Ciston, S. (2026). Generating the language of AI harms: Mapping guardrails using critical code studies. AI & SOCIETY.

Link: https://doi.org/10.1007/s00146-026-02922-0

Open access: Yes

Notes: Even while companies might like to pretend that they are, the guardrails set up by digital platforms and AI to deal with vioelnt and harmful behaviour are neither neutral nor technical. Rather, they are culturally contingent and value-based—thus showcasing what such companies deem worthy of being communicated. In this paper, Ciston explores how large model guardrails are socially produced and technically enforced through language. The author does this by focusing on three moments: 1) gathering and filtering its models’ pre-training data, 2) developing values and taxonomies, and 3) fine-tuning on models to match its taxonomies. Overall, they found that “the machine learning pipeline is riddled with ideological choke points where quantitative methods rely on qualitative judgments. Concepts and values are still defined, encoded, and executed programmatically, whether or not creators are willing to articulate how they define an idea, make a decision, or from what subject position they operate” (p. 18). In this way, Ciston argues that guardrails are both culturally narrow and technically incapable of dealing with the complexity of harmful and violent behaviour.