrument and they pull further ahead. The gap between the stars and everyone else widens.
That is the story we have mostly told ourselves about technology and work, and for a long time it was probably the right story. As Erik Brynjolfsson put it, “For most of the past 30 years, computers and digital technologies have helped higher-skilled workers more than less-skilled workers, which has led to a growing gap in wages and income inequality.”
Then a study came along and ran the pattern in reverse.
A quick note before going further: I am not an economist or a labour-market researcher, and this piece is reading and reflection on one study, not advice. The numbers below come from a single field study of one company. It watched what happened rather than running a clean experiment, so it is a clue about how a tool behaved in one place, not a settled rule about every workplace.
What the study actually found
The spine here is a 2023 paper called “Generative AI at Work,” by Brynjolfsson at Stanford with Danielle Li and Lindsey Raymond at MIT. They tracked what happened when a big software firm gave its customer-support team an AI chat assistant, the same kind of technology behind ChatGPT. The sample was large: 5,179 agents. The tool was rolled out to different groups at different times, which gave the researchers a clean before-and-after to measure against.
On average, the tool raised the number of issues an agent resolved per hour by about 14 percent. That is the headline number, and on its own it is unremarkable. The interesting part is where the 14 percent came from. The gain was not spread evenly. Novice and lower-skilled agents improved about 34 percent. The most experienced agents? Close to nothing. They got a little faster and, if anything, slightly worse on quality.
One number stuck with me more than the percentages. Agents with two months on the job, using the tool, performed about as well as agents with six months of experience and no tool. The beginners did not creep up the curve. They jumped most of the way up it.
Why it ran the other way
So why did the tool help the people at the bottom and leave the people at the top flat? The researchers’ explanation is the part I find most interesting. The assistant had been trained on the company’s own records, which means it had quietly absorbed how the best agents handle a hard conversation: the phrasing that calms someone down, the order to ask questions in, the move that gets to a resolution fast. Then it offered that to everyone in real time. The researchers describe the system as capturing tacit knowledge. That is the stuff a veteran knows but would struggle to write down, the know-how that usually only passes on by sitting next to someone for a year. The tool packaged it and handed it out. Which explains the veterans. The bot, trained on millions of transcripts, learned what the best agents were doing right and spread it to everyone. If the suggestions are built from your own habits, the tool has little to teach you. Brynjolfsson’s noted: “It was fascinating to see that this technology goes the other way around,” he said, “it’s a good sign.”
I would hold all this carefully, though. This is one firm, where there happened to be a roughly correct way to handle a support ticket. Other research points in a different direction. The BCG and Harvard “jagged frontier” experiment with 758 consultants found AI lifted weaker performers more on tasks it was good at. But on a task picked to sit outside its strengths, AI users were about 19 percent less likely to reach the right answer.
What this looks like inside my own writing day
I can feel both halves of this from my own desk. I use AI tools hands-on, and I have spent a fair bit of time helping new writers get going with them. The first thing I have noticed is that AI only really earns its keep once you already know what you are trying to do. If you cannot say what you want from it, it is hard to get anything good out of it.
Then there is the part that stings a little. I used to have a research edge: two screens going, a knack for the exact search that would surface the one source nobody else had found. That edge is mostly gone now. The tool finds sources faster and wider than I ever did. The skill I spent years sharpening got cheap almost overnight, which is the ‘veteran’s’ predicament in miniature.
What did not move is the part I would have struggled to write down in the first place. Deciding which of those sources actually matters. Choosing the angle. Knowing when a clean paragraph is hiding a weak idea. The tool does the fetching fast. The judgment call still won’t move without me sitting there making it, and when I hand that part to the tool the work goes flat in a way I can feel before I can explain.
The catch worth keeping in mind
My read is that a smaller skill gap is mostly good news. It lets people with good ideas act on them without first paying years of dues. I have watched people with no coding background put together a small working project, because the part that used to gatekeep the whole thing, the syntax and the know-how, is now cheap. It makes generalists more nimble. It rewards taste and curiosity over the slow muscle memory of one trade.
But the imbalance is real. If your edge was the know-how, the thing that took you a decade to build, you should expect the floor to come up toward you. The thing the tool does not hand anyone is the idea, or the vision, or the decision about what is worth doing at all. It can give a beginner a veteran’s moves. It cannot give them a reason to make them.
If your work is the thing being reshaped here, and that lands somewhere closer to worry than curiosity, it is worth talking to someone, a mentor or a counsellor, rather than sitting with it alone.
