ConceptRetrieval & RAG
AI Hallucination & Grounding
At a glance
Confident, fluent, wrong, and how grounding + citations contain it.
- Who this is for
- Engineers and technical readers learning the terms used in AI systems.
- Topics
- Retrieval & RAG
- Concept
In short
An AI hallucination is model output that reads as fluent and confident but is false or unsupported, such as an invented figure or a citation to a case that does not exist. Bigcircle contains it outside the model: answers come from retrieved evidence or are withheld, citations are validated against a whitelisted index, and a draft says less when retrieval returns too little. Our guide on why RAG systems fail shows why retrieval alone does not end it.
Key facts
Definition
Hallucination is model output that reads as assured and is factually wrong or unsupported, and grounding is the practice of answering only from retrieved evidence, with a citation, or abstaining.
How Bigcircle does it
- On the trademark platform, Bigcircle enforces citation integrity outside the model, with retrieval over a whitelisted index, schema-constrained decoding and a validator that rejects any citation that does not resolve.
- When retrieval returns too little there, Bigcircle's correct behaviour is a draft that says less, and a failed validation retries using only retrieved evidence instead of the model's memory.
- On a wrong grounded answer, Bigcircle recommends opening the logged context first, because the fix is in retrieval when the passage was missing and in generation when it was present and ignored.
Comparison
A retrieval miss means the right passage never arrived, so the fix is chunking, hybrid search or reranking, while generation drift means the passage was in context and the model contradicted or changed it.
Key number
As of its 16 October 2026 update, Vectara's leaderboard scores about 109 models with HHEM-2.3, from 1.8% unsupported summaries at best to above 20% at worst.
An AI hallucination is output from a language model that reads as fluent and confident but is false or unsupported by any source: an invented statistic, a citation to a case that does not exist, a policy clause nobody wrote. In LLMs the false answer comes out of the same process, with the same confidence, as the true one, so nothing in the text marks it. Grounding is the countermeasure: the model answers from retrieved evidence, cites it, and declines when the evidence is not there.
The definition matters less to a builder than one consequence of it. A hallucination cannot be spotted by reading the output, only by checking it against something outside the model. Every defence on this page is a way of arranging that check.
Why do LLMs hallucinate?#
A model generates the likely next token given everything before it. Training rewards fluent, plausible continuations, not verified ones, so when the likely words are false, the model writes them as smoothly as it writes facts. Ask for a citation behind a niche claim and you can get a perfectly formatted reference that does not exist, because a well-formed citation is exactly what the objective rewards.
Kalai and colleagues at OpenAI add a second cause in a September 2025 paper. Most benchmarks grade answers right or wrong, so a model that says "I don't know" scores zero while a model that guesses sometimes scores. Trained and evaluated against such tests, models learn to guess when uncertain. They also argue that some error is statistical rather than fixable: if a model cannot tell certain false statements from true ones, pretraining alone will produce them.
Take two things from that for a production system. Assume any answer the model gives from memory alone can be wrong in this way, however it reads. And check your own evals for the same flaw: if an abstention scores as a failure, your test suite rewards guessing too.
Grounding and cite-or-abstain#
Grounding moves the model from recalling to reading. You retrieve passages, tell the model to answer only from them, and require a citation to the span behind each claim. That constrains what the model can say and makes every claim checkable. The strict form is cite-or-abstain: if no passage supports an answer, the system answers that it could not find one. Count an honest "not in the documents" as a correct result in the product's metrics, not as a failure, or the metrics will push the team back toward guessing.
Retrieval does not end hallucination#
The common advice is that RAG fixes hallucination. It moves the problem rather than removing it, in two ways.
A grounded model can still drift from its sources: the right passage is in the context and the answer contradicts it, adds to it, or rephrases it into something else. And a grounded model can cite something real that does not apply. On our trademark drafting platform, if a matter is tagged with the wrong jurisdiction, retrieval returns authorities that are real, well formed and correctly cited, and wrong for the matter. Every validator passes them, and the attorney catches them or nobody does. No hallucination score finds that, because nothing was invented.
Fine-tuning is the other fix teams reach for, and on citations it makes things worse. On the same platform, fine-tuning improved structure and register, and it improved how a citation is phrased, which is the wrong thing to improve when the authority behind it might not exist. Citation integrity there is enforced outside the model: retrieval runs over a whitelisted index that is the entire set of citable authority, schema-constrained decoding fixes the shape of the draft, and a validator rejects any citation that does not resolve.
Retrieval miss or generation drift#
When a grounded system gives a wrong answer, split the failure before fixing anything. Open the logged context for that query and check whether the answer was in it.
If it was not, the right passage was never retrieved, and the model filled the gap from memory. That is a retrieval bug, and the fix is in chunking, hybrid search or reranking. If it was, the model had the evidence and ignored or changed it. That is a generation bug, and the fix is in the prompt, the model, or a stricter cite-or-abstain rule. Teams that skip the check tune prompts to fix retrieval and rebuild retrieval to fix prompts, and both changes miss.
How to measure it, and how far to trust the measure#
Groundedness scoring breaks an answer into claims and checks each one against the retrieved context. Two kinds of checker are common. Purpose-built detectors, such as Vectara's HHEM, are small models trained to judge whether a summary is supported by its source. As of its 16 October 2026 update, Vectara's leaderboard runs HHEM-2.3 over about 109 models, with the best at 1.8% unsupported summaries and the worst above 20%. The other kind is an LLM judge checking each claim against the sources, the approach behind Google DeepMind's FACTS Grounding benchmark, which averages several judge models to reduce the bias of any one.
Neither is reliable on hard cases. FaithBench, a 2024 set of summaries chosen because popular detectors disagreed on them and labelled by human annotators, found that even the best detectors were near 50% accurate. Use automated scores for trend lines, regression tests in your evals, and flagging answers for review. Do not let one automated score be the last check on an answer that carries legal or financial weight.
Design so the model can say less#
Two of our builds make silence the default when evidence runs out. On the trademark platform, when retrieval returns too little, the correct behaviour is a draft that says less, not one that fills the gap from the model's memory, and a failed validation triggers a retry that may draw only on retrieved evidence. On Prospex, our own research engine, a claim with no supporting passage does not appear at all. A short answer is visibly short, while an invented one looks like any other.
Then match review to stakes. A tool that is right 95% of the time still makes an attorney check 100% of it, so on high-stakes work the goal is a constraint the reviewer can verify, not a rate they have to trust.
Hallucination is not only a text-model problem. On the on-device meeting transcription we built with a SalesTech client's team, now in production, the speech model's hallucinated phrases go into a CRM and into prompts, where they break the output rather than degrading it. Cleaning runs once on the transcript, before the guidance model and the CRM read it, because cleaning once at the source is worth more than repairing the same text at every consumer.