# How to Reduce LLM Hallucinations in Your AI Product

> Why LLMs hallucinate and how to reduce it: allow I don't know, quote sources first, verify with citations, run best-of-N and keep humans on high stakes work.

- Canonical: https://zeroraised.com/blog/how-to-reduce-llm-hallucinations
- Published: 2026-11-02T09:26:40Z
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Growth
- Tags: artificial-intelligence, saas, developer-tools, growth

You cannot remove hallucinations from an LLM product, but you can make them rarer, cheaper to detect and less damaging when they happen. OpenAI's research argues that models guess because standard training and evaluation reward guessing over admitting uncertainty ([OpenAI](https://openai.com/index/why-language-models-hallucinate/)). Anthropic's documentation offers a set of prompting and verification techniques, and is explicit that they reduce hallucinations without eliminating them ([Anthropic](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations)).

**TL;DR**

- OpenAI defines hallucinations as plausible but false statements from a language model.
- Its research says training and evaluation reward guessing over acknowledging uncertainty.
- In SimpleQA, a model that abstained more often had a far lower error rate.
- Anthropic recommends allowing "I don't know", quoting sources first and verifying with citations.
- These techniques reduce hallucinations but do not remove them, so validate critical output.

**Disclosure:** ZeroRaised publishes this guide and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 2 Nov 2026; ZeroRaised has no commercial relationship with the companies named.

**Hallucination** is text that is factually incorrect or inconsistent with the context the model was given. **Grounding** is the practice of tying an answer to supplied sources so the model has something specific to quote rather than something plausible to invent.

## Why do LLMs hallucinate?

Standard training and evaluation "reward guessing over acknowledging uncertainty", so a model that always produces an answer scores better than one that sometimes declines ([OpenAI](https://openai.com/index/why-language-models-hallucinate/)).

The multiple-choice analogy makes it concrete: a student who guesses has a chance of being marked right, while leaving it blank guarantees a zero. Models learn the same habit. In the birthday example, a guess has a 1-in-365 chance of being correct.

Some facts cannot be predicted from patterns. OpenAI notes that arbitrary low-frequency facts, like a pet's birthday, "cannot be predicted from patterns alone and hence lead to hallucinations" ([OpenAI](https://openai.com/index/why-language-models-hallucinate/)).

The paper's fix is to penalise confident errors more than uncertainty and give partial credit for appropriate uncertainty. That is a training argument, but it translates into product design: reward saying less when the evidence is thin.

## What does the data say about abstention?

The GPT-5 system card's SimpleQA results, cited in OpenAI's post, show how differently two models can behave on the same benchmark ([OpenAI](https://openai.com/index/why-language-models-hallucinate/)).

| Model | Abstention | Accuracy | Error |
| --- | --- | --- | --- |
| gpt-5-thinking-mini | 52% | 22% | 26% |
| OpenAI o4-mini | 1% | 24% | 75% |

The two models have similar accuracy, 22% and 24%, but wildly different error rates, 26% and 75%. By abstaining about half the time, the newer model made roughly a third as many wrong answers as o4-mini, for almost the same accuracy.

This is the strongest argument for an explicit abstention path. A refusal is a bad experience for one query; a confident wrong answer is a trust problem, and in some categories a liability.

## Which prompting techniques reduce hallucinations?

Anthropic's guidance is a checklist rather than a single trick ([Anthropic](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations)). The techniques worth adopting first:

1. **Allow "I don't know".** Anthropic says explicit permission to say it does not know can drastically reduce false information. Put it in the system prompt and show a visible path for it.
2. **Extract quotes first for long documents.** For inputs over 20,000 tokens, Anthropic recommends pulling word-for-word quotes before answering.
3. **Verify with citations and retract unsupported claims.** Ask the model to cite the passage behind each claim, then drop any claim without a supporting quote.
4. **Use chain-of-thought verification.** Anthropic suggests reviewing the summarised reasoning when an answer looks wrong, which can reveal faulty logic or assumptions.
5. **Use best-of-N comparison.** Generate several answers and compare them. Anthropic notes that inconsistencies between runs can indicate a hallucination.
6. **Refine iteratively.** Feed the draft back with the specific problems found.
7. **Restrict the model to provided documents.** When the answer must come from a contract, a policy or a manual, forbid outside knowledge.

Anthropic's caveat: "although these techniques significantly reduce hallucinations, they don't eliminate them entirely. Always validate critical information" ([Anthropic](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations)).

