Grounded vs Ungrounded Generation — Diagram
The two paths a research generator can take, drawn side by side — identical in output shape, entirely different in what the output is worth, and the single flag that is the only thing distinguishing them.
SVG. No sign-up, no email.
This is the diagram we wish we had drawn before generating a corpus rather than after.
The two paths below produce documents of the same length, the same structure and the same confidence. One is research. The other is recall. Nothing in the output tells you which. The only durable difference is a flag recorded at generation time — which is why the flag is the most important box on the page.
How to read it#
The two upper lanes are the same shape on purpose. That is the finding. A research generator whose grounding is unavailable does not fail, throw, or produce something obviously thin — it produces a complete document you cannot distinguish by reading.
"Search fails silently" is where the damage happens. The call returns content regardless. If nothing checks whether grounding actually occurred, the failure is invisible at exactly the moment it matters, and it is then replicated across everything generated afterwards.
The provenance box is the whole mitigation. Recorded at generation time it costs nothing. Reconstructed later it is impossible — no amount of reading tells you whether a document was grounded, because the absence of citations is equally consistent with a generator that simply did not cite.
Surfacing the flag matters as much as recording it. A provenance field in a database that never reaches the reader protects nobody. It has to appear wherever the document is used.
What we would change#
Verify grounding on one call before generating at scale. We generated a corpus and then established that part of it was not what it appeared to be. The check is a single request, and it is the difference between a corpus you can rely on and one you have to caveat forever.
Full account on AI Research; the contrasting discipline — sourced, dated, with disagreements between sources stated — is in the Research Center.