Why AI Helps Some People and Hurts Others
Give a mediocre copywriter a good AI model and the change is immediate.
Within minutes they produce work that would have taken years of practice. Headlines get sharper. Structure improves. The awkward sentences disappear. The finished piece starts looking a lot like the work of someone who knows what they are doing.
Give the same model to a great copywriter and the difference is harder to see at first.
They reject the first headline, then the fifth. They tell the model the customer does not want what it thinks she wants. They add details about the offer that were not in the brief. They notice a paragraph sounding exactly like every competitor in the market. They push further into the psychology, change the angle, cut the language that sounds generated, and keep going until the piece says something worth saying.
Same model. Very different use of it.
I have come to think of AI as an extension of the mind operating it, and the research points somewhere more complicated than the usual prediction that giving everyone access to extraordinary intelligence levels the playing field.
The Floor Rises
There is good evidence that AI makes weaker performers dramatically better.
Researchers studying 5,179 customer support agents found that access to a generative AI assistant raised productivity 14% overall. The gain came almost entirely from novices and lower-skilled workers, at about 34%. Experienced agents saw little benefit.
The mechanism was interesting. The AI appeared to transmit patterns used by the company’s best agents to newer ones. Knowledge that previously took years to accumulate arrived while the work was happening.
A separate experiment published in Science found the same shape in professional writing tasks. ChatGPT made people faster and improved quality, with the larger gains among weaker writers. The spread narrowed.
Where the task is well defined and the model knows how to do it, this makes sense. A mediocre writer gets better prose. A junior employee inherits senior patterns. Someone staring at a blank page gets a competent draft in seconds. That is a genuine democratization of capability.
Then the task gets harder.
Tomorrow’s forecast, in your inbox.
One email. Five minutes. Written for owners, not engineers.
What Happens When the Machine Is Wrong
Harvard researchers working with Boston Consulting Group studied 758 consultants using GPT-4 on realistic tasks.
Inside the model’s range, the results were extraordinary. Consultants using AI completed 12.2% more tasks, worked 25.1% faster, and produced work rated substantially higher in quality. Lower performers gained the most.
Then the researchers included a task deliberately chosen to sit outside the model’s capability, where it would produce a plausible and incorrect answer.
On that task, consultants using AI were 19% less likely to reach the correct solution than those working without it.
The machine was useful enough that people trusted it where they should not have.
A follow-up paper looked at what happened when consultants tried to check the AI’s work on that task. Analysing GPT-4 logs from more than seventy of them, the researchers found that pushing back did not produce a concession. The model apologised, made a correction, then restated its original position with more supporting data and tighter reasoning. Challenged, it argued harder rather than disclosing its limits.
That is the part worth sitting with. Generating a plausible answer has become cheap. Recognising when a plausible answer should be rejected has not, and the machine will defend the wrong one with increasing sophistication while you try.
Yesterday’s issue covered the field experiment with Kenyan entrepreneurs that found the same split from the other direction: higher performers improved by more than 15%, lower performers declined by nearly 10%, on similar advice. What separated them was which recommendations they chose to act on.
Where the Value Moves
Copywriting makes this easy to see, since I have spent a long time doing it.
Being able to write good copy used to be the valuable thing. Years went into headlines, offers, sales arguments, pacing, objections, calls to action. AI has erased a surprising amount of that advantage. Give a strong model enough context and very good copy comes back almost immediately.
I still have to know what context to give it. I have to understand the business, know what makes the offer different, understand the buyer beyond the demographic description. I have to recognise when the model has produced something polished and generic, notice the sentence that reads well and is not true, catch the psychological premise underneath the whole piece being wrong.
And sometimes I know a piece is not right before I can explain why.
That ability becomes more valuable when producing another version costs nothing.
When competent output becomes abundant, the ability to recognise exceptional output becomes scarce.
It extends well past writing. Give someone with taste unlimited design options and they have more to curate from; give the same options to someone without it and they produce enormous volumes of work they believe is excellent. Give a curious person AI and they explore at a scale that was not previously available. Give the same tool to someone attached to being right and they have a remarkably capable machine for building arguments in favour of what they already think.
The Skills Nobody Puts on Their AI Résumé
Most of the preparation happening right now is technical. Learn prompting, learn agents, learn automation, learn the newest model. Those skills matter, and they keep getting easier. Techniques that felt advanced eighteen months ago now happen by default.
The harder ones sit inside the person operating the machine. Curiosity. Taste. Knowing enough about a domain to notice when an answer does not make sense. Sitting with several possibilities without needing one to win immediately. Being willing to find out you were wrong. Synthesising a great deal of information without treating all of it as equally important. Knowing what you actually want before asking a machine to help you get it.
And keeping enough confidence in your own judgment to challenge an intelligence that can make almost any position sound convincing.
There is evidence for that last one. Microsoft researchers studying 319 knowledge workers found that higher confidence in AI went with less critical thinking, while higher confidence in one’s own ability went with more of it.
That relationship gets more consequential as these systems get more persuasive.
The Forecast
AI narrows the difference in execution and widens the difference in judgment.
The floor keeps rising. People with less experience will keep producing work that once required years of practice, and the evidence for that is already strong.
The second-order effect is the interesting one. As baseline execution gets easier, value moves toward the abilities needed to direct, evaluate and occasionally reject what the machine produces.
The dividing line shows up at the edge of the model’s competence: ambiguous problems, unusual situations, high-stakes decisions, creative work where good is subjective, and the moment when it produces an excellent argument for the wrong answer.
Which changes what being good at AI means. The largest advantage will belong to people who bring something to the interaction before the first prompt: domain knowledge, curiosity, taste, self-awareness, and a developed sense of what good looks like.
The technology keeps improving. The quality of the mind operating it matters more because of that, rather than less.
Sources: Brynjolfsson E, Li D, Raymond LR. Generative AI at Work. NBER Working Paper 31161 · Noy S, Zhang W. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 2023 · Dell’Acqua F, McFowland E, Mollick E, et al. Navigating the Jagged Technological Frontier. Organization Science · Randazzo S, Joshi A, Kellogg KC, et al. GenAI as a Power Persuader. HBS Working Paper 26-021 · Otis N, Clarke R, Delecourt S, Holtz D, Koning R. The Uneven Impact of Generative AI on Entrepreneurial Performance. Management Science · Lee H, Sarkar A, Tankelevitch L, et al. The Impact of Generative AI on Critical Thinking. Microsoft Research / CHI 2025