## How do you build these checks into a product?

Prompting is the first layer. The product is the second, and it is where a small team gets most of the reliability.

**Ground the answer with retrieval.** Fetch relevant passages, pass them in and instruct the model to answer only from them. The [guide to RAG](https://zeroraised.com/blog/what-is-rag) covers retrieval, and the [guide to system prompts](https://zeroraised.com/blog/how-to-write-a-system-prompt) covers the restriction.

**Make the sources visible.** Show the passage beside the claim. A user who can check an answer can catch errors your tests missed. If there is no source, show uncertainty rather than a polished paragraph.

**Validate in code where the output is structured.** If the answer is a number, date, category or record, parse it and check it against rules. The [guide to structured output](https://zeroraised.com/blog/how-to-get-structured-output-from-an-llm) explains how. A schema will not make a fact true, but it stops impossible values.

**Reserve best-of-N for high stakes.** Several answers and a comparison cost more, so use them where a wrong answer is expensive: quotes, prices, legal adjacent text, financial figures.

**Keep a human on irreversible decisions.** The [guide to human approval](https://zeroraised.com/blog/how-to-add-human-approval-to-ai-agents) describes the pattern. A person should confirm anything that moves money, files a document or changes a customer record.

**Measure the failure rate.** Build a small set of questions with known answers, run it after every prompt or model change, and track abstention and error separately. The [guide to evaluating LLM output](https://zeroraised.com/blog/how-to-evaluate-llm-outputs) covers the method.

## What this means for bootstrapped founders

You do not need a research team to ship a more honest product. You need to decide, per feature, whether the model may guess.

1. Add an abstention instruction to every prompt that answers factual questions.
2. Return a visible "not enough information" state instead of filling the gap.
3. Ground answers in retrieved documents and show the source.
4. Validate structured fields in code before they reach a customer.
5. Use multiple samples and comparison for high-value answers.
6. Put a human on money, legal, medical and irreversible actions.
7. Keep a regression set of known questions and re-run it on every model change.

The economics are simple. A hallucination costs support time, refunds and trust, and a bootstrapped product has little of all three to spare. The [guide to reducing LLM costs](https://zeroraised.com/blog/how-to-reduce-llm-costs) covers the other side of the same trade-off: reliability work and cost work are often the same work.

OpenAI argues that accuracy will never reach 100%, but hallucinations are not inevitable, because a model can abstain, and it can be easier for a small model to know its limits ([OpenAI](https://openai.com/index/why-language-models-hallucinate/)). Make uncertainty a first-class outcome, not an error to hide.

## Where ZeroRaised fits

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have shipped an AI product, [submit it here](https://zeroraised.com/submit) and say where it can be wrong and how you handle uncertainty. Honest limitations are a selling point. The [newsletter](https://zeroraised.com/newsletter) carries analysis like this, and the [pricing page](https://zeroraised.com/pricing) explains the listing options. We make no promise of traffic or results.

## Sources

- [Anthropic: reduce hallucinations](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations)
- [OpenAI: why language models hallucinate](https://openai.com/index/why-language-models-hallucinate/)

## Frequently asked questions

### What is an LLM hallucination?

OpenAI defines hallucinations as plausible but false statements generated by language models. Anthropic's docs describe hallucination as text that is factually incorrect or inconsistent with the context the model was given, which means an answer can be wrong even when it sounds fluent.

### Why do language models hallucinate?

OpenAI argues that standard training and evaluation reward guessing over acknowledging uncertainty. A model that always answers looks better on many benchmarks than one that sometimes says it does not know, so it learns to produce a confident guess even when the fact cannot be predicted from patterns.

### Does telling the model to say I don't know help?

Yes. Anthropic's docs say allowing Claude to say I don't know can drastically reduce false information. The SimpleQA results in the GPT-5 system card also show a version that abstained far more often produced a much lower error rate than one that almost never abstained.

### Do citations stop hallucinations?

Citations help because they force the model to point at a source you can check, and Anthropic recommends retracting any claim without a supporting quote. They are not a guarantee, because a citation can still be wrong or misread, so critical information needs validation.

### Can hallucinations be eliminated completely?

No. Anthropic says its techniques significantly reduce hallucinations but do not eliminate them entirely and that critical information should always be validated. OpenAI argues accuracy will never reach 100%, though it also says hallucinations are not inevitable because a model can abstain.