Positive attitudes toward AI narrowed that gap and weakened a credibility difference between male and female authors.
Researchers randomly assigned participants to read a technical text about large language models attributed to either a male or female human author or a male or female AI author. Participants rated the author’s competence and the credibility of the text, and also reported their attitudes toward AI.
Human authors received higher competence ratings than AI authors, while author gender did not affect competence ratings. Among people skeptical of AI, texts attributed to male authors were rated as more credible than texts attributed to female authors; this difference became smaller among people with more positive attitudes toward AI.
How authors were judged
The study included 219 participants in a 2×2 design that varied author type—human or AI—and author gender—female or male. Participants attributed greater competence to human authors than to AI authors. This difference was smaller among participants who viewed AI more positively.
Author gender did not affect perceived competence. For message credibility, however, AI-skeptical participants rated texts attributed to male authors as more credible than those attributed to female authors. The gender difference decreased as attitudes toward AI became more positive. The abstract does not report a general author-type difference in message credibility.
Why the judgments matter
The findings show that people’s judgments of technical information are not based only on the text itself. Perceptions of the author—whether the author is human or AI, and in some cases whether the author is presented as male or female—also matter.
The results also indicate that attitudes toward AI shape these evaluations: people who view AI more positively show a smaller human advantage in perceived competence and a weaker gender difference in perceived credibility. This is relevant when information is presented as coming from a person or an AI system.
Evidence and caveats
This was a randomized study with 219 participants, using four combinations of author type and gender. All participants read the same technical text about large language models, which helped the researchers compare how the stated author affected ratings.
The evidence comes from participants’ ratings in one study and one technical-information setting. The abstract does not specify the participants’ demographic makeup or whether the findings apply to other subjects, types of messages or real-world decisions. The study measured perceived competence and credibility rather than the objective quality or accuracy of the text.
// Source
Social Sciences & Humanities Open · 2026 · DOI: 10.1016/j.ssaho.2026.103339
Authors: Joachim Kimmerle, Annalena Ulsperger, Ulrike Cress
Institutions: University of Tübingen, Leibniz-Institut für Wissensmedien
