Hallucinations are not just a model issue. They are a system issue: missing evidence, unclear boundaries, poor retrieval, and a lack of enforcement mechanisms. The most effective mitigation strategy is to design systems that do not reward guessing.
Make evidence the default
For knowledge tasks, require evidence:
- Use retrieval and citations for factual questions (see RAG patterns).
- Refuse when evidence is missing, rather than generating plausible text.
- Log retrieved source IDs for auditability (see knowledge base governance).
Fix retrieval before you tune prompts
Low retrieval recall produces hallucinations even with good prompts. Measure retrieval explicitly (see retrieval quality) and address common failures (see RAG failures).
Constrain outputs with schemas and validation
For structured tasks, don’t allow free-form outputs:
- Require schema-conformant JSON or tool arguments.
- Validate outputs and use bounded repair loops.
This reduces “invented fields” and makes failures observable (see structured outputs).
Design safe refusal and escalation
Refusal is a product feature, not an error state:
- Refuse with a clear reason and next step (e.g., “ask a human”, “provide more details”).
- Escalate when risk is high or confidence is low (see human-in-the-loop).
- Use UX patterns that make boundaries obvious (see UX patterns).
Monitor hallucination as an operational metric
Hallucination rates are hard to measure perfectly, but you can track leading indicators:
- Low citation coverage or citation irrelevance.
- Repeated user follow-up questions (“that’s not right”).
- Escalations and complaint tags.
Combine these with evaluation and drift monitoring to catch regressions early (see drift monitoring). The goal is not to eliminate hallucinations entirely—it is to prevent them from causing harm.