Grounding tells the model what to look at.
It does not check what the model did with it.
Grounding reduces what AI invents. It does not check what AI changed. Between the material a model is given and the output it produces, there is a gap where meaning drifts, conditions disappear, and conclusions outrun the source that was supposed to support them. Grounding does not close that gap. Reading the output back against the source does.
What grounding actually does
Grounding covers a family of techniques: retrieval-augmented generation, document grounding, cited sources. All of them constrain what material the model works from. Instead of writing from what it picked up in training, the model is handed the relevant material and expected to produce output that reflects it.
This is genuinely useful. It reduces the rate at which models produce claims that have no basis in any source. That is the hallucination problem, and grounding meaningfully addresses it.
What grounding does not do is check the output. The model receives the material. It does not reproduce it. It compresses, foregrounds, and decides: what to include, what to drop, how to frame what remains, how confidently to state what it worked out. That process moves meaning, and grounding is not designed to catch it.
Claims generated with no basis in any source. The model invented something. Grounding reduces this by anchoring the model to retrieved material.
The model saw the evidence. It did not faithfully represent it. A condition was dropped. A risk was softened. A conclusion the evidence never reached was asserted. Grounding provides no signal that this happened.
The same evidence. A different output.
A grounded model working from the same clause can produce outputs that range from faithful to badly distorted, all without inventing anything. The model has the material. What it does with it is the problem.
The vendor is obligated to deliver a remediation plan within 30 days of any material breach, subject to written approval from both parties before implementation.
The vendor is expected to deliver a remediation plan within 30 days of any material breach.
The model did not invent the remediation plan or the 30-day timeline. It was grounded on them. What it changed was the nature of the obligation and the need for approval from both sides. That is exactly the kind of shift that decides whether a clause protects you or not. Grounding gave no signal that anything had changed.
Why better models don't close this gap
A reasonable assumption is that as models improve, faithfulness improves with them. What we see does not support this.
Language models are optimized to produce fluent, coherent, confident output. That is the objective. Fluency and faithfulness are related but not the same, and where they pull apart, fluency wins. A more capable model produces more convincing output, not necessarily more faithful output. The drift gets harder to spot, not rarer.
Fluency is not faithfulness. A more fluent output can be a more convincingly distorted one. Better AI makes the grounding gap harder to see, not smaller.
The pressure on model development runs the other way. Users reward output that reads authoritatively, that resolves ambiguity, that presents conclusions cleanly. Keeping the conditions, hedges, qualifiers, and unresolved tensions of the original works against the fluency that makes output readable. The models are doing exactly what they were built to do.
The layer grounding doesn't provide
The check sits outside the AI. After the model produces output, something separate reads both the material the model was given and the output it produced, then works out what came through faithfully, what changed meaning, what was left out, and what has nothing behind it at all.
This is what Plumb does. It takes two things: the material the AI worked from, and the output exactly as produced. It returns a report on what held. The report does not score confidence or predict accuracy. It traces each claim in the output back to the material that was supposed to support it, and says what it finds.
Grounding and checking afterwards are not competing approaches. They solve different problems in the same workflow. Grounding constrains what goes in. The check reads what comes out. Both are necessary. Neither substitutes for the other.
The grounding gap is the difference between telling a model what to look at and checking that what it produced still says what the source said. Grounding constrains the input. It does not check the output. A grounded model can still soften a condition, drop a qualifier, omit a risk, or state a conclusion the source never reached, fluently and with no signal that anything changed.
No. Grounding reduces hallucination, meaning claims the model invents with no basis in any source. It does not stop meaning from shifting, material from being dropped, or conclusions the source never supported. These happen while the model is working directly from the grounded material. The model sees the material. It does not reproduce it. It compresses, foregrounds, and decides, and that process moves meaning in ways grounding does not address.
No, and the gap may get harder to spot as models improve. Better models produce more fluent, more authoritative output, and that fluency makes drift harder to catch on a quick read. What these models are built for is coherent, confident output, not faithful reproduction of the source. A more capable model closes neither the gap nor the risk that a smoothly written output has quietly changed what the source said.
Hallucination is when a model produces claims with no basis in any source. It invented something. The grounding gap is a different failure: the model worked from real material and still changed what it said. The output is not invented. It is unfaithful. Grounding addresses hallucination. It does not address the grounding gap.
Reading the output back against the source once the AI is done. A separate step reads both the source and the output and works out what came through faithfully, what changed meaning, what was left out, and what was invented. This is what grounding does not provide. Plumb is built to do it: sitting outside the AI, taking both, and returning a report before the output reaches anyone who will act on it.