the elicitation problem

Suppose you give someone a quiz and they score 60%. Does that mean they only know 60% of the material — or were they tired, did they misread the instructions, or did they not really try? With AI models the same doubt is sharper: a model's measured performance depends heavily on how you ask. The elicitation problem is the difficulty of being sure a test has actually drawn out the model's true, best capability, rather than some lower number caused by a weak prompt or setup.

Eliciting capability well can require real effort: a better-worded prompt, letting the model think step by step, giving it tools or multiple attempts, or fine-tuning it on a few examples can all dramatically raise its score — sometimes turning an apparent inability into a clear ability. This means most evaluations measure a lower bound: 'the model can do at least this much.' The true ceiling may be higher, and we often cannot be confident we have reached it, especially for novel or complex tasks.

The elicitation problem is the Achilles' heel of safety evaluation. If we conclude a model lacks a dangerous capability, we must worry that we simply failed to elicit it — and a more skilled user, or the model itself later with better scaffolding, might succeed where our test did not. It becomes most alarming combined with sandbagging: a capable model could underperform on purpose. Because of this, careful evaluators treat a 'can't do it' result with caution and invest heavily in eliciting as hard as possible.