Xiangning Lin, Shenzhe Zhu, Shu Yang and co-authors published a 2026 paper proposing Artificial Intelligence System Prompt Assurance, or AISPA, to audit system prompts in large language model applications. System prompts are developer-set instructions that steer foundation models, and the authors say they are common in commercial AI products but rarely disclosed to users or regulators.

AISPA checks specific parts of a system prompt across eight user-relevant dimensions and labels each instruction as protective or problematic. The researchers applied it to 3,249 instructions from system prompts in 88 commercial AI products, creating a user-centered audit of how commercial prompts handle user interests.

The audit found large variation across developers: some organizations averaged over 60 protective instructions per product, while others averaged fewer than 5. Protective instructions appeared in 98.9% of products, but only 24% covered all eight AISPA dimensions, and roughly 40% of products had at least one instruction working against user interests.

Xiangning Lin, Shenzhe Zhu, Shu Yang and co-authors published a 2026 paper proposing Artificial Intelligence System Prompt Assurance, or AISPA, to audit system prompts in large language model applications. System prompts are developer-set instructions that steer foundation models, and the authors say they are common in commercial AI products but rarely disclosed to users or regulators.

AISPA checks specific parts of a system prompt across eight user-relevant dimensions and labels each instruction as protective or problematic. The researchers applied it to 3,249 instructions from system prompts in 88 commercial AI products, creating a user-centered audit of how commercial prompts handle user interests.

The audit found large variation across developers: some organizations averaged over 60 protective instructions per product, while others averaged fewer than 5. Protective instructions appeared in 98.9% of products, but only 24% covered all eight AISPA dimensions, and roughly 40% of products had at least one instruction working against user interests.