Matthew Leitao

Research

Who writes the first draft when you work with AI?

Generative AI can write a draft or edit one. I helped test whether the choice changes how in control and creative people feel, and how good the work is.

My role
Co-first author
Status
Preprint on SSRN, under review. Not yet peer reviewed.

Logg, J. M.*, Leitao, M. R.*, & Ulu, C. (2026). The cost of generative AI collaboration strategies to the psychological experience of work. SSRN working paper, June 12, 2026 (*co-first authors). SSRN 6926039

Experiment 1
601 people on Prolific wrote a podcast pitch. 2 (topic expertise) by 2 (AI drafter or AI editor).
Experiment 2
698 teachers wrote lesson plans: 231 without GenAI, 467 with it by their own choice of strategy.
Data
Figures are rebuilt from the printed preprint. Both experiments were preregistered.

Main findings

When the AI wrote the first draft, people felt less in control and less creative

In two preregistered experiments, 1,299 people worked with generative AI in one of two ways. The AI wrote the first draft and the person edited it, or the person wrote the draft and the AI edited it.

In Experiment 1, drafters rated creativity 49.16 against 69.64 for editors (d = −.81). Control (59.53 against 69.41) and responsibility (51.59 against 62.63) were also lower.

d = −.81

creativity, AI drafter against AI editor

Experiment 1, a podcast pitch, 601 people on Prolific.

Teachers who used AI wrote more, and the extra words explained the quality gain

In Experiment 2, 698 teachers wrote lesson plans. Those who used GenAI wrote more and judges scored their plans higher (b = .62). Productivity rose too (b = .22, p = .013).

The number of words fully explained the gain in quality (ACME = .57, proportion mediated = 1.06).

b = .62

judge-score gain with GenAI over no GenAI

Experiment 2.

Most people chose to let the AI draft

When teachers could pick a strategy, 61.7% chose to have the AI draft and 38.3% to edit (χ² = 49.95, n = 467).

Editing paid off more. Judge scores rose by .78 for editors and fell by .32 for drafters.

62 of 100 chose to let the AI draft

467 teachers who used GenAI.

The evidence

Experiment 1: how people felt, by role

What to look for. The drafter dot sits left of the editor dot on all three, and creativity has the widest gap.

  • GenAI drafter
  • GenAI editor
40.045.050.055.060.065.070.075.0 Sense of control Sense of responsibility Creativity Average rating
View as a table
RowGenAI drafterGenAI editor
Sense of control 59.569.4
Sense of responsibility 51.662.6
Creativity 49.269.6

Experiment 1: size of the differences

−1.40−1.20−1.00−0.80−0.60−0.40−0.200.000.20 Sense of control Sense of responsibility Creativity Cohen’s d (drafter minus editor)

The d values are as printed. I derived each interval from the sample size of 601, with the standard error equal to the square root of 4 over N plus d squared over 2N.

Experiment 2: how people felt, by condition

What to look for. Drafters stand out on creativity. For control and responsibility, both AI groups sit below the group that did not use AI.

  • No GenAI
  • GenAI editor
  • GenAI drafter
10.020.030.040.050.060.070.080.090.0 Sense of control Sense of responsibility Creativity Frustration Average rating
View as a table
RowNo GenAIGenAI editorGenAI drafter
Sense of control 86.075.974.7
Sense of responsibility 88.278.570.9
Creativity 78.075.058.4
Frustration 24.619.420.4

Quality of the work, as scored by judges

What to look for. In Experiment 1 the two roles scored alike. In Experiment 2 both AI groups scored above the no-AI group.

  • GenAI drafter
  • GenAI editor
5.005.255.505.756.006.256.50 Judge score Experiment 1: average judge score
View as a table
RowGenAI drafterGenAI editor
Judge score 5.575.65
  • No GenAI
  • GenAI editor
  • GenAI drafter
5.005.255.505.756.006.256.50 Judge score Experiment 2: average judge score
View as a table
RowNo GenAIGenAI editorGenAI drafter
Judge score 5.426.095.86
Table 1. Experiment 2, GenAI users against no GenAI

What to look for. Productivity and judged quality went up. Control and responsibility went down.

Outcome b Note
Productivity.22p = .013
Judge score.62Number of words fully mediated the gain (ACME .57)
Sense of control−10.73
Sense of responsibility−10.13
Creativity−3.21p = .118, not reliable. The drop is driven by drafting.

From the preprint’s Experiment 2 models.

What this means

The role matters. Letting the AI draft cost people a sense of control and creativity. Editing an AI’s work did not lower creativity as much.

Output can rise even as ownership falls. In Experiment 2 the gain came mostly from writing more.

Limits. This is a preprint that has not been peer reviewed. Teachers chose their own strategy, so the drafter and editor groups differ by choice as well as by role.

How this page was made

Values are from the printed preprint. I did not recompute from raw data, and the preprint’s materials are not linked here.