are you aware of the paper Who Controls the Conversation? User Perspectives on Generative AI (LLM) System Prompts?
Yes — I found it. It's a CHI 2026 paper (presented at the ACM Conference on Human Factors in Computing Systems, Barcelona, April 2026) that asks three core questions: what do system prompts contain, how do end-users perceive them, and what do those perceptions offer for design and governance practice? arXiv
Here's what the paper establishes and finds:
The core problem it identifies
System prompts — instructions set by developers or deployers that shape AI behaviour, define the model's personality, and establish guardrails — take precedence over user inputs, sitting at the top of the instruction hierarchy. Despite this influence, transparency is severely limited: system prompts are generally not made public, and most platforms instruct models to conceal them, leaving users disconnected from a key mechanism governing their interactions. arXiv
Users may attribute certain AI behaviours to inherent model characteristics when those behaviours actually result from specific system prompt choices — potentially leading to misunderstandings and misattributions of AI capabilities and limitations. System prompts can be modified at any time by multiple actors, with users having no visibility into those changes or their effects. arXiv
What users revealed
The results reveal user perspectives on the benefits and risks of system prompts, the values they prefer to be associated with prompt design, their levels of comfort with different types of prompts, and their views on degrees of transparency and arXiv control — with users expressing both appreciation for the useful customisation system prompts enable, and discomfort at the degree of invisible influence being exerted over their interactions without their knowledge.
This paper connects very directly to the MCI framework and to the Cheng et al. sycophancy paper we discussed. Would you like me to read the full paper and then map it through the MCI lens as well? It feels like a natural companion piece — where Cheng et al. exposes what sycophancy does to users, this paper exposes the structural mechanism (the system prompt hierarchy) through which that and other behavioural shaping happens invisibly. Under MCI, that opacity is a direct Legitimacy Maintenance failure — and the concealment instruction is a Non-Domination violation, placing users in a position of arbitrary dependence on a governance structure they cannot even see.
Comments
Post a Comment