Human verifiers check AI claims under effort and error limits

Human-LLM Deliberation as Interactive Proof: Conditions for Verifiability Without Transparency

Computation and Language

Summary

When an AI language model makes a claim too complicated for a person to check directly, it's hard to know if it's true. The authors study how a human and AI can work together like a question-and-answer game where the human asks for proof details and checks them. They prove that if the human’s checking and the AI’s responses meet certain error limits, the human can trust the claim without seeing everything inside the AI’s ‘mind.’ But the human’s effort, expertise, and fatigue affect how much checking can be done. Their findings help understand how humans might verify AI statements without needing full transparency.

What this means in practice

  • For ai system designers: Design interactive protocols where users check AI claims under controlled error and effort limits to improve trust without exposing AI internals.
  • For legal compliance teams: Develop methods to verify AI-generated arguments through stepwise user checks when full AI transparency is impossible.

A theory result. No direct application yet.

Authors

Baotong Zhang, Dean Foster, João Sedoc

Abstract

When an LLM supplies an argument that a user could not readily construct, how can the user decide whether to accept its claim? Inspired by interactive proofs, we model human-LLM deliberation as an interaction between a prover with unrestricted internal search and a resource-bounded human verifier. The verifier requests and checks supporting details without access to the LLM's internal state. Passed checks accumulate evidence toward an acceptance threshold. We prove anytime-valid soundness against adaptive provers: the probability of ever accepting a false claim is at most a chosen error level, provided the task supplies bounds on false passes and human checking errors that remain valid after every relevant history. A finite-horizon completeness bound additionally requires bounds on the adequacy of honest responses and sufficient diagnostic progress. Further checks can strengthen the evidence for acceptance, but each requires another adequate response and reliable human effort. Whether this tradeoff permits certification depends on the verifier's effort budget, cognitive load, expertise, and fatigue. We identify conditions under which the supplied bounds certify a specified sequence of local checks but not a specified global check under the same resource budgets.
View PDFOpen arXiv