Handling refusals and hallucinations

Handling refusals and hallucinations

Why it matters: Refusals and hallucinations are manageable if you handle uncertainty and scope correctly.

Goals

• Recognize refusal vs uncertainty vs hallucination.
• Ask follow-ups safely.
• Remove unsupported claims.

Key definitions

• Refusal: The model declines due to policy or missing info.
• Uncertainty: The model indicates it may be wrong or needs more data.
• Hallucination: A confident but unsupported claim.

Workflow (step-by-step)

1. If refused: reframe the request to a safe alternative (general guidance, template, questions).
2. If uncertain: ask clarifying questions or request sources.
3. If hallucinating: remove the claim; ask for evidence or mark as unknown.
4. Add a “do not fabricate” guardrail to prompts.
5. Use a verification checklist for any factual output.

Example (good vs bad)

✅ Good: You handle refusals by reframing and handle uncertainty by asking for inputs/sources.

❌ Bad: You push for an answer and accept invented details.

Checklist

• Refusal path exists.
• Uncertainty triggers questions.
• Unsupported claims removed.
• Verification applied.

Metrics / criteria

• Unsupported claims shipped: 0.
• Measurable: You can say whether the output is correct/complete (pass/fail or a score).
• Effort: You can produce the result in 15–30 minutes using the workflow.
• Clarity: Another person can run your prompt and get a similar outcome.

Common mistakes + fixes

• Pitfall: Arguing with refusals. Fix: Reframe to safe requests: templates, questions, general guidance.
• Pitfall: Vague instructions. Fix: Add constraints, examples, and a success checklist.
• Pitfall: One-shot prompting. Fix: Iterate: draft → critique → revise → verify.
• Pitfall: No validation. Fix: Add QA steps, tests, and explicit “what counts as done”.

Student tasks

1. Apply the workflow to one real work task related to handling refusals and hallucinations.
2. Create a small artifact: prompt, checklist, rubric, table, draft, or decision note.
3. Run one quality check and record what changed after the check.
4. Write a 3-sentence reflection: what worked, what failed, and what you will reuse.

Useful external sources

• OpenAI prompt engineering guide: https://platform.openai.com/docs/guides/prompt-engineering — Prompt structure, iteration, constraints, and evaluation habits.
• NIST AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework — Risk framing, governance language, and verification mindset.
• OWASP Top 10 for Large Language Model Applications: https://owasp.org/www-project-top-10-for-large-language-model-applications/ — Practical AI safety risks such as prompt injection, leakage, and insecure output handling.
• Google People + AI Guidebook: https://pair.withgoogle.com/guidebook/ — Human-centered AI design, user control, feedback, and failure modes.

Bibliography

• OpenAI prompt engineering guide. https://platform.openai.com/docs/guides/prompt-engineering
• NIST AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework
• OWASP Top 10 for Large Language Model Applications. https://owasp.org/www-project-top-10-for-large-language-model-applications/
• Google People + AI Guidebook. https://pair.withgoogle.com/guidebook/