What it means
A grounded answer is one built from supplied sources, with citations pointing back to them. An ungrounded answer comes from the model's training and cannot be verified — it is an assertion with no provenance.
Grounding is the practical response to hallucination. It does not eliminate the failure: a model can still misread a source, over-generalize from it, or cite a real document that does not actually support the claim. What grounding changes is that the error becomes *checkable*. A reader can follow the citation and see.
The stronger implementations add abstention — instructing the model to say the sources do not contain the answer rather than filling the gap from memory. Getting a model to reliably decline is harder than getting it to answer, but it is what separates a system you can trust from one you have to double-check.
Why it matters
Grounding is what makes AI usable in regulated and high-stakes settings. The distinction it creates — between an unauditable assertion and a traceable one — is frequently the difference between a deployment that passes review and one that doesn't.
In practice
Require citations and check that they resolve. Then test the abstention path explicitly: ask something your sources genuinely do not cover and see whether the system says so or invents an answer. That single test tells you most of what you need to know.