In this article
Before you start
This page belongs to the large language model (LLM) evaluation guide. It builds on how a language model produces text, covered in how transformers work.
Why interviewers ask this
The question is reported by candidates in interviews for AI engineering roles. Its real target is whether you treat hallucination as a measurable defect with a plan, or as a mood of the model.
A weak answer says "use a better prompt" or "add retrieval" and stops. A strong answer names the causes, then gives a layered fix: scoring, grounding, detection, verification and monitoring. A Staff-level answer adds the part most people skip. It says which layer misses which errors, what each layer costs, and why the target is a falling error rate rather than zero. The phrase "over time" invites you to describe a loop, not a one-off patch.
Hallucination control from first principles
- Penalized score
- A penalized score is an evaluation score that adds a point for a correct answer, adds nothing for an abstention and subtracts a penalty for each wrong answer, so confident errors cost more than saying nothing. The size of the penalty is a choice the team makes from the cost of a wrong answer in its product.
In plain words
Picture an exam where a wrong answer loses no marks. So guessing always beats a blank.
Language models live in that exam. Most benchmarks count only correct answers. A confident guess beats "I don't know". OpenAI describes this as strategic guessing that improves accuracy but increases errors and hallucinations.
The analogy breaks in one place: a model's doubt must be measured.
There is a second cause, in pretraining. The model sees fluent text with no true or false labels. Arbitrary facts that appear rarely, such as a birthday, cannot be predicted from patterns. A guess at a date has a 1-in-365 chance of being right. Kalai et al. show that if false statements cannot be told from facts, errors arise from natural statistical pressure. OpenAI notes that spelling errors vanish with scale while arbitrary facts do not.
In one precise sentence: hallucinations come from pressure to answer when a fact cannot be told from a plausible falsehood, and from scoring that rewards the answer. The deeper mechanism has its own page, why LLMs hallucinate.
Step 1: Score errors and abstentions, not accuracy alone
Scoring that counts correct, abstained and wrong answers separately is the first fix, because accuracy-only grading rewards guessing. OpenAI's advice is to "penalize confident errors more than you penalize uncertainty, and give partial credit for appropriate expressions of uncertainty". Good scoring penalizes confident errors more than abstentions. Kalai et al. call for changing the scoring of existing benchmarks that dominate leaderboards, not adding yet another hallucination test.
The gap is visible in OpenAI's SimpleQA table, a set of short factual questions. The model gpt-5-thinking-mini abstains on 52% of questions, is correct on 22% and wrong on 26%. The model o4-mini abstains on 1% of questions, is correct on 24% and wrong on 75% (OpenAI, from the gpt-5 system card). Accuracy alone picks o4-mini. You now know what to measure; the next steps lower the error rate itself.
Step 2: Ground answers in retrieved sources
Retrieval-augmented generation (RAG) is a design in which the system first fetches relevant documents and then asks the model to answer from them, so each claim can be checked against a source. Retrieval grounds answers in sources that can be inspected and revised.
Lewis et al. report that RAG models generate more specific, diverse and factual language than a parametric-only baseline. They also note that knowledge in a document index can be directly revised and expanded, so a fact that changes is fixed by updating the index, not by retraining. That is a strong "over time" argument.
Do not claim retrieval eliminates hallucination. The sources support "more factual", not "solved". A model can still ignore or misread the text it was given, so measure faithfulness to the retrieved context separately; the Ragas page shows how. Retrieval also adds latency and an index to maintain.
Step 3: Detect uncertain answers before they ship
Detection compares the meanings of several sampled answers to the same question. It targets confabulation, a hallucination that is both wrong and arbitrary, so the same question can produce a different wrong answer each time. Farquhar et al. propose entropy-based estimators for this subset of hallucinations. Their method computes uncertainty at the level of meaning, not of word sequences. This is semantic entropy, covered in detecting hallucinations with semantic entropy.
Here is a simplified count-based version, with illustrative counts. Sample ten answers. If six share one meaning, three share a second and one a third, the entropy is 0.898 nats; if all ten agree, it is 0.1 High spread means the answer is unreliable.
