Skip to content

November 15, 2026 / Issue 91 / 5 min read

Wharton tested three ways to write with ChatGPT. People who only approved its work did worse than with no AI at all. Keep the idea and the outline yourself

The daily briefing video for this issue, 4:29.
Subscribe on YouTube for a daily briefing video(opens in a new tab)

A study published Tuesday in PNAS Nexus tested a question most AI rollouts skip: which part of the job should the person keep? Kartik Hosanagar of the Wharton School and Daehwan Ahn of the University of Georgia(opens in a new tab) had 285 students write a 1,000-word story on their own, then write another with ChatGPT under one of three setups. In "Human Confirmation," the AI came up with the idea, outlined, drafted and edited, and the student only approved or rejected each step. In "Human Creativity," the student picked the idea and wrote the outline, and the AI drafted and edited. In "Copilot," students used the AI at whatever stage they liked.

Total writing time did not differ significantly between the three setups, so keeping the idea and the outline cost nothing in speed. Quality did differ. Four raters scored every story on a 1 to 7 scale, and the approve-only stories improved less than the Copilot stories, mostly because they were less interesting and less coherent. Students who had written the strongest stories on their own did worse under approve-only. That group also reported less pride in their stories and a drop in how much flexibility they felt they had, and their stories were the most alike of any group. The press release(opens in a new tab) sums it up: approve-only stories were rated below the students' unaided work, while the other two setups beat it.

That matters because many AI deployments are built exactly like the approve-only setup: the model produces, a person signs off, and the sign-off is called keeping a human in the loop. It looks like the cheapest design, since the person does the least. In this test it saved no time over the alternatives and produced the weakest work. Most AI business cases count hours saved. This study suggests the human's place in the work decides what those hours are worth. Uniform output is a cost as well, for anyone whose writing, pitches or product copy has to stand out.

The limits are real. The writers were students, the task was fiction, the model was ChatGPT 3.5, several generations old, and everyone wrote without AI first, so before and after comparisons mix in learning and fatigue. It does not prove the same holds for code, analysis or customer email. But it is a randomized test of how to divide the work, which is rarer and more useful than another survey of how many people use AI.

If you write with AI, or manage people who do, try this on one recurring piece of work this week: a proposal, a product description, a monthly report. Write the main point and a rough outline yourself, in your own words, before you open the AI tool, then let it draft and edit. If you run a team, find any workflow where a person's only job is to approve AI output, and move that person to the start of the work instead of the end.

Action items

A study in PNAS Nexus by Kartik Hosanagar of Wharton and Daehwan Ahn of the University of Georgia had 285 students write stories with ChatGPT under three setups. Those who only approved the AI's ideas, outline and draft produced the weakest stories, with no time saved over setups where the student kept the idea and the outline. The cheapest-looking way to keep a human in the loop, signing off at the end, may be the one that costs you quality.

If you write with AI, or manage people who do, try this on one recurring piece of work this week: a proposal, a product description, a monthly report. Write the main point and a rough outline yourself, in your own words, before you open the AI tool, then let it draft and edit. If you run a team, find any workflow where a person's only job is to approve AI output, and move that person to the start of the work instead of the end.

Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.

Also worth knowing