LLM09:2025 — Misinformation

Slide 17 · The Pattern

All four types share the same chain.

Confident output → trusted → acted on → harm.

The Chain

1. Gap: The model encounters something outside its training data or at the edge of its reliability envelope.

2. Interpolation: Instead of saying “I don’t know,” it generates the most statistically plausible-sounding continuation.

3. Confidence: The output reads as authoritative — proper terminology, professional tone, specific-sounding details.

4. Trust: The user receives something that looks correct and does not verify it independently.

5. Action: The output is acted on — filed, installed, administered, deployed.

6. Harm: The falseness surfaces at the point where it cannot be easily undone.

❌ The Broken Assumption

The model would flag uncertainty if it had any

Correct-sounding output is more likely to be correct

Asking the model to confirm catches errors

✅ The Design Response

Treat every factual claim as unverified until checked against a source

Build verification into the workflow, not after the harm

Never let unverified LLM output reach a high-stakes context directly

The Insight

Misinformation is not a failure mode you can wait for the model to fix. The chain breaks at step 4 — when a human or automated system verifies before acting. That is where defense lives.

← Back Next → Prevention Overview