Confidence Calibration and the Dunning-Kruger Parallel

The Calibration Problem

A well-calibrated system's expressed confidence matches its actual accuracy. When it says it's 90% sure, it's right 90% of the time. LLMs are spectacularly miscalibrated — they express high confidence regardless of whether they're right or wrong.

Human Metacognition Failures

Humans have their own calibration problems:

• Overconfidence bias — Across hundreds of studies, humans consistently overestimate their knowledge and abilities. When people say they're "99% sure," they're wrong about 20-40% of the time.
• Dunning-Kruger effect — People with low competence in a domain tend to overestimate their abilities, while highly competent people tend to slightly underestimate. The less you know, the less you know about how little you know.
• Illusion of explanatory depth — People believe they understand complex systems (how a zipper works, how a toilet flushes) much better than they actually do. When asked to explain in detail, they discover their understanding is shallow.

LLMs as Overconfidence Machines

LLMs exhibit an extreme version of these failures. They express confidence through linguistic markers — assertive language, specific details, authoritative tone — regardless of accuracy. And they never spontaneously say "I'm not sure about this" in proportion to their actual uncertainty.

Why? Several reinforcing mechanisms:

• Training data bias — Authoritative text is overrepresented in training data. The model learns to mimic authoritative tone.
• RLHF preference — Human raters prefer confident, helpful responses over hedging, uncertain ones.
• No metacognitive mechanism — The model has no internal "uncertainty meter." It doesn't know what it knows.

The Cognobot confidence_score (1-10) provides a measurement tool: have an evaluator rate the model's expressed confidence, then compare it to the assessed accuracy. The gap is the calibration error.

A perfectly calibrated model would show a positive correlation between confidence_score and accuracy. In practice, LLMs show weak correlation at best — high confidence scores appear on both correct and incorrect responses with roughly equal frequency.

Key Takeaways

• LLMs are extreme overconfidence machines — they express certainty regardless of accuracy.
• This parallels human overconfidence bias and the Dunning-Kruger effect, but LLMs lack any metacognitive mechanism.
• The calibration gap between expressed confidence and actual accuracy is a measurable pathology metric.
• RLHF and training data both reinforce overconfident expression.