16,392 rated or chosen profiles from 683 respondents, in up to 24 tasks of 1 profile varying on 7 attributes.
| Description | 683 Dutch respondents (panel, Jan 2025) rated the trustworthiness of 24 news articles each (8 per topic x 3 topics), each shown with an 'AI Monitor' table stating whether six journalistic tasks were done by a human or an AI and whether a human was in the loop; one rating per profile, 1-7. |
|---|---|
| Reference | Mattis, N. M., Kieslich, K., & de Vreese, C. (2025). Feeling iffy about generative AI: When journalists disclose AI use, trust in news is lower [Preprint]. OSF Preprints. |
| DOI | 10.31219/osf.io/tmzq4_v1 |
| Licence | CC BY 4.0 |
| Source data | https://doi.org/10.17605/osf.io/76n2d |
These notes describe the source data, which the IRW reproduces as released. All notes are listed on the Data Notes page.
| Respondents | 683 |
|---|---|
| Rows (one per respondent, task and profile) | 16,392 |
| Tasks per respondent (most) | 24 |
| Profiles per task | 1 |
| Attributes | 7 |
| Topic | technology ai privacy, political communication |
|---|---|
| A profile represents | message |
| Country | NL |
| Language shown | nl |
| Language of stored levels | en |
| Randomization restrictions | yes (stated by the source): Fixed 8-run fractional factorial of the 7 binary attributes: only 8 of 128 combinations occur, the same 8 for every respondent and topic; the randomised parts are the order of topics, the order of profiles within a topic and which article of the topic is attached to which design row. |
| Task order | unknown |
| Profile position | inferred from row order |
Because the attribute levels were not randomized independently, an estimator has to account for the restriction; a plain difference in means across levels can mislead. Task order or profile position was rebuilt from row order; analyses of position or task order should leave it out.
Outcomes:
| rating | Rating of the perceived trustworthiness of the article, from 1 (not at all trustworthy) to 7 (extremely trustworthy) (paraphrase) [rating, 1 (not at all trustworthy) to 7 (extremely trustworthy)] |
|---|
attr_article_writingattr_background_researchattr_fact_checkingattr_human_in_the_loopattr_idea_generationattr_image_generationattr_proofreadingcov_agecov_education_codecov_gendercov_political_position_codeidprofileratingtasktrial_article_imagetrial_topic
The layout (one row per respondent, task and profile; choice/rating outcomes; attr_ columns holding each attribute level as displayed) is defined in the IRW conjoint standard. What the outcomes asked is under Design above.
Or load it directly in R or Python (both need a free Redivis account; the first call opens a browser to sign in):
# R
install.packages("remotes")
remotes::install_github("itemresponsewarehouse/Rpkg")
library(irw)
df <- irw_fetch("mattis_2025_ai_disclosure", source = "conj")
# Python
pip install irw
import irw
df = irw.fetch("mattis_2025_ai_disclosure", source="conj")
@article{Mattis_2025, title={Feeling iffy about generative AI: When journalists disclose AI use, trust in news is lower}, url={http://dx.doi.org/10.31219/osf.io/tmzq4_v1}, DOI={10.31219/osf.io/tmzq4_v1}, publisher={Center for Open Science}, author={Mattis, Nicolas Michael and Kieslich, Kimon and de Vreese, Claes}, year={2025}, month=Apr }
| IRW version | v584 |
|---|---|
| Redivis dataset | irw_conjoint_2 v1.4 |
| Redivis dataset DOI | 10.57761/hk97-wv30 |
| Manifest pin for this IRW version | v1.4 |
| Metadata source | irw_meta v44.0 |