Document is current

Article History

Received: 5 December 2025

Accepted: 22 April 2026

First Online: 11 May 2026

Declarations

:

: This research included several studies involving human participants, as well as analyses of publicly accessible facial images. All procedures involving human participants were reviewed and approved by the appropriate Institutional Review Boards/Research Ethics Committees (Research Ethics Committee of the Queensland University of Technology, approval no. 2000000651; Economics and Business Ethics Committee of the University of Amsterdam, approval no. EB-6781; see appendix) and were conducted in accordance with institutional, national, and international ethical guidelines. No research involving animals was conducted.

: Two categories of data were used in this research. First, publicly available Crunchbase images and professional information did not require informed consent because these data are publicly displayed, were analyzed only in de-identified and aggregate form, and involved no direct interaction with individuals. Second, all participants in the human-judgment studies (the expert sample and the trained sample) provided informed consent prior to participation, were informed of the study purpose and procedures, and retained the right to withdraw at any time. No minors participated in the studies, and participation involved no more than minimal risk.

: The AI model was trained and evaluated using 40,728 publicly accessible facial images and professional information from Crunchbase profiles. These data consist of voluntarily posted professional headshots and occupational histories. No direct interaction with individuals occurred, no sensitive personal data were collected, and no identifiable results are reported. All data were de-identified and analyzed in aggregate. Because these data constitute public digital trace information and were used in accordance with institutional guidelines, informed consent from individuals depicted in the images was not required.

: An online classification experiment was conducted with 650 adult entrepreneurship experts who evaluated pairs of anonymized facial images. Participants provided informed consent before beginning the study, were informed of their right to withdraw at any time, and no personally identifying information beyond general demographic characteristics was collected. Classifications were removed whenever participants indicated that they recognized an individual in a displayed image. Ethical approval for this study was granted by the appropriate IRB (Research Ethics Committee of the Queensland University of Technology, approval no. 2000000651).

: A second human-participant experiment assessed whether brief exposure to labeled image examples improved classification accuracy. A total of 133 adult participants enrolled in university entrepreneurship and business courses completed the task. Participation was voluntary, informed consent was obtained, and no identifiable personal information was collected. Training materials consisted of publicly available images subject to the same ethical safeguards described above. Ethical approval for this study was granted by the relevant IRB (Economics and Business Ethics Committee of the University of Amsterdam, approval no. EB-6781).

: Across all components of the research, the authors applied safeguards including de-identification of data, removal of familiar-image classifications, evaluation of gender and race model bias, and clear communication of potential societal implications and ethical risks associated with AI-based facial inference. These measures align with contemporary guidelines for responsible AI research and the ethical use of human-centric data. This focus on high ethical standards is also a subject of the paper (Sect.  ).

: The authors declare no competing interests.