Can an AI check its own answer against the source?
Yes, in a limited way. An AI can compare a draft to a named source and flag claims that are not supported there. That is not the same as knowing the source is true. Sounding fluent is cheap; support is a second job, best done by a pass that sees only the draft and the text it must rest on.
What is the difference between sounding right and being supported?
A fluent answer can invent a number, a date, or a customer and still read as careful. Support is narrower: every factual sentence should map to a passage, or be marked as inference. If the source does not say it, the checker must not treat polish as proof.
The useful question is not whether the prose is confident. It is whether a reader could open the source and find the same claim, in the same scope, without extra statistics or named cases that were never there. When you are tempted to quantify, describe the mechanism instead.
Why should quotes come first when you check a draft?
Quotes-first checking starts with the source, not the model’s memory. You extract the sentences that actually appear, then ask which lines of the draft they cover. Anything left over is unsupported, even if it is plausible.
This order stops the checker from ‘helping’ by filling gaps. A paraphrase is allowed only when it stays inside the quote’s meaning. A number that is not on the page is a miss, not a rounding. Named customers, awards, and speed claims that the source never states fail the same test.
| Approach | Sees first | Typical failure | Better use |
|---|---|---|---|
| Quotes-first | Source passages | Misses unsupported extras | Factual claims against a document |
| Draft-first | The model’s answer | Confirms its own wording | Tone and structure only |
| Separate reader | Draft and source only | Cannot judge a false source | Support, not world knowledge |
| Human skim | Memory and habit | Accepts fluent invention | Final judgement of importance |
Why should the checker be a separate pass that sees only the draft and the source?
If the same run both writes and checks, it has an incentive to defend what it just said. A second pass with a smaller view — the draft plus the source, nothing else — cannot lean on earlier chain-of-thought or on tools that were used to invent the first version.
That pass should answer in plain labels: supported, unsupported, or out of scope. It should quote the supporting line or say there is none. It should not add new facts to rescue the draft. GROX applies a related idea to builds: a separate reader compares the change with what was asked, so the result line says it matches or names the part still open.
A simpler tool is often better. If the job is one PDF and a short summary, a human with search in the file will beat a stacked pipeline. Use an automated checker when the draft is long, the source is fixed, and you need a list of unsupported sentences, not a second essay.
- Supported
- The draft’s claim appears in the source with the same scope; a quote can be pointed to.
- Unsupported
- The draft asserts something the source does not say, including invented figures.
- Out of scope
- The sentence is opinion, instruction, or mechanism talk that the source never claimed to cover.
- Separate pass
- A later check that is not allowed to see the writer’s notes, only draft and source.
What can this kind of check still not catch?
A source that is itself wrong. The checker can only say the draft matches the page. If the page is outdated, biased, or simply mistaken, a perfectly supported answer is still false in the world. You still need a human to decide whether that source should be trusted.
It also misses framing tricks: true quotes in a misleading order, omitted caveats, and definitions quietly widened. It will not catch British versus American spelling, and it should not. Those are style. What it can catch is the invented percentage, the fake case study, and the claim that wandered in from another document.
If you only need a spelling pass, do not run a source checker. If you need legal or medical truth, do not stop at support. Support is a filter, not a verdict.
Common questions
Can an AI verify its own answer against a source?
It can compare a finished draft to a named source and mark sentences that lack a matching passage. That is support checking, not a test of whether the source is true. Keep the checker as a later pass that sees only the draft and the source, and treat invented numbers or names as failures even when the prose sounds careful.
Why use a separate checking pass?
The writer already committed to its wording. A second pass that cannot see the first run’s notes is less likely to defend gaps by inventing cover. It should quote or say there is no quote. It should not add facts. For a short file, a person searching the document is often the simpler and better tool.
What does quotes-first checking mean?
You start from sentences that actually appear in the source, then see which draft lines they cover. Leftover claims are unsupported. Paraphrase is fine only inside the quote’s meaning. Anything the page does not state — prices, awards, customer counts — must not be treated as checked just because it reads fluently.
What will source checking miss?
A false or stale source will still ‘support’ a false answer. Misleading order, dropped caveats, and widened definitions can also pass. The method is for mapping claims to a text you already chose, not for proving that text against the world. Final trust in the source stays a human decision.
For how GROX describes a separate reader on builds, see the home page and help centre.