Read original
hf-paperspapers86

AISPA: User-Centric System Prompt Auditing for Large Language Model Applications

AI Summary

The paper introduces Artificial Intelligence System Prompt Assurance (AISPA), a user-centric framework for auditing system prompts across eight user-relevant dimensions. The authors review 3,249 instructions from 88 commercial AI products, labeling instructions as protective or problematic. They report substantial variation among developers: some average more than 60 protective instructions per product, while others average fewer than five. Although 98.9% of products contain at least one protective instruction, only 24% cover all eight dimensions. System prompts have also become longer and more user-protective over time. However, about 40% of products contain at least one instruction that may work against user interests.

Why it's worth reading

System prompts increasingly shape user rights and product behavior while remaining opaque. AISPA offers a concrete audit framework and quantitative evidence that protective controls can coexist with instructions that undermine user interests.

Deep Read

1. What happened

Original facts: The paper introduces Artificial Intelligence System Prompt Assurance (AISPA), a framework for auditing system prompts from a user-centered perspective. It applies the framework to 3,249 instructions from 88 commercial AI products.

2. Core technology

Original facts: AISPA examines individual parts of a system prompt across eight dimensions considered relevant to users. Each instruction is classified as protective or problematic. Analysis: This converts normally hidden product rules into structured audit objects that can be compared across products and developers.

3. Key evidence and numbers

Original facts: Some organizations average more than 60 protective instructions per product, while others average fewer than five. At least one protective instruction appears in 98.9% of products, but only 24% cover all eight dimensions. Approximately 40% contain at least one instruction that works against user interests. The paper also reports that system prompts have become longer and more protective over time. Not independently verified: These figures come from the supplied abstract. The sampling procedure, annotation agreement, and statistical details were not checked against the full paper.

4. Why it matters

Analysis: System prompts can define disclosure boundaries, refusal behavior, commercial priorities, and instruction hierarchy. Measuring only whether a product has any safety-oriented instruction may therefore miss gaps and conflicts. AISPA’s emphasis on coverage and problematic instructions makes those hidden differences more visible.

5. Practical impact

Analysis: Product teams could use the eight dimensions as a prompt-review checklist. Regulators and enterprise buyers could request audited coverage information, while independent researchers could compare applications using a shared classification scheme. The framework may also help prioritize prompt fragments for human review.

6. Limitations and uncertainty

Original facts: The abstract does not specify the eight dimensions, how products were sampled, how instructions were extracted, or how problematic classifications were validated. Analysis: System prompts may change by version, region, account tier, and context, so a static audit may not represent long-term product behavior. Prompt text alone also cannot establish that a model will follow an instruction or that it causes a particular user outcome. Unverified inference: The reported share of problematic instructions may depend on the researchers’ definitions, the visibility of prompts, and the product mix; it should not automatically be treated as a market-wide estimate.

7. Original sources

Tags

系统提示词AI审计AI治理透明度用户权益商业AI安全评估