Witness urges citation-first tools and enforced UX checks to make AI outputs trustworthy for lawmakers

During a subcommittee exchange about applying large language models to legislative materials, a witness recommended practical design standards for trustworthy outputs: require traceable citations, surface items the model could not resolve, and force manual confirmation on unclear items before permitting downstream use.

The witness contrasted fully generative systems such as ChatGPT, which "just kind of aggregate and spit something out," with citation-forward tools like Perplexity that attach sources to each claim. The witness told the panel: "Perplexity has a citation and a, like, tags on every single thing. So you can go check that citation." They argued that if the House were to set a standard that every line should trace back to a source, there would be no technical barrier to enforcing it, though human judgment would still be required to interpret and verify sources.

The witness emphasized operational UX controls as a second line of defense: build workflows that do not allow users to proceed until unresolved items are acknowledged and verified. A questioner asked whether these safeguards would eliminate the need for human review; the witness replied that the answer depends on engineering choices and that design can reduce but not completely remove human accountability. The transcript contains no formal votes or adopted rules; the discussion remained advisory and exploratory in the provided segments.