Centaurs and Cyborgs on the Jagged Frontier

"we took hundreds of consultants and randomized whether they were allowed to use AI"

Randomizing is the point. Compare AI users to non-users in the wild and you mostly measure who adopts new tools; a coin flip makes the groups comparable, so the difference between them is caused by the AI.

The tasks were ordinary consulting work — idea generation, market segmentation, marketing copy, a persuasive memo — for a fictional shoe company, vetted for realism by a shoe-company executive. The tooling was deliberately unremarkable: plain GPT-4, no fine-tuning, no elaborate prompting. The gains are what you get from handing capable people an ordinary chatbot.

"Imagine a fortress wall, with some towers and battlements jutting out into the countryside, while others fold back towards the center of the castle."

The frontier is the set of tasks the AI can do; jagged means its difficulty ordering does not match yours. A sonnet, formally intricate, is easy; an exactly-50-word poem is hard, because models read text as tokens — chunks that are often fragments of words — so counting words is awkward for them.

But jaggedness is not the dangerous part. Invisibility is. A border you could see would be a user manual. Because you cannot see it — and because a model states errors in the same fluent register as successes — the output never tells you which side you are on.

"consultants using ChatGPT-4 outperformed those who did not, by a lot. On every dimension. Every way we measured performance."

The headline numbers: 12.2% more tasks finished, 25.1% faster, 40% higher quality by the graders' ratings. Those measure throughput, speed, and rated quality on tasks with no single right answer — real, but not the same as accuracy.

Then the finding people underrate: AI acted as a "skill leveler". The consultants who scored worst at baseline gained most, 43%; the top performers gained less. That cuts two ways: if finished output is what matters, leveling is wonderful; if the gap between competent and excellent is how novices become experts, it is an unsolved problem wearing the costume of a compliment.

"human consultants got the problem right 84% of the time without AI help, but when consultants used the AI, they did worse"

BCG built one task to fall outside the frontier: a problem where the AI would give a confident, plausible, wrong answer a competent human could see through — which, Mollick notes, was hard to arrange. The result inverts everything above: 84% correct unaided, 60–70% with AI.

The mechanism has a name, from Dell'Acqua's separate study of recruiters: "falling asleep at the wheel". When the assistant is right almost every time, checking stops feeling worth the effort, so people stop — including when checking was the job. AI users also wrote up their wrong answers better: fluency and correctness came apart, and only one of them was visible.

"Centaurs have a strategic division of labor, switching between AI and human tasks"

Two working styles, not two kinds of person. The centaur draws a line and allocates across it by strength: I choose the statistical approach, the AI builds the chart. The cyborg refuses the line — "Cyborgs don't just delegate tasks; they intertwine their efforts with AI" — starting a sentence for the model to finish, handing over fragments, crossing continuously.

What makes the pair useful is what both refuse to do: hand over a whole task and walk away. That refusal separates the consultants who won on both halves of the experiment from the ones who fell asleep — and there is no manual for locating the wall, only walking into it on work you know well enough to catch the errors.