Skip to content

AI Anti-pattern

The Wrong Tool

Asking the model for tasks it handles poorly and accepting the result.

Why it matters

AI models have a jagged capability frontier: excellent at some tasks, unreliable at others that look similar. Raw arithmetic, post-cutoff facts, and precise character counts are in the valley, not on the peak. Treating valley outputs as authoritative produces errors that compound silently.

  • Jagged frontier: Noy & Zhang (2023) on capability variance across task types
  • Willison: the model will hallucinate with confidence on tasks it handles poorly

What this looks like

Repeated asks for raw arithmetic, post-cutoff factual recall, exact character or token counts, or long-tail statistics, accepted without a verification request.

This is a generic example. Run your own analysis to see instances from your actual AI conversations.

See this in your own conversations

Paste a conversation with Claude, ChatGPT, or Gemini to find out if you show signs of The Wrong Tool.

Run your own analysis →