Grounding answers and reducing hallucination.
A language model can sound just as confident when it is wrong as when it is right. Here is what the field does about that, in plain language.
A made-up but confident-sounding answer has a name in the field: a hallucination. The counter-move is grounding, tying an answer to specific evidence so each claim can be traced back and checked. We walk it the way the field is built: what a hallucination is, why it happens, how grounding and attribution work, how support is measured and shaky answers are caught, and where it is heading. The papers behind each idea are linked, so you can read the source yourself.
The renewal closed on March 12. 1
I couldn’t confirm who signed it.
- Status
- Closed won
- Close date
- March 12
A fluent answer that nothing actually supports.
The field name is a hallucination. The counter-move is grounding: tying each claim to evidence you can open.
A language model writes fluent text by predicting what words come next. That is its gift and its trap. When it knows the answer, it sounds confident. When it does not, it can still sound just as confident and be wrong. A made-up but confident-sounding answer is what the field calls a hallucination. The model is not lying and not guessing in the human sense. It is predicting plausible text, and plausible is not the same as true.
The survey literature (Ji et al., 2022) draws a line worth keeping. A factual error is a claim that is false in the world. A faithfulness error is a claim the source did not actually support, even if it happens to be true elsewhere. Both matter, and grounding speaks to both.
The fix is not to make the model smarter. It is to change what the answer is allowed to stand on. A grounded answer is tied to specific evidence: a retrieved passage, a record, a source the reader can open and check. Attribution is the sharp version, where every claim carries a pointer to the exact passage that backs it, like a footnote. When each claim points back to a source, a wrong claim has nowhere to hide.
Claim
The account renewed at the higher tier in March.
Evidence
Renewal effective March 1. Plan: Enterprise.
Confirm that the source supports each detail, including the plan and date.
Fluency hides uncertainty, and support is fine-grained.
A wrong answer arrives in the same confident voice as a right one, and a paragraph is rarely all right or all wrong.
Fluency hides uncertainty.
The model writes a wrong answer in the same confident voice as a right one. Nothing in the prose says "I am unsure," so tone tells you nothing about whether a claim is true.
Knowledge is baked in, not filed away.
A model's knowledge is fixed at training time and stored as patterns, not facts in a filing cabinet. It can blend two real things into one that never existed.
Plausible beats true.
Training rewards text that looks like the data, not text checked against the world. A smooth fabrication can score well, which is why fluency and accuracy can come apart.
One label hides the mix.
A paragraph is usually a blend of supported and unsupported claims, so a single thumbs-up or thumbs-down hides the truth. Measuring support has to go claim by claim.
Often you cannot peek inside.
Many real deployments cannot read internal token probabilities or an external database at answer time. Catching a shaky answer then has to work from the outside.
Saying "I do not know" is its own problem.
A good system has to recognize when it lacks enough evidence and say so, rather than fill the gap. Teaching a model to abstain at the right moment is a hard problem in itself.
Reduction is a toolkit, not a single trick.
Ground in evidence, attribute each claim, measure support, catch errors by consistency, have the answer check itself, and know when to abstain. Layers, each doing one job.
1Ground in retrieved evidence
Answer from the source, not from memory.
A direct defense is to give the model the relevant source material and ask it to answer from that, not from frozen memory. Grounding shrinks the room to hallucinate, because the answer is built on text that is right there to check.
The retrieval mechanics, how the right passages are found in the first place, get the full treatment on the sibling page. Here the point is narrower: once the evidence is in hand, the answer is supposed to stand on it. See Retrieval-augmented generation for the "look it up first" half of the story.
| Claim | Evidence |
|---|---|
| Renewed in March | Effective date: March 1 |
| Enterprise tier | Selected plan: Enterprise |
| Three-year term | Not found in retrieved material |
Do not add the unsupported term length.
2Attribute every claim
Each claim carries its receipt.
Beyond answering from sources, the answer attaches each claim to the passage that backs it. Attribution turns "show your sources" into something that can be checked mechanically, claim by claim.
There is a second pattern worth naming. RARR (Gao et al., 2022) takes an answer a model already produced, researches it, and revises the parts that no evidence supports, while keeping the rest intact. So attribution can come before the answer, or after it as a repair pass.
- Signed by Dana
- Contract, signature block.
- Effective March 1
- Order form, effective-date field.
- Discount carries over
- No supporting passage found.
Ask for the missing pricing document or omit the assertion.
3Measure support, fact by fact
Turn "is it accurate" into a number.
FActScore (Min et al., 2023) breaks a long answer into atomic facts and scores the share that a reliable source supports. This turns a vague "is it accurate" into something you can track over time, and it exposes the supported-and-unsupported mix that a single label hides.
The tally below uses illustrative sample data only. It is a way to picture the measure, not a rating of any product.
| Atomic claim | Judgment |
|---|---|
| Renewal is in March | Supported |
| Enterprise tier selected | Supported |
| Three-year term | Unsupported |
| Dana signed | Supported |
75% in this illustrative sample.