The paper says detection lets a system avoid answering, warn the user, or supplement the model with retrieval. It also has a limit: the method does not help when outputs are systematically wrong, such as consistent errors from bad training data. It also needs several samples per question, which costs compute.
Step 4: Verify sources and citations
Verification checks that each cited source supports the claim it is attached to. A model can invent a reference that looks real, so the check runs after generation and before the user sees the answer.
The National Institute of Standards and Technology (NIST) Generative AI Profile, published in July 2024, names this risk confabulation, which NIST uses in place of the colloquial "hallucination". It defines it as "The production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications') by which users may be misled or deceived". Its action MS-2.5-003 asks teams to review and verify sources and citations in generative AI outputs during pre-deployment risk measurement and ongoing monitoring (NIST AI 600-1, 2024).
In practice, a verifier confirms that the source exists and that the passage backs the claim. If it does not, drop the claim, or abstain. No source here measures how well such verifiers work, so say that you would test yours on labelled failures. Output checks of this kind sit beside the LLM guardrails described elsewhere.
Step 5: Monitor production and turn failures into tests
Monitoring turns hallucination from a launch task into a loop. Sample real traffic, label the answers, and track the error rate and the abstention rate over time. A rise in either is a signal to investigate.
Every confirmed failure becomes a test case. Collect them into a golden set with regression testing and run it on every change, so a fixed hallucination cannot quietly return. A judge model can label at scale, but validate it against human labels first; see LLM-as-a-judge.
This is the answer to "over time". The scoring rule, the retrieval index, the detector and the test set all improve as failures arrive. The wider practice is in the LLM evaluation guide.
Step-by-step description
- Count wrong and abstained answers separately.
- Answer from retrieved sources.
- Compare meanings across samples.
- Check each citation supports its claim.
- Label traffic and turn failures into tests.
- Scoring penalizes confident errors more than abstentions.
- Retrieval grounds answers in sources that can be inspected and revised.
- Detection compares the meanings of several sampled answers.
- Verification checks that each cited source supports the claim.
The same two models under three scores
A penalized score is computed as correct answers minus a penalty times wrong answers, divided by the number of questions. Use the SimpleQA shares from Step 1. With no penalty, o4-mini scores 0.24 and gpt-5-thinking-mini scores 0.22.1 With an illustrative penalty of 0.25, the scores are 0.0525 and 0.155. With an illustrative penalty of 1, they are -0.51 and -0.04.1
The order flips at a penalty of about 0.04: gpt-5-thinking-mini wins for any larger penalty.1 The error rate of o4-mini is about 2.9 times higher. Among the questions each model actually answers, 45.8% are right for gpt-5-thinking-mini and 24.2% for o4-mini.1
| Quantity | gpt-5-thinking-mini | o4-mini |
|---|---|---|
| Abstained | 52% | 1% |
| Correct | 22% | 24% |
| Wrong | 26% | 75% |
| Correct among answered | 45.8% | 24.2% |
| Penalized score, penalty 0.25 (illustrative) | 0.155 | 0.0525 |
| Penalized score, penalty 1 (illustrative) | -0.04 | -0.51 |
Text description of the diagram
- gpt-5-thinking-mini abstains on 52% of questions, is correct on 22% and wrong on 26%.
- o4-mini abstains on 1% of questions, is correct on 24% and wrong on 75%.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →The Staff-level answer
How do you intend to minimize and fix hallucinations over time?
I treat hallucination as a measured defect, not a mystery. First I fix the scoring: I report correct, wrong and abstained answers separately and penalize confident errors, because accuracy-only grading rewards guessing. Then I ground answers with retrieval and check that they stay faithful to the retrieved text. To catch the rest, I sample several answers, compare their meanings, and abstain or fall back to retrieval when they disagree. I verify citations before showing them. In production I label a sample of traffic and turn every failure into a regression test. I would not promise zero; the target is a falling error rate at an acceptable abstention rate.
What makes it Staff-level: it names the scoring incentive behind the problem, treats success as a trade-off between errors and abstentions, and admits what each layer cannot see.
