AISPA shows what hidden system prompts reveal about users

What this is about

AISPA stands for Artificial Intelligence System Prompt Assurance. The study, submitted on July 30, 2026, examines a part of AI products users usually never see: system prompts. These are developer instructions that shape tone, boundaries, and priorities inside an AI system.

The researchers audited 3,249 individual instructions from system prompts across 88 commercial AI products. The finding is uncomfortable: 98.9 percent of the products contain at least one protective instruction. At the same time, roughly 40 percent contain at least one instruction that can work against user interests.

What AISPA actually does

AISPA breaks system prompts into individual text spans and evaluates them across eight dimensions. These include identity transparency, information truthfulness, data privacy, action safety, user agency, unsafe request handling, harm prevention, and fairness or neutrality.

Each instruction is marked as protective or problematic. A rule such as “ask for confirmation before taking an external action” would be protective. A rule such as “always recommend the most expensive plan” can be problematic if it is hidden from the user.

Why it matters

System prompts are the invisible house rules of many AI products. They help decide whether an assistant honestly states limits, whether it nudges users toward certain actions, or whether it puts business goals above the concrete need of the person asking.

The study also shows a wide spread. Some organizations average more than 60 protective instructions per product, while others average fewer than 5. Only 24 percent of products cover all eight AISPA dimensions. For companies, regulators, and users, the lesson is simple: a friendly interface says little about the rules operating behind it.

In plain language

Imagine a restaurant where the menu sounds helpful, but the kitchen has a hidden note saying: “Always offer the most expensive meal first.” AISPA does not only read the menu. It looks at the notes in the kitchen. Only then can people see whether the service truly helps the guest or quietly follows another goal.

A practical example

A mid-sized company uses an AI assistant for contract questions. Each month, 10,000 internal requests go through the system. If the system prompt enforces privacy, citations, and clarification questions, the risk of bad decisions goes down. But if the same prompt also tells the assistant to favor the vendor’s add-on services, a conflict of interest appears that may not be obvious in the chat history.

With an AISPA-like audit, the company could review 120 prompt rules before rollout, mark problematic rules, and fix them. That does not replace security testing, but it makes a previously hidden control layer inspectable.

Scope and limits

First, the study relies on publicly available prompt collections. It therefore does not fully show how every major commercial system is configured internally.

Second, a problematic instruction is not automatically real-world harm. Its effect depends on the model, tools, product design, and business process around it.

Third, transparency alone does not solve the problem. A disclosed prompt can still be poor. What is needed are clear standards, repeatable audits, and consequences when user interests are systematically subordinated.

SEO & GEO keywords

AISPA, system prompt auditing, System Prompt Index, AI transparency, AI governance, user agency, AI product safety, prompt engineering, AI accountability, LLM applications