Reducing Hallucinations

Why Models Hallucinate

Hallucination — producing confident, fluent, factually wrong content — is an emergent property of how language models work. They are trained to produce plausible text, not verified text. When a model encounters a question it doesn't have reliable knowledge to answer, it doesn't say "I don't know" by default — it produces what sounds like a good answer. This is particularly problematic for specific facts: citations, statistics, dates, names, product details, and legal specifics are all high-risk hallucination zones.

Grounding With Context

The single most effective hallucination mitigation is grounding — providing the source material in the prompt and instructing the model to answer only from that material. "Answer the question using only the provided text. If the answer is not present, say 'I don't know.'" This instruction dramatically reduces fabrication because the model has a concrete source to reference rather than relying on reconstructed training-time patterns. This is the basis of RAG architectures.

Prompting for Honesty

Several prompt patterns push models toward appropriate epistemic humility: - Explicit permission to not know: "If you're not certain, say so rather than guessing." - Confidence labels: "Rate your confidence in this answer: high / medium / low." - Uncertainty surfacing: "What assumptions are you making in this answer?" - Source citation: "For each claim, indicate whether it comes from the provided text or from your general training knowledge."

Temperature and Model Selection

Low temperature (0.0–0.2) reduces hallucination by reducing randomness — the model sticks to its highest-confidence outputs rather than sampling from the long tail. Larger frontier models hallucinate less than smaller ones on knowledge-intensive tasks, as their broader training data provides more reliable grounding. For high-stakes factual applications — legal, medical, financial — always combine grounding with a low temperature and consider adding a verification step where the model checks its own answer against the source material.