---
id: con-centaurs-and-cyborgs
schema: concept/1
layer: garden
status: established
name: Centaurs and cyborgs
aka: ["centaur work", "cyborg work", "self-automators", "ways of working with AI"]
summary: >
  Ethan Mollick's names for two ways of working well with AI: centaurs split a job, keeping some tasks
  and handing others to AI, while cyborgs go back and forth with it inside a single task. Vivienne Ming
  adds a weaker third style, the self-automator, who hands over the thinking.
first_seen: "2023-09"
relations:
  builds_on: [con-jagged-frontier]
  evolved_from: []
  related: [con-cognitive-offloading, con-productive-friction, con-human-in-the-loop]
  parallels:
    - idea: "centaur chess"
      field: "chess"
      note: "After losing to Deep Blue, Garry Kasparov promoted a format where a person and a computer play as one team; the name 'centaur' comes from there."
implementations: []
competencies: [cmp-steering, cmp-check-the-result, cmp-frame-the-outcome]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "Widely taught words for how someone is working with AI, drawn from the 2023 Boston Consulting Group experiment; gives learners a way to notice and choose their style."
---

In 2023, researchers gave 758 Boston Consulting Group consultants access to GPT-4. Writing about the results, Ethan Mollick noticed two patterns among the people who used it well. Centaurs drew a clear line, for example doing the analysis themselves and handing the slide drafts to AI. Cyborgs blended the two, writing a sentence, asking AI to extend it, editing, asking again.

Both styles keep a person judging the work, and both depend on knowing which tasks fall inside the AI's strengths, the jagged frontier. Vivienne Ming describes a third style, the self-automator, who hands over the whole task and the thinking with it. She calls it the weakest of the three. In a forecasting experiment she has described in interviews but not yet published, the cyborgs did best.

Centaur work is easier to check; cyborg work can do better when you stay alert.

## Example

A policy analyst writing a briefing works as a centaur on the numbers, which she checks herself, and as a cyborg on the wording, trading paragraphs with AI until the tone fits.

## How it connects

- Builds on the **jagged frontier**: both styles depend on knowing where AI is strong.
- The self-automator is **cognitive offloading** gone too far; **productive friction** pushes back.
- Both styles keep a **human in the loop**.

## Try it

On your next task with AI, notice which style you used. Then do a similar task in the other style and compare the results and how much you had to check.