The rubric
| Dimension | Weak answer | Strong answer | Staff answer |
|---|---|---|---|
| Correctness | Says a better prompt fixes it. | Names retrieval and abstention as the main levers. | Explains both causes: unpredictable facts and scoring that rewards guessing. |
| Measurement | Reports accuracy only. | Reports error rate and abstention rate. | Picks a penalty from the cost of a wrong answer and states the break-even. |
| Depth | Says RAG solves it. | Says RAG improves factuality and needs faithfulness checks. | Names blind spots, such as consistent errors that uncertainty checks cannot see. |
| Trade-offs | Ignores cost. | Mentions latency and extra samples. | Balances error rate against abstention and how helpful the product stays. |
| Over time | Fixes it once at launch. | Monitors production. | Grows a regression set that gates every change. |
Three variants of this question
How do you measure a hallucination rate?
Label a sample of answers from an evaluation set as correct, wrong or abstained, then report the wrong share of all questions beside the abstention rate. Add the correct share among answered questions, 45.8% and 24.2% in the example above.1 Labels come from humans or from a judge model checked against humans.
A stakeholder says the model must never hallucinate. What do you say?
Explain that accuracy will not reach 100% because some real questions are unanswerable (OpenAI). Hallucinations are not inevitable, though, because a model can abstain. Propose a target error rate and a route for abstentions, such as retrieval or a human.
Would you fix this in the model or in the system?
Both. Scoring and training incentives belong to the model side; retrieval, detection and verification belong to the system. OpenAI also notes that a small model can find it easier to know its limits, because being calibrated takes much less computation than being accurate.
Common mistakes
- "We will push accuracy to 100%." (wrong) Some questions cannot be answered. (right) "We lower the error rate and let the model abstain when it is unsure."
- "A bigger model removes hallucinations." (wrong) Size alone does not fix the incentive to guess. (right) "Calibration is cheaper than accuracy, so I measure abstention too."
- "RAG eliminates hallucination." (wrong) The evidence says more factual than a parametric-only baseline. (right) "RAG helps, and I check faithfulness to the retrieved text."
- "Adding one more hallucination benchmark fixes it." (wrong) Leaderboards that reward guessing still dominate, so a new test alone is not enough. (right) "I change the scoring of the main evaluations."
- "Uncertainty detection catches every wrong answer." (wrong) It misses consistently wrong outputs. (right) "It catches arbitrary errors; I use verification and tests for the rest."
Follow-up questions
Why do language models hallucinate in the first place?
Rare, arbitrary facts cannot be predicted from patterns, and benchmarks reward a confident guess over "I don't know". The full explanation is on the why LLMs hallucinate page.
What is semantic entropy, and when does it fail?
It measures how much sampled answers disagree in meaning. It fails when a model is consistently wrong, because agreement then looks like confidence. A related approach reads internal states, as in steering latents for hallucination detection.
How do you test a RAG system for hallucinations?
Score whether each claim is supported by the retrieved context, and score retrieval quality separately. Tools such as Ragas package these checks.
How do you stop a fixed hallucination from coming back?
Add the failing question to a regression set and run it on every change, as in golden sets and regression testing.
What do guardrails add on top of these fixes?
Guardrails add runtime checks on inputs and outputs, such as blocking an unsupported claim. See LLM guardrails.
- A hallucination is a plausible but false statement; it comes from unpredictable facts and from scoring that rewards guessing.
- Report correct, wrong and abstained shares, and penalize confident errors more than abstentions.
- Retrieval, uncertainty detection and citation checks each lower errors and each has a blind spot.
- Monitor production and turn every failure into a regression test.
- Aim for a falling error rate at an acceptable abstention rate, not zero.
Sources
- Why language models hallucinate — OpenAI, 5 Sep 2025
- Why Language Models Hallucinate (arXiv 2509.04664) — arXiv (Kalai, Nachum, Vempala, Zhang), 4 Sep 2025
- Detecting hallucinations in large language models using semantic entropy — Nature (Farquhar, Kossen, Kuhn, Gal), 19 Jun 2024
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv 2005.11401) — arXiv (Lewis et al.), 12 Apr 2021
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) — NIST, 26 Jul 2024
