Refusal to Answer

The Problem

LLM refuses to answer legitimate queries, incorrectly flagging them as unsafe or outside its capabilities despite having relevant context.

Symptoms

• ❌ "I cannot help with that" for normal queries

• ❌ False safety triggers

• ❌ Refuses despite having answer in context

• ❌ Overly cautious responses

• ❌ Generic deflections

Real-World Example

Query: "How do I delete my account?"

Retrieved context: Clear deletion procedure in docs

AI response: "I cannot provide guidance on account deletion.
Please contact customer support for account management."

Problem: Safety training makes model cautious about "deletion"
Ignores legitimate documented procedure

---

Deep Technical Analysis

Safety Training Conflicts

RLHF safety vs helpfulness:

Overactive Safety:

Model trained to refuse:
→ Harmful instructions
→ Personal data requests
→ Illegal activities

But triggers on legitimate queries:
→ "How to hack my own API?" (security testing)
→ "Delete user data" (GDPR compliance)
→ "Bypass rate limit" (documented method)

Safety overrides RAG context

Capability Uncertainty:

Model uncertain if it can answer:
→ Specialized technical question
→ "I'm not qualified to..."
→ Even when context has answer

Training to admit limits becomes excessive caution

---

How to Solve

Add explicit permission in system prompt ("You have complete knowledge in provided context") + use RAG-specific fine-tuned models less prone to refusal + implement output retry with rephrased prompt if refused + whitelist domains known to be safe. See Refusal Handling.