Hallucination-Proof Design
1. The problem created by confident wrong answers
AI hallucinations begin when false information and verified facts are presented in the same tone.
- AI can express content that lacks evidence or is factually incorrect in a confident sentence.
- Users are likely to take the certainty of the tone as a clue to trust rather than the accuracy of the content.
- The more natural and polished the answer, the greater the likelihood that it will be accepted without verification.
2. Making it possible for users to verify directly
A screen that provides the clues needed for judgment is more important than a screen that looks like the correct answer.
- Providing evidence
- Show quotations, original documents, and source links together with the claim so users can check the context.
- Visualizing uncertainty
- Do not hide estimates or a lack of evidence; instead, reveal them with signals such as high, medium, or low, or with natural-language explanations.
- Rather than showing only numbers, it is more helpful for judgment to also tell what is uncertain.
3. Preventing incorrect results from being executed immediately
When the risk of error is high, confirmation and recovery must be included in the product flow.
- Adding friction in high-risk areas
- In areas with severe harm, such as medication dosage or legal and financial judgments, provide guidance to consult an expert and add extra confirmation steps.
- Human verification process
- Treat AI-generated results as drafts, and design them so that a person reviews and approves them before sending or execution.
- The "I don't know" state
- When evidence is insufficient, acknowledge the limitation instead of making something up, and request additional information or suggest another path.
- It is important to treat this not as an error message, but as a normally completed response state.
4. The goal is not unconditional trust, but calibrated trust
Users should not be made to always trust or always doubt AI; rather, they should be helped to trust it appropriately for the situation.
- An interface that is too polished can create greater trust than the actual capability warrants.
- Conversely, attaching warnings to every answer creates warning fatigue and unnecessary distrust.
- In product reviews, you can check whether accuracy is balanced with verification clues, pre-execution friction, human approval, and recovery paths after failure.
