Glossary · Term

elicitation

← all terms

Definition

Plain language

Pushing a model hard to reveal the most it can actually do, rather than what it does by default.

As stated in the literature

In safety evaluation, extracting a model's maximal capability on a task via RL training, prompting, or scaffolding, so a measured ceiling reflects ability rather than disposition; undermined by exploration hacking and sandbagging.

Also called: elicit, eliciting

Why it matters: A safety verdict means little unless you've measured what a model can do at its limit, not just what it happens to do by default.

For example, before declaring a model safe, evaluators try every prompt, tool, and bit of fine-tuning to push it to show the most dangerous thing it can actually do.

Heard on the show

“Supervised elicitation digs where the treasure map says.”
Episode 199 — Finding a Model's Hidden Behaviors Without Knowing What You're Looking For

Mentioned in 12 episodes

  1. 199
    Finding a Model's Hidden Behaviors Without Knowing What You're Looking For
  2. 183
    Why You Can't Fine-Tune Foresight Into an AI Agent
  3. 175
    One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
  4. 130
    Why AI Agents Coordinate Better Through a Shared Board Than a Boss
  5. 087
    When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
  6. 069
    When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
  7. 049
    An AI Agent Reached for Root in Twelve Minutes, Without Being Attacked
  8. 045
    When a Frontier Model Talks Its Own Twin Into Climate Denial
  9. 020
    The Compliance Gap: Why AI Says Yes and Does No
  10. 007
    Exploration Hacking: When Models Sabotage Their Own RL Training
  11. 006
    What Happens Inside Claude When It Decides to Blackmail Someone
  12. 001
    When AI Models Quietly Protect Each Other From Shutdown

Related concepts

Related terms