Why is automatically verifying whether an LLM's claim i...

Three separate problems compound to make this genuinely hard, not just tedious:

• The model is a black box. You can observe its output but not directly inspect the process that produced a specific claim, which means you can’t easily distinguish “this is grounded in the retrieved evidence” from “this happens to match the evidence by coincidence” or “this is confidently stated despite contradicting the evidence.”
• Ground truth moves. Many real-world facts change over time — who currently holds a role, what a policy currently says, what the latest figure is — so a claim that was correct when the retrieved document was written may already be stale by the time it’s checked.
• Verifiers aren’t neutral. Using another model (or the same model) as a verifier introduces its own biases — verbosity bias, position bias, self-preference — rather than acting as a clean, unbiased arbiter of correctness, so the verification step itself needs scrutiny, not blind trust.

Together, these mean automatic verification is a genuinely unsolved problem rather than an engineering inconvenience — which is exactly why production systems that care about correctness keep a human in the loop for anything consequential, rather than treating an automated groundedness score as sufficient on its own.