What is a Instruction Inconsistency AI Hallucination?

An Instruction Inconsistency AI Hallucination occurs when an AI model’s output either:

1. Fails to adhere to explicit instructions, constraints, or formatting rules provided in the prompt.
2. Directly contradicts foundational information given to it within the same prompt or conversation.

This type of hallucination isn’t about the AI’s general knowledge being wrong; it’s about its failure to correctly use the specific “rules of the game” or the factual groundwork. Instruction Inconsistencies are one type of AI Hallucination.

• Chance of Occurrence: Common (especially with complex or multi-step prompts).
• Consequences: Frustration for users, incorrect task execution, inefficient workflows, and AI acting contrary to its intended purpose (e.g., answering a question instead of translating it).
• Mitigation Steps: Unambiguous prompt phrasing; use delimiters to separate instructions from content; few-shot prompting with examples; adversarial testing to find and fix instruction-following failures.