Cyborg Authoring

Thinking ToolsAny stepAlso known as Centaurs and cyborgs

From the centaur and cyborg distinction of Ethan Mollick (2023; Co-Intelligence, 2024), grounded in the field experiment of Fabrizio Dell'Acqua, Mollick, and colleagues (2023).

A writing routine: a person sets the outline and claims, an AI drafts sections under tight limits, and the two trade critique until a final human pass.

One section of a churn report, passed between a person and an AI model

At each of the analyst's turns, pick a move: the one that keeps her judgment on the claims and the final words, or the shortcut that hands them to the model. The model's reply shows what each one produces. Illustrative.

Analyst

AI model

Analyst

AI model

Select a stage to read it, or play the sequence.The setup. A subscription firm's churn rose last quarter, and the analyst owes a report in four days. Pick how she starts the model.Pick a move to see the reply.A confident 1,500 words, with an industry benchmark nobody can source.Handed over: the model chose the claims, and some of them are invented.A 250-word draft that follows the claims, with every sentence hedged.Bounded: her claims and data, its drafting.The draft. The draft is usable. It says may and could in every sentence, and one claim lacks its number.Edit and feed back. Pick her next move.Pick a move to see the reply.Done.Accepted as is: the hedges and the gap go out under her name.A tighter draft: plain claims, and the 14 percent in place.Edited and fed back: her edits teach the next draft her standard.The redraft. The second draft reads close to her own voice, because her edits told it how.Critique, reversed. She wrote section 3 herself. Pick what she asks for.Pick a move to see the reply.Claim 3 blames the price rise, but the March onboarding change fits the data too.Critique traded: the model tests her reasoning, and finds a gap.A section 3 that agrees with everything in section 2.More drafting: nobody checks the argument.A rival cause. She checks: the churn rise starts two weeks after the onboarding change, before the price rise reached most customers.Final pass. Pick how she finishes.Pick a move to see the reply.A smoother report, with the onboarding point softened back into a maybe.The last pass handed back: the model undoes her strongest finding.Two numbers corrected, and the opening now leads with the onboarding finding.A human final pass: every claim and number is one she stands behind.Sent. The report goes out Thursday. It leads with the onboarding change, which the first prompt would never have found.

Moves that kept her judgment: 0 of 4Moves that kept her judgment: 1 of 4Moves that kept her judgment: 2 of 4Moves that kept her judgment: 3 of 4Moves that kept her judgment: 4 of 4

  1. The setup (Analyst). A subscription firm's churn rose last quarter, and the analyst owes a report in four days. Pick how she starts the model.
    • Write a report on why our churn went up last quarter. A confident 1,500 words, with an industry benchmark nobody can source. Handed over: the model chose the claims, and some of them are invented.
    • Draft section 2 in 250 words from my three claims and this table. Add no new claims. A 250-word draft that follows the claims, with every sentence hedged. Bounded: her claims and data, its drafting.
  2. The draft (AI model). The draft is usable. It says may and could in every sentence, and one claim lacks its number.
  3. Edit and feed back (Analyst). Pick her next move.
    • Looks fine. Paste it into the report. Done. Accepted as is: the hedges and the gap go out under her name.
    • Cut the hedges in paragraph two. Claim 3 needs the 14 percent from the table. Redraft. A tighter draft: plain claims, and the 14 percent in place. Edited and fed back: her edits teach the next draft her standard.
  4. The redraft (AI model). The second draft reads close to her own voice, because her edits told it how.
  5. Critique, reversed (Analyst). She wrote section 3 herself. Pick what she asks for.
    • Here's my section 3. What is its weakest claim, and what would a skeptic ask? Claim 3 blames the price rise, but the March onboarding change fits the data too. Critique traded: the model tests her reasoning, and finds a gap.
    • Write section 3 as well. A section 3 that agrees with everything in section 2. More drafting: nobody checks the argument.
  6. A rival cause (AI model). She checks: the churn rise starts two weeks after the onboarding change, before the price rise reached most customers.
  7. Final pass (Analyst). Pick how she finishes.
    • Run the whole report through the model once more for polish, then send it. A smoother report, with the onboarding point softened back into a maybe. The last pass handed back: the model undoes her strongest finding.
    • Read it aloud, check every number against the table, and rewrite the opening myself. Two numbers corrected, and the opening now leads with the onboarding finding. A human final pass: every claim and number is one she stands behind.
  8. Sent (Analyst). The report goes out Thursday. It leads with the onboarding change, which the first prompt would never have found.

