![[2602.18483] Red Teaming LLMs as Socio-Technical Practice: From Exploration and Data Creation to Evaluation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18483] Red Teaming LLMs as Socio-Technical Practice: From Exploration and Data Creation to Evaluation
Summary
The article examines red teaming as a socio-technical practice in evaluating large language models (LLMs), highlighting the importance of dataset creation and evaluation in ensuring AI safety.
Why It Matters
Understanding red teaming in the context of LLMs is crucial for improving AI safety and reliability. This study sheds light on the socio-technical aspects that influence dataset creation, which is essential for assessing potential harms and enhancing model evaluations.
Key Takeaways
- Red teaming is essential for evaluating the safety of generative AI models.
- Current practices often overlook socio-technical factors in dataset creation.
- Empirical evidence from practitioners reveals gaps in risk conceptualization.
- Adversarial datasets play a critical role in assessing model performance.
- Opportunities exist for HCI researchers to enhance red teaming methodologies.
Computer Science > Computers and Society arXiv:2602.18483 (cs) [Submitted on 10 Feb 2026] Title:Red Teaming LLMs as Socio-Technical Practice: From Exploration and Data Creation to Evaluation Authors:Adriana Alvarado Garcia, Ruyuan Wan, Ozioma C. Oguine, Karla Badillo-Urquiola View a PDF of the paper titled Red Teaming LLMs as Socio-Technical Practice: From Exploration and Data Creation to Evaluation, by Adriana Alvarado Garcia and 3 other authors View PDF HTML (experimental) Abstract:Recently, red teaming, with roots in security, has become a key evaluative approach to ensure the safety and reliability of Generative Artificial Intelligence. However, most existing work emphasizes technical benchmarks and attack success rates, leaving the socio-technical practices of how red teaming datasets are defined, created, and evaluated under-examined. Drawing on 22 interviews with practitioners who design and evaluate red teaming datasets, we examine the data practices and standards that underpin this work. Because adversarial datasets determine the scope and accuracy of model evaluations, they are critical artifacts for assessing potential harms from large language models. Our contributions are first, empirical evidence of practitioners conceptualizing red teaming and developing and evaluating red teaming datasets. Second, we reflect on how practitioners' conceptualization of risk leads to overlooking the context, interaction type, and user specificity. We conclude with three opport...
