Grounding is tying generated text to sources it can actually point to. An ungrounded claim is one the model produced with nothing real behind it, and ungrounded output is where hallucination starts. Grounded writing can show its work; ungrounded writing only sounds like it can.
Why it matters
A model will happily write a confident sentence with no source under it. It reads exactly like a sourced one, which is the whole problem: fluency is not evidence.
For a brand, ungrounded output is not a style issue, it is a trust risk. The fix is to hold generated text to what it can point at, so a claim that cannot be backed is caught before it ships, not after a customer repeats it.
Grounding is also what makes content citable. Text an engine can trace to a source is text it can quote with confidence; text that floats gets skipped.
How Jinn treats it
Full source-verification is a hard problem, so Jinn ships a concrete, checkable slice of it: a deterministic check in the content quality spine that flags where a draft leans on something it never grounds, like an acronym it never expands or an explanation it defers to later.
It scores higher for more grounded and fails safe: an empty or mostly-quoted draft comes back as unavailable, never as a pass. It is a structural proxy, not a full fact-check, and it is honest about being one.
Related terms
Grounded writing can show its work.
Ungrounded output is where hallucination starts. See how Jinn ties content back to what is real.