4Catch errors by consistency
Ask it a few times and watch.
Ask the model the same thing several times. If the answer is grounded in something real, the samples tend to agree. If it is a fabrication, they tend to wander. Self-consistency (Wang et al., 2022) uses this for reasoning: sample several solution paths and keep the answer they converge on, with reported accuracy gains on reasoning benchmarks attributed to that paper.
SelfCheckGPT (Manakul et al., 2023) turns the same idea into detection. It samples several responses with no external database and no access to internal probabilities, then flags the claims the samples do not back each other up on. A way to catch fabrications from the outside.
| Sample | Answer |
|---|---|
| 1 | Closed in March |
| 2 | Closed in March |
| 3 | Closed in January |
The samples disagree; none becomes true merely by winning a vote.
5Have the answer check its own work
A draft that audits itself.
Chain-of-Verification (Dhuliawala et al., 2023) has the model draft an answer, then write its own targeted fact-check questions, answer those independently so they are not swayed by the draft, and revise. The draft audits itself against the questions before anyone sees it.
The independence step is the quiet key. Answering the check questions without looking back at the draft is what keeps the model from simply agreeing with itself.
- Draft
Write a candidate answer.
- Extract check questions
For example: “What is the effective renewal date?”
- Verify independently
Consult evidence rather than relying on the draft’s wording.
- Revise
Remove or correct unsupported details.
6Know when to abstain
An honest "I do not have that."
When the evidence is weak, the strong move is to say "I do not have enough to answer that" rather than guess. In a grounded system this is concrete: if nothing retrieved clears a support bar, the answer is a clean decline, not a confident invention.
Phrased neutrally, abstention is about declining when the evidence is thin. It is a judgment on whether a claim is supported, nothing more, and it is what makes the difference between a system you can trust and one that always has something to say.
What discount did we promise?
- Available evidence
- A note says pricing was discussed, but records no amount.
- Missing evidence
- The final quote or an explicit confirmation.
Ask for the quote or confirm with the owner.
Trust that comes from evidence, not tone.
At the field level: answers you can check, a number for accuracy, fewer confident wrong answers, an honest "I do not know," and a trail a reviewer can retrace.
This family of techniques is how the industry makes an assistant's answers checkable and keeps confident errors out. Here is what that feels like as an everyday outcome for a seller: you ask about an account and get an answer you can check, with the source right there. Fewer confident-sounding answers turn out wrong. And when the record does not hold it, you get an honest "I do not have that" instead of a made-up one.
Answers you can check.
Every claim can carry the line it stands on, so trust comes from evidence rather than tone. Ask about an account and the source can be right there. See Deep context.
A number for accuracy.
Support can be measured fact by fact, so quality stops being a vibe and becomes something a team can track over time.
Fewer confident wrong answers.
Consistency checks and self-verification catch a class of fabrications before a person ever sees them, which is exactly what you want on a live call.
An honest "I do not know."
A system that declines when evidence is thin is more useful, and more trusted, than one that always answers. And when answers are grounded, a reviewer can retrace how a claim was reached.
This explainer covers published research. To see how AI REV Voice helps with sales conversations, explore Prism.
The direction the field is heading.
We describe the direction, not a roadmap of our own.
From one score to a live signal.
Moving past a single number toward a per-claim "is this supported" signal attached to the answer itself, so support shows up sentence by sentence.
Knowing when it actually knows.
Better signals for when a model truly knows versus is improvising, so the "I do not know" line fires at the right moments rather than at random.
Light enough for every answer.
Self-checking and consistency methods cost extra compute today. The work is making them light enough to run on every answer, not just the high-stakes ones.
Grounding in tables and records.
Tying answers to tables, records, and structured data, not just paragraphs, so attribution holds across the messy shapes real work comes in.
Shared ways to measure support.
Public, shared ways to measure factual support and faithfulness, so progress is comparable across systems rather than self-reported.
The papers behind this page.
- Ji et al. (2022). Survey of Hallucination in Natural Language Generation. arXiv:2202.03629.
- Wang et al. (2022). Self-Consistency Improves Chain of Thought Reasoning in Language Models. arXiv:2203.11171.
- Gao et al. (2022). RARR: Researching and Revising What Language Models Say, Using Language Models. arXiv:2210.08726.
- Manakul et al. (2023). SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models. arXiv:2303.08896.
- Min et al. (2023). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. arXiv:2305.14251.
- Dhuliawala et al. (2023). Chain-of-Verification Reduces Hallucination in Large Language Models. arXiv:2309.11495.
- Lewis et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401.
Keep reading.
Sibling explainers in the Research section.
Talk. Done.
Reading is one thing. Hand a phone number over and ask an account a question out loud, then check the answer for yourself.
No app required. Just your phone number.