Language models sometimes produce confident, incorrect statements. No technique eliminates this, but several reduce it.
Grounding
Provide source material through retrieval or documents, and instruct the model to answer only from it.
Permission to Say "I Don't Know"
Explicitly allow and encourage the model to state when information is missing.
Citations and Quotes
Ask the model to cite or quote supporting passages. Verify citations programmatically where possible.
Constrain Outputs
Use structured output, enumerated options and tools for facts such as prices, dates and account details.
Verification Steps
- A second pass that checks claims against sources.
- Consistency checks: ask several times and compare.
- Automated checks against databases.
Choose Suitable Models
Newer and more capable models generally hallucinate less, and some are trained to express uncertainty better.
Design the Experience
Show sources to users, highlight uncertainty, and make it easy to report errors.
Measure
Build test sets that probe for hallucination, including questions that can't be answered from the sources, and track rates over time.