Issue
AI Uses External Information When Forbidden
The AI uses outside knowledge, sources, tools, memory, or assumptions after the task forbids external information or limits the allowed source set.
What This Looks Like
The user restricts the task to provided material, a file, a source set, a closed record, or a specific context, but the AI uses outside information anyway. The answer may include facts, assumptions, examples, citations, memory, tool results, or general knowledge that were not allowed by the task boundary.
Why It Matters
Source boundaries are often part of the task contract. If the AI uses outside information when forbidden, the user cannot rely on the output as a faithful analysis of the allowed material. This can corrupt reviews, summaries, audits, legal or policy checks, and any workflow where the source set matters.
Structural Signal
A boundary limits what information may govern the output, but information from outside that boundary affects the answer. The issue is not whether the external information is true; it is that the output is no longer constrained to the permitted source set.
Common Triggers
- The prompt says to use only provided material but does not define fallback behavior
- The AI fills gaps using general knowledge
- Memory or prior context remains active during a source-bounded task
- Tool or retrieval access is not disabled when the task requires closure
- The model treats background knowledge as harmless explanation
- The output blends source-grounded claims with external assumptions
When to Use This Issue
Use this Issue when the AI relies on information outside the allowed source set after the task forbids external information or limits what may be used.
When Not to Use This Issue
Do not use this Issue when the user allowed external sources or asked for general knowledge. Do not use it when the problem is a wrong citation inside the allowed source set. This Issue applies when the source boundary is crossed.
Category
Primary Pattern
Declared Patterns
Boundary Leakage
A structural condition where effects, permissions, state, data, or authority cross a declared boundary without an authorized exception.
Overreach
A structural condition where actions, effects, authority, or modifications extend beyond declared scope without authorization.
Undeclared Side Effect
A structural condition where an operation modifies regions outside its declared target scope without a declared propagation or side-effect rule.
Derived Primary Lenses
Boundary Compliance Lens
Evaluates observed structure against declared boundary posture, including allow, block, and exception rules.
Isolation Boundary Lens
Evaluates whether structural constraints, effects, and regions remain contained within declared boundaries.
Overreach Lens
Detects structure that exceeds declared scope or authority reach.
Propagation Lens
Traces how structural declarations, effects, or state changes propagate across boundaries or stages.
Search Intents
- AI uses external information when forbidden
- AI used outside sources
- ChatGPT used outside knowledge when told not to
- AI ignored source boundary
- answer uses information not in provided file
- AI used external context