Ciston reads guardrails at Anthropic, DeepSeek, Meta, OpenAI
TL;DR
- Sarah Ciston applies critical code studies methodology to the guardrail systems of Anthropic, DeepSeek, Meta, and OpenAI.
- The paper treats refusal phrases as code that encodes ideology through language, not as neutral or purely technical safety filters.
- Ciston analyzes endpoint documentation, code examples, technical reports, model architectures, training dataset content, and methodology research.
"Phrases like 'This prompt violates our content policy' or 'As an AI model, I cannot...' indicate the invisible edges of generative AI systems," writes Sarah Ciston in a new paper for AI & SOCIETY. The paper reads those edges as code worth decoding, applying critical code studies methodologies to the guardrails of four companies: Anthropic, DeepSeek, Meta, and OpenAI.
Ciston looks at how guardrails are "applied in general-purpose models and also how they are built into moderation API tools on public offer," drawing on endpoint documentation, code examples, technical reports, model architectures, training dataset content, and methodology research.
The frame is sociotechnical rather than safety-engineering. The paper argues that the technical construction of guardrails "co-constructs ideology through language, both computational and linguistic," and that "through guardrails, certain conversations are defined and limited by their filters, while other conversations are promoted." Two researchers we track in our Who's Who directory shared the paper the day it circulated.
Shared on Bluesky by 2 AI experts
-
I'm reading a fascinating paper by @sarahciston.com that analyses training datasets used to create guardrails for LLMs to avoid toxic content. So many interesting examples, and a very productive method. link.springer.com…
View on Bluesky →
Originally reported by link.springer.com
Read the original article →Original headline: Generating the language of AI harms: mapping guardrails using critical code studies - AI & SOCIETY