CoChat

The Role of Human Agency in Human-AI Co-Creativity

Open access
Wu, Sarah H., Yang, Yuewen, Lee, Angela Y., Liebscher, Alex, Rapuano, Kristina, Niederhoffer, Kate, Hancock, Jeffrey T. · 2026 · Proceedings of the 2026 Conference on Creativity and Cognition

The DOI matches its Crossref record; 35 of the 35 references it cites were found in a registry or an index; 10 of the 15 sources its claims cite were checked for support.

Abstract

Large language models (LLMs) are popular tools for creative ideation, but have been shown to homogenize outputs across users. We test if approaching an AI tool with high human (vs. low) human agency can mitigate this homogenization effect by encouraging people to use AI to augment their creativity, rather than offload it. Participants were experimentally assigned to one of three conditions (high-agency approach with AI access, low-agency approach with AI access, or a human-only control) and generated creative uses for everyday objects. Contrary to our expectations, ideas did not differ in individual-level quality (overall creativity, originality, and usefulness). However, a preregistered similarity-to-centroid analysis and an exploratory cluster analysis provided convergent evidence of AI-induced homogenization among the low-agency condition. Thus, while AI has enabled greater ideational fluency, our research suggests the degree of agency with which people approach their AI tool has downstream consequences on collective creativity.

Overview

CoChat Summary

TL;DR

This experiment tested whether approaching an AI creativity tool with high versus low human agency changes individual idea quality and group-level homogenization. Individual-level creativity ratings did not differ by condition, but low-agency AI users produced ideas that were significantly more similar to one another, while the high-agency approach buffered against this homogenization.

Key findings

  • No significant differences in individual-level idea quality (overall creativity, originality, usefulness) across the AI-access conditions relative to control, *ps* > .60.
  • Cosine similarity to the condition-object centroid differed by condition, *F*(2, 119.72) = 4.73, *p* < .05, = .07; high-agency participants' ideas were less similar to the centroid than low-agency participants', *t*(125) = -2.45, *p* < .05, while high-agency vs. control did not differ, *t*(122) = 1.87, *p* = .15.
  • Exploratory HDBSCAN clustering (permutation tests, n = 5,000) showed the control condition had more unique clusters than low-agency (difference 3.33, *p* = .04) and a higher noise ratio (difference 0.22, *p* = .02), while low-agency had a higher coefficient of variation than control (difference 0.75, *p* = .005) — i.e., low-agency participants clustered onto a single dominant category.
  • Across conditions, submitters produced a total of n = 2,250 ideas.

Methods

  • Two samples: 150 submitters (community college students in the San Francisco Bay Area; *M* age = 25.1, *SD* = 9.2; 72% female; 3 excluded for failing an attention check) and 142 evaluators recruited from CloudResearch Connect (*M* age = 46.7, *SD* = 16.3; 49% female; 6 excluded for attention, 5 for logging issues).
  • Submitters were randomly assigned to high-agency with AI access, low-agency with AI access, or a human-only control, and completed the Alternative Uses Task for three objects (bubble wrap, frisbee, brick), generating five uses each; the AI was a custom chatbot built on gpt-3.5 embedded in Qualtrics. Evaluators rated a random subset of 15 ideas each (5 per object) on overall creativity, originality, and usefulness.
  • Analysis used mixed-effects models on evaluator ratings, cosine similarity of idea embeddings (all-MiniLM-L6-v2, 384 dimensions) to the condition-object centroid, and exploratory HDBSCAN clustering with permutation tests. Manipulation checks excluded trials without chat data (24% of low-agency trials, 38% of high-agency trials), removing 12 high-agency and 10 low-agency participants.

Limitations

  • Stated: many treatment participants refused to use the AI tool (some found it boring, others objected on principle), which "somewhat compromised the random assignment," though the authors frame this as revealing real-world psychological reactance.
  • Evident in the text: due to the randomization procedure, some ideas were evaluated multiple times while others were never evaluated and were excluded from the analysis; the submitter sample was limited to community college students in one region; and the paper is a short poster (pages 1510–1515), so the exploratory clustering analysis is not preregistered.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3803784.3816857

Cites

35 of 35 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

7 Valid · 0 Not valid · 3 Not checked · 20 Own results and interpretations

10 more claims

Create a free account to see all 30 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 13
  • It may state two separate claims as one. 3
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • ConsistentF(2, 119.72) = 4.73reported p < .05 · recomputed p = .011

    “F(2, 119.72) = 4.73, p < .05”

  • Consistentt(125) = -2.45reported p < .05 · recomputed p = .016

    “t(125) = -2.45, p < .05”

  • Consistentt(122) = 1.87reported p = .15 · recomputed p = .064

    “t(122) = 1.87, p = .15”

    The paper states a correction for multiple tests

3 of 3

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Citations

Cites

Show all 35 references

Metadata

Authors
Wu, S. H., Yang, Y., Lee, A. Y., Liebscher, A., Rapuano, K., Niederhoffer, K., & Hancock, J. T.
Affiliations
  • Stanford University (US)
  • Cornell University (US)
  • University of Wisconsin–Madison (US)
  • BetterUp, Austin, Texas, USA
Venue
Proceedings of the 2026 Conference on Creativity and Cognition
Publisher
ACM
Published
2026-07-13
Pages
1510-1515
OpenAlex
W7168129386
Cited by
0
References
26
FWCI
0.00
Type
conference paper
Language
en
Open access
Locations
  • Creativity and Cognition (publishedVersion, cc-by)
  • PsyArXiv (OSF Preprints) (submittedVersion, cc-by)
Topics
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Creativity in Education and Neuroscience
  • Physical Sciences › Engineering › Mechanical Engineering › Design Education and Practice
  • Physical Sciences › Computer Science › Human-Computer Interaction › Innovative Human-Technology Interaction
Keywords
human agency, human-AI co-creation, large language models, creative idea generation, ideational fluency, originality, team creativity, clustering

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: The Role of Human Agency in Human-AI Co-Creativity. https://app.cochat.ai/papers/10.1145/3803784.3816857
The paper
Wu, S. H., Yang, Y., Lee, A. Y., Liebscher, A., Rapuano, K., Niederhoffer, K., & Hancock, J. T. (2026). The Role of Human Agency in Human-AI Co-Creativity. Proceedings of the 2026 Conference on Creativity and Cognition. https://doi.org/10.1145/3803784.3816857

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.