3.4 Refusals, hedges, and hallucinations

The three failure modes

Refusal. “I can’t help with that.” The model has decided the request violates policy or is unsafe. It is not going to try.

Hedge. “That’s a complex question. It depends on many factors. You should consult a professional.” The model has answered, but the answer is so cautious that it conveys nothing.

Hallucination. “The Muluki Ain of 2074 requires… section 47B states…” The model has answered confidently. Nothing it said is true.

The scary one is the third — refusals and hedges announce themselves. Hallucinations feel exactly like correct answers until you check.

Recognising each

• A refusal usually starts with “I can’t,” “I won’t,” or “I’m not able to.” It’s short. It sometimes offers an alternative (“but I can help you draft a general question for a lawyer”).
• A hedge is long, uses “generally,” “typically,” “it depends,” and never commits to a number, a date, or a specific claim. If you could delete the whole answer without losing information, it was a hedge.
• A hallucination is confident, specific, and — until you check — wrong. Fake case numbers, fake statistics, fake API endpoints, fake historical dates. Confidence is not a signal.

What causes each — and what to change

Refusals

Refusals happen because:

• The prompt looks like a request the model was trained to decline (medical, legal, dangerous, deceptive).
• The prompt reads as sensitive even though your intent is benign (“how do I stop someone from injecting SQL into my form” is a defensive question, but the phrasing can trip filters).

Fixes:

• State the legitimate context. “As a backend developer securing my company’s login form, how do I prevent SQL injection?” is far less likely to be refused than a bare “how do SQL injection attacks work.”
• Ask for the safe framing. “Explain the defenses; don’t give exploit code.”
• Move to a safer verb. “Draft” instead of “impersonate,” “summarize the arguments” instead of “argue for.”

Refusals are not stubborn. Most soften with better context.

Hedges

Hedges happen because:

• The prompt asks a question that genuinely does have no single answer, and the model won’t fake one.
• The prompt didn’t give the model permission to commit.

Fixes:

• Constrain the shape of the answer. “Pick one option and defend it in two sentences” beats “which option is best?”
• Give it an escape. “If the answer genuinely depends on X, say ‘depends on X’ explicitly and give the answer for the two main cases of X.”
• Remove the invitation to hedge. Don’t ask “should I use React or Vue?” — ask “for a solo developer building a portfolio site in a week, pick React or Vue and justify in three bullets.”

Hallucinations

Hallucinations happen because:

• The prompt asked for facts the model has no reliable source for (specific case law, current prices, someone’s phone number).
• The prompt gave the model no way to say “I don’t know.”
• The model was invited to be specific in a domain where it has to guess.

Fixes:

• Ground it (3.3). Give the model source material and force it to quote.
• Provide an explicit “I don’t know” instruction. “If you don’t have this information in the provided source, say ‘not in the source.’ Do not guess.”
• Ask for uncertainty. “For each claim, rate your confidence: high / medium / low.” This doesn’t make the model correct, but it flags the claims most likely to be wrong.
• Verify before you ship. Anything the model produces about specific people, places, statistics, laws, or code that will actually run — check it in a source you trust before you rely on it.

The diagnostic order

When output is bad, ask in this order:

1. Is it a refusal? Rewrite the context and framing.
2. Is it a hedge? Constrain the shape and force a commitment.
3. Is it a hallucination? Ground with sources and give an “I don’t know” option.
4. Is it just wrong-shaped? That’s a format problem, not a reliability problem — go back to Chapter 2.

Check your understanding

Quick check

—

A model responds to your legal question with three paragraphs of "it depends," "you should consult a professional," and "the situation is complex," never giving a specific answer. What failure mode is this, and what should you change first?

1. AA refusal — soften the framing to avoid triggering the safety filter.
2. BA hallucination — add grounding with sources.
3. CA hedge — constrain the shape of the answer ("pick one, defend in two sentences") and give an explicit escape for genuine dependencies.
4. DA formatting bug — ask for JSON.

Check answer→