AI That Doesn't Hallucinate? Use Your Own Sources

Ask a general-purpose chatbot to summarize a contract or pull a number out of a report, and you get a fluent answer that sounds right. The problem is the word "sounds." A language model predicts plausible text; plausible and correct are not the same thing. For anything where being wrong has a cost — money, legal terms, medical details — fluency isn't enough. You need answers you can check.

So is there an AI that doesn't hallucinate? Honestly, no: every language model can produce something that sounds right and isn't. What you can get is an AI that answers from your sources and shows you, quote by quote, what it actually found there. That's the job of Facts mode: answer from sources you trust, and make the key claims checkable.

Grounding: answer from the source, not from memory

A normal chat answers from the model's training. Facts mode answers from your material — the documents you attach or text you paste. If your documents don't contain the answer, it can also search the web (the AI decides when; only its search query — not your files — is sent to a search provider). The model is told, in effect: "Here are the sources. Answer from these. If they don't say, say so."

That last clause is the important one. "The documents don't contain that information" is a correct, useful answer — not a failure. A system that would rather invent something than admit a gap is the opposite of trustworthy.

Quotes you can verify

Grounding alone isn't enough, because a model can still paraphrase a source into something it doesn't actually say. So Facts mode goes further: it is instructed to back each factual claim with an exact quote — a citation — and that quote is checked against the source before you see it. If the words really appear in your document, the citation stands. If they don't, the claim is flagged as unverified rather than passed off as fact.

The check itself is code, not a promise in a prompt: a literal text match against what you uploaded (ignoring only differences in capitalization, spacing, and quote-mark style):

• The claim is backed by a quote that exists in the source → shown as verified.
• The quote can't be located → shown as unverified, so you know to look closer.

What this doesn't cover: a sentence the model writes without quoting anything isn't checked the same way. The verified quotes are the parts you can lean on; treat the rest as the model's summary of them.

Math that's actually correct

Language models are famously shaky at arithmetic — they predict digits the way they predict words. So when a Facts answer needs to add a column, compute a percentage, or reconcile figures, the model is instructed not to "do the math" in its head but to hand the expression to a real calculator — a safe arithmetic evaluator, not the model's guess — and use the exact result. Each calculation is shown in the answer, so you can see that it was computed.

The difference shows up precisely where it matters most: totals, ratios, and reconciliations, where a confident-but-wrong number is worse than no number at all.

What this does and doesn't replace

Facts mode is a research assistant that shows its sources, not an oracle. It's at its best when you want to:

• Pull specific figures or clauses out of a long document.
• Cross-check a claim against text you actually have.
• Get a summary where the key points trace back to quotes you can check.

It is not a substitute for professional judgement on high-stakes decisions. But it changes the question from "do I trust this answer?" to "let me check the quote it gave me" — and that's a question you can actually act on.

The goal isn't an AI that's never wrong. It's an AI whose answers you can verify in seconds.