HESEIA Sentence Bias Dataset

Ethical Review

This professional development course and data collection study was reviewed and approved by the Universidad Nacional de Córdoba (Argentina) and endorsed and run as an official course by the regional Ministry of Education of the Córdoba province. Below we summarize the ethical considerations of the project. Participation in the project was entirely voluntary. Teachers enrolled in the professional development course of their own accord and had the option to engage with the data collection and co-design components based on their interest and institutional support. While no financial compensation was provided, the course was offered free of charge, officially accredited by the Ministry of Education, and provided participants with access to training, meals during in-person classes, pedagogical resources, and university teaching assistants to all teachers. The course offered 36 hours of content of critical perspective on AI for teachers and involved at most two hours of sentence writing for the benchmark dataset according to the lesson plans. The actual cost of the course was 300USD per teacher, which was covered by the sponsors we thank in the acknowledgments. Teachers were free to design the interaction with students as they wanted in a way adapted to the age and to the school subject. Teachers were given three options in order to develop their class. (1) use EDIA, (2) use another tool that better suited their pedagogical approach, or (3) use our unplugged activity. Options (2) and (3) were excluded from our data collection. This approach improved teachers’ agency and creativity. We place high value on the teacher–student relationship; therefore, we emphasized that the topics and activities should align with the interests, needs, and preferences of both teachers and students. Informed consent was obtained from everyone involved. An in-person lecture on data governance, private and sensible information, opt in and opt out was taught during the course, including reading aloud the informed consent and discussing it. The teachers provided feedback on how to adapt the text of the informed consent to an age appropriate version. The researchers in this paper did not interact directly with students. The course offered an unplugged alternative that teachers could use to register the activity on paper instead of digitally. If the school opted for the use of the digital tool EDIA, the data collected could be visualized and deleted. EDIA software was reviewed and approved by the ethics board of the feminist network on AI (FAIR) and it is described in detail in detail in Alonso Alemany et al. (2023). To ensure the protection of personal information, all data collected through this project was pseudo-anonymized and not linked to individual identities, only to optional gender and age information. This process followed the principle of data minimization, which emphasizes collecting and processing only the information strictly necessary to meet the research objectives. This approach aligns with the Argentina National Personal Data Protection Law and the Comprehensive Protection Law for the Rights of Children and Adolescents, both of which informed the ethical design of the study. Computing infrastructure: All computing infrastructure for the PD course software and all experiments were self-hosted with the help of Universidad Nacional de Córdoba (Argentina). The study avoided exposing participants to overtly offensive content. Instead, it focused on fostering critical reflection about language and fairness through the examination of everyday school and life content. However, we are aware that discussions of bias can still evoke discomfort or bring attention to marginalizing experiences. Teachers designed the activities for their classroom context and encouraged reflective discussions within a supportive environment. The course offered open access to all teaching materials, and opportunities for students and teachers for (optional) presentation of their experiences and findings at schools (with or without) university tutors, in the course webpage, and at the university closing class. This project's main goal was to involve teachers and students as critical agents in understanding and questioning the biases embedded in AI technologies.

Intended Use

Bias Benchmark for model evaluation. Specifically on biases and stereotypes annotated by teachers and students from Córdoba - Argentina.