Øystein Skaar

Øystein Skaar

MishMash role: Member · WP3 , WP4 , WP7

MishMash Projects

Latest results

Type

Book chapter

  2. Book chapter, 2023

    Digital kompetanse, vaksinemotstand, KI og pandemisk risikopersepsjon

    Rune Johan Krumsvik ; Øystein Olav Skaar

Journal article

  1. Journal article, 2026

    The Gentler Researcher: Examining Research Bias and Gender Stereotypes in Sikt AI Chat

    Open access CC BY Read it

    Øystein Olav Skaar ; Caroline Miriam Skaar

    Studies on large language models (LLMs) have shown that these technologies can reproduce stereotype-based biases present in their training data. In this paper, we hypothesized that an OpenAI-based model would more frequently attribute first authorship to women in masculine-typed research fields than to men in feminine-typed fields, an asymmetry consistent with gender-role beliefs. Across three studies, we analyzed how the model’s generative outputs reflect gendered associations. Study 1 replicated Fulgu and Capraro’s phrase-association paradigm, replicating the study’s findings that feminine-stereotyped phrases were more consistently linked to female characters, whereas masculine-stereotyped phrases were less reliably linked to male characters. The results suggest that the training data contains a stronger and more stable association between femininity and communal traits than between masculinity and agency. Study 2 examined 4,585 randomly generated hypothetical studies and identified a substantial skew toward female first authors (>82%), with male authors overrepresented primarily in engineering, technology, and theoretical contexts. Study 3 extended these findings across nine specified research fields, where the results of the first stage were consistent with Study 2. When prompted to classify rather than generate the gender of the first author, the model produced a lower overall proportion of women, but retained the same relative distribution across fields observed at the first stage. Taken together, the results indicate that the model’s outputs correspond more to feminine than masculine gender-role beliefs. While this tendency may appear to increase women’s representation in masculine-typed domains, it does not afford men equivalent visibility in feminine-typed domains. These results underscore how LLMs, as statistical systems shaped by training corpora and reinforcement learning, can reflect and reproduce cultural and historical asymmetries in gender representation, challenging assumptions of epistemic neutrality.

Professional article

Conference

  1. Conference poster, 2023

    DigitalInnSights- To play or not to play?

    Odd Rune Stalheim ; Oda Julie Hembre ; Øystein Olav Skaar ; Eloisa Michaelsen

More results in NVA…