Who Controls the Conversation? User Perspectives On Generative AI (LLM) System Prompts
A strong and timely CHI contribution: it turns system prompts from an obscure backend mechanism into a user-facing design and governance issue, and backs that move with both a real-world prompt taxonomy and survey evidence that users are not as unaware or passive as industry practice assumes.
Axes Lens
Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.
Contribution shape
- Knowledge form
- descriptive knowledge typical · 92/268
- Novelty type
- framework typical · 59/268
- Abstraction level
- field typical · 41/268
- Generalization target
- user population typical · 75/268
- Validation mode
- mixed methods typical · 136/268
Evidence profile
- Evidence strength
- strong typical · 158/268
- Claim alignment
- strong typical · 231/268
- Overclaim risk
- medium typical · 210/268
Review Summary
This paper is compelling because it reframes system prompts as something more than hidden implementation detail. In mainstream practice, prompts are usually treated as confidential developer or deployer instructions, with user exclusion taken for granted. The authors show why that framing is incomplete. They first map what system prompts actually contain in the wild, using public official and community sources plus manual and computational analysis to derive a seven-topic taxonomy. They then connect that taxonomy to user perspectives through a survey of 109 participants. That combination is the paper's real strength: it does not only speculate normatively about transparency and control, but grounds those questions in observed prompt-design practice and then asks users how they feel about those practices. The most interesting result is not simply that users want more transparency. It is that users are already somewhat aware of system prompts, can articulate both benefits and risks, and differentiate among prompt topics rather than treating all prompting as equally acceptable. That matters for HCI because it suggests prompt governance should be designed as a nuanced interaction and accountability problem, not a binary reveal-or-hide decision. The paper also productively surfaces tensions: users value safety and reliability, but also worry about embedded bias, hidden agendas, and organizational value imposition. Those tensions make the work useful for both design and governance communities. Its limits are also clear. The taxonomy is based on public prompts rather than confidential production prompts, and the survey captures stated preferences rather than real use of transparency or control tools. So the paper should be read as foundational descriptive work, not as a final answer on deployment design. Even so, it is a significant contribution because it opens a concrete research agenda around stakeholder-informed prompt design, disclosure practices, and ecosystem governance for a mechanism that increasingly shapes AI behaviour.
What Changed
Canon before
System prompts typically operate as confidential developer or deployer instructions that govern the behaviour of LLMs, taking precedence over user inputs and largely hidden from end-users. The dominant assumption is that system prompts are technical guardrails or safety layers managed by organizations, with minimal user awareness or involvement in their design or control. Existing transparency practices in AI vary widely, with most providers maintaining confidentiality and instructing models not to disclose prompts. User perspectives on prompts, their content, and governance mechanisms have been underexplored.
Departure from common sense
The paper challenges the expectation that end-users are mostly unaware of system prompts and indifferent to them. Instead, it shows that most surveyed participants either knew of or suspected system prompts, recognized both benefits and risks, and expressed interest in transparency and meaningful influence over these hidden instructions.
Actual novelty
The paper's main novelty is a grounded empirical framing of system prompts as both a design object and a governance object. It combines a taxonomy of seven recurring system-prompt topics derived from real-world prompt corpora with a survey of 109 users about awareness, comfort, values, transparency, and control preferences, thereby linking prompt-design practice to user-centred expectations.
Evidence
Evidence comes from a two-part mixed-methods study. First, the authors compile prompts from official releases and community repositories, then use manual coding plus computational concept induction to derive and validate a seven-topic taxonomy. Second, they run a survey with 109 participants to measure awareness, perceived benefits and risks, comfort with prompt topics, and preferences for transparency and control. This supports descriptive claims well, while broader governance implications remain more interpretive.
“ To gain insight into user perspectives on such, we conducted an empirical survey study (N = 109), building on our analysis of current practice”
actual novelty · 3 Methodology · confidence 0.96
“9% of respondents were fully unaware that system prompts exist. Roughly the same percentages knew of (43.1%) or suspected such a mechanism (45%) (see Figure 8). The vast majority of participants (98.2%) acknowledged potential benefits of system prompts, with only two participants (1.”
departure from common sense · 6.2 Participants are aware of system prompts, and identify broad risks and benefits · confidence 0.95
“We first sought to understand what topics and approaches appear in real-world system prompts across different AI deployments. Towards this, we compiled a dataset from multiple public sources, including officially released prompts and ones from community repositories”
limitation · 3 Methodology · confidence 0.90
“To investigate the nature of system prompts and user perceptions on such prompts, we employed a two-part mixed-methods approach that combines computational text analysis with empirical user research.”
validation scope · 3 Methodology · confidence 0.97
Limits
Method limits
The prompt taxonomy is derived from publicly available official and community-sourced prompts, so it may not capture confidential production prompts used in proprietary deployments. The survey measures stated perceptions and preferences rather than observed behaviour, and the framing materials may shape participant responses. The domain is also changing quickly, so prompt practices and user expectations may shift.
Deployment limits
The findings are most applicable to LLM systems that use system-prompt style control and to users similar to the surveyed population. They do not directly establish how transparency or control interfaces would work in deployed products, nor whether users would exercise such controls effectively in practice.
Boundary conditions
The contribution is strongest as a descriptive account of current prompt topics and surveyed user attitudes in the studied context. It should not be read as proving that all users want the same degree of disclosure or control, or that the identified taxonomy exhausts all prompt practices across future models, languages, or organizational settings.
Position in field
This paper helps establish system prompts as a legitimate HCI and governance research object rather than merely a hidden technical implementation detail. Its contribution is important because it connects real prompt practices to user expectations, showing that prompt design has implications for transparency, accountability, trust, and agency. That makes it a useful foundation for later work on participatory prompt design, disclosure mechanisms, and governance standards.