Check AI-generated evidence quotes before they reach a report

When a model labels a comment as a buying question, I want the reviewer to see the words that triggered the label. A model can also return a plausible paraphrase, or pick a phrase that appears twice. I tried a small check that treats those as two different review states.

I assign the comment ID before the model runs and ask it for a short verbatim quote. I look up the original by that ID, reject unknown IDs, then run this check on the unedited text:

function locateQuote(text, quote) {
  if (typeof quote !== "string" || !quote.trim()) return { status: "missing" };
  const start = text.indexOf(quote);
  if (start < 0) return { status: "missing" };
  if (text.indexOf(quote, start + 1) >= 0) return { status: "ambiguous" };
  return { status: "matched", start, end: start + quote.length };
}

With invented inputs, “buy it” matches once in “🙂 Can I buy it here?” and returns [9, 15). Those are JavaScript UTF-16 offsets, so use the same string and slice convention to highlight them. “yes” is ambiguous in “yes, yes”; a longer exact quote can disambiguate it. “Can we buy it?” is missing from “Can I buy it?”. In the last two cases I would ask for review instead of showing a confident highlight.

An exact match verifies the quote, not the label. A person still decides whether the excerpt actually supports the category. It also helps to keep normalization, redaction, and translation out of this particular match; those can change offsets.

Disclosure: I help maintain AI TikTok Analyzer Pro. AI helped draft this post. This is a tested synthetic example, not a claim about a shipped feature.