Elicitation Bias

Elicitation bias is the systematic distortion of capability evaluation that arises from the choice of elicitation method rather than from properties of the system being evaluated. It is the evaluation equivalent of sampling bias: just as a survey that contacts only landline phones underrepresents young voters, a capability evaluation that tests only zero-shot prompting underrepresents the system's true reasoning capacity.

The bias operates at multiple scales. At the methodological scale, it privileges techniques that are easy to standardize over techniques that are more powerful but harder to replicate. At the institutional scale, it privileges evaluations produced by well-resourced labs with access to frontier techniques over evaluations produced by regulators or civil society with limited technical capacity. At the epistemic scale, it creates a recursive blind spot: the community's understanding of what AI systems can do is shaped by the methods the community knows how to use, and the community's incentive to develop new methods is shaped by its understanding of what remains undiscovered.

Elicitation bias has practical consequences for AI governance. A regulatory framework that mandates specific evaluation protocols — say, standardized benchmarks with fixed prompting — may institutionalize a particular elicitation bias, making it legally required to underestimate capabilities. The framework would be measuring compliance with a biased instrument, not safety.

The concept connects to capability elicitation as its critical shadow, and to epistemic infrastructure as a diagnosis of how institutionalized measurement practices can become self-validating even when they are systematically wrong.