An example

An analyst at a subscription meal-kit company had four days to explain why churn rose last quarter. A colleague's first attempt, a single prompt asking an AI model for the report, had produced 1,500 fluent words citing an industry benchmark no one could trace.

She split the report into sections and wrote the outline and the three claims herself, with two paragraphs from an earlier report as a sample of the voice. For each section she gave the model the claims, the data table, a word limit, and one rule: add no claims. She marked up each draft, cutting the hedges and adding a missing figure, and sent the edits back. She wrote the third section herself, then asked the model for its weakest claim. It pointed out that the March onboarding change fit the data as well as the price rise did, and when she checked, the rise in cancellations began two weeks after it.

The last pass was hers alone: she read it aloud, checked every number, and rewrote the opening around the onboarding finding.

Why it works

AI drafting is fast and fluent, and its failures are the ones a reader cannot see: invented facts, borrowed confidence, and the quiet substitution of plausible claims for true ones. The routine keeps the work a person is accountable for, the claims, the data, and the final words, in human hands, and gives the model the part it does well, turning a settled outline into prose.

The interleaving is what separates it from simple delegation. Short sections with tight limits keep each draft small enough to check. Feeding edits back trains the next draft toward the writer's standard. Reversing roles, so the model critiques the person's section, uses it as a tireless skeptic that will name a rival explanation without worrying about anyone's feelings. And the final human pass catches what every earlier pass let through, which is usually a number or a hedge.

The thinking behind it

Ethan Mollick Ethan Mollick, a professor at the Wharton School who studies entrepreneurship, innovation, and the use of AI at work, describes two ways of dividing work with AI in Co-Intelligence: Living and Working with AI (2024). Centaurs split tasks cleanly between person and machine; cyborgs interleave the two within a single task, passing it back and forth. The routine here is a cyborg pattern for writing. The distinction is grounded in a 2023 field experiment by Fabrizio Dell'Acqua, Mollick, and colleagues.

How to run it

  1. Split the piece into sections. Cut it into parts small enough that one draft can be checked in a few minutes.
  2. Write the skeleton by hand. Set the outline, the claims, and the data each section rests on, plus a sample of the voice.
  3. Draft under stated limits. Give the model one section at a time with a word limit and a rule against new claims.
  4. Edit and feed the edits back. Mark up each draft and send the changes with it, so the next one starts closer.
  5. Ask it to attack the human sections. Have the model name the weakest claim in a section the writer drafted, and the question a skeptic would ask.
  6. Check what it finds. Test every rival explanation or gap against the data before changing a word.
  7. Finish with a human pass. Read the whole piece aloud, verify every number, and write the opening and close by hand.

Where it goes wrong

The limits slip. A prompt that asks for a whole report, or allows the model to add context, lets invented facts in at the start, and they survive every later edit because they read well. Keep the rule against new claims in every request.

The edits stop being fed back. A writer who fixes drafts silently gets the same hedges and errors every round. Tell the model what changed and why.

The final pass goes back to the model. One last request for polish can soften the strongest finding back into a maybe or reintroduce a number the writer corrected. Once the human pass begins, nothing goes back through the model.

When to use something else

  • Pyramid Principle. When the hard part is the structure of the argument, which has to be settled before anyone drafts.
  • Red Teaming. When the conclusion needs a determined adversary from outside the team, not a drafting partner.
  • SCQA. When the piece is short and the job is to frame the situation, complication, question, and answer.

Where it sits in SOLVE eX

  • Any step: Tools that serve every step: habits of thought and checks on the thinking itself.

Go to the source

Back to the catalog