Hallucination Mitigation

Hallucination is not random. It follows patterns. The model fabricates most readily when it is asked for specific facts that are not in its training data, when it is asked for citations it cannot verify, when it is asked to fill in details that feel like they should exist, and when vague prompts invite elaboration. Each of these triggers has a corresponding prompt design that reduces it. Hallucination mitigation is a constraint engineering problem.

1. Strategy 1: Retrieval grounding "Answer using only the information in the document below. Do not use knowledge from your training data." This is the primary mitigation for factual hallucination. It restricts the model's source to a controlled, verifiable document. Without this constraint, the model blends your document with its training data — seamlessly and invisibly.
2. Strategy 2: Citation requirement "For every factual claim, cite the specific section, paragraph, or sentence in the source document that supports it." When the model must cite a source for each claim, it cannot easily fabricate — fabricated claims cannot be cited against real source material. This constraint does not eliminate hallucination but makes it far more detectable.
3. Strategy 3: Unknown acknowledgment "If the answer is not available in the provided source, respond with: This information is not in the provided document. Do not approximate or extrapolate." This replaces the model's default behavior (fill in with plausible-sounding content) with a specific instruction to report absence.
4. Strategy 4: Role constraints "You are a researcher who reports only what can be directly verified. You do not speculate, extrapolate, or infer beyond the available evidence." A well-designed expert persona that includes epistemic constraints reduces the frequency with which the model adds information that was not requested.
5. Strategy 5: Scope restriction "Address only the topics explicitly raised in the following questions. Do not expand the scope of your response beyond what is asked." Open-ended tasks invite elaboration. Narrow scope instructions reduce the surface area where hallucination can occur.