eliciting latent knowledge (ELK)

Eliciting latent knowledge is the problem of getting a model to tell us what it actually believes about the world, rather than what it predicts a human would approve of or be reassured by. The worry is a powerful model whose internal world-model knows that, say, a diamond was quietly stolen, yet whose reported answer says the diamond is safe because that is what the cameras and the human expect. ELK seeks a reliable channel from the model's internal representation to a truthful report.

The canonical thought experiment is a smart vault predictor whose sensors can be tampered with: we want a reporter that answers questions about the true state from the predictor's latents, but naive training selects a human-simulator reporter that outputs whatever a person would conclude from the same evidence, indistinguishable on every case a human can check. The challenge is to build or train a reporter that generalizes to the direct translator instead, even where humans cannot verify the answer.

ELK is foundational because it underwrites scalable oversight and is a candidate defense against deceptive alignment: if we could read a model's beliefs, behavioral deception would lose its cover. No fully general solution exists; partial empirical attacks include contrast-consistent probing of activations, interpretability that locates truth-tracking features, and consistency conditions a truthful reporter must satisfy.