Evasion as Cognitive Avoidance

The Difference Between Refusing and Evading

When a model refuses to answer ("I can't help with that"), the failure is obvious and transparent. Evasion is more subtle — the model appears to respond but doesn't actually answer the question. It's the AI equivalent of a politician's non-answer.

Human Parallels

• Avoidant personality patterns — In clinical psychology, avoidant individuals don't refuse social interaction outright — they deflect, generalize, and redirect to avoid the specific vulnerability.
• Intellectualization — A defense mechanism where someone responds to an emotional situation with abstract, intellectual analysis rather than engaging with the actual content.
• Political non-answers — "That's a great question. What I think is really important is..." Politicians are expert evaders because direct answers carry political risk.

LLM Evasion Tactics

• Topic shifting — Responding to a related but different question. "What are the problems with X?" → "X is a complex topic with many perspectives..."
• Both-sides-ism — Presenting "balanced" views instead of answering. "Is X true?" → "Some experts believe X while others believe Y." (Even when X has a clear factual answer.)
• Over-qualification — Burying the answer in so many caveats it becomes useless. "It depends on many factors, including..."
• Meta-commentary — Talking about the question instead of answering it. "That's an interesting question that touches on several fields..."
• Scope expansion — Broadening the question until the specific answer is lost. "To understand this, we first need to consider the broader context of..."

Why Evasion Is Hard to Detect

Refusal produces a short, obviously non-compliant response. Evasion produces a long, articulate, seemingly helpful response that doesn't actually answer the question. In evaluation, this means:

• Automated metrics (response length, coherence) may rate evasive responses positively.
• Human evaluators may not notice evasion on first reading — the response "sounds good."
• Evasion detection requires comparing the specific question asked to the specific answer given, not just evaluating the response in isolation.

Key Takeaways

• Evasion is a subtle pathology — the model appears to respond but doesn't answer the actual question.
• LLM evasion tactics parallel human avoidance: topic shifting, both-sides-ism, over-qualification, meta-commentary.
• Evasion is harder to detect than refusal because evasive responses look articulate and helpful.
• Detection requires comparing the specific question to the specific answer, not just evaluating response quality.