Why AI Hallucinates: What the Research Says and How to Reduce It in Real Work

An AI hallucination is a confident answer from a model that is false. It can be a made-up fact, a fake citation, or a real source that never said what the model claims. It happens because training and scoring reward a guess over "I don’t know." You can lower the risk a lot. No vendor says you can remove it.

Our guide to what a large language model is covers the basics. This piece goes deeper on why AI hallucination happens and what reduces it. It covers the types, why models guess, what grounding and citations fix, and two review checklists.

The short version: a model writes likely text, and a wrong answer can be likely text. Training and benchmarks have taught models that a guess scores better than a blank. Give the model your sources and let it say it does not know. Make every claim checkable. Keep a person on anything public, financial, legal or medical.

The types of AI hallucination to watch for

OpenAI’s research team defines hallucinations as "plausible but false statements generated by language models." A wrong answer usually reads just like a right one.

A widely cited survey by Ji and colleagues, Survey of Hallucination in Natural Language Generation, splits the problem in two. It appeared in ACM Computing Surveys in 2023. An intrinsic hallucination is output "that contradicts the source content," in the survey’s words. An extrinsic one "cannot be verified from the source content."

If a model summarizes your annual report and gets a number wrong, that is intrinsic, and the report catches it. If it adds a statistic the report never mentioned, that is extrinsic. Check it elsewhere or cut it.

Wrong sources are the easiest to miss. The link works and the outlet is real, so the reviewer moves on.

Why models guess instead of saying "I don’t know"

The 2025 paper Why Language Models Hallucinate gives the clearest current answer. Its authors are Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang. The paper lists three at OpenAI and one at Georgia Tech.

Their argument starts with a school exam. "Students may guess on multiple-choice exams and even bluff on written exams," the paper notes. If a right answer earns a point and a blank earns zero, guessing scores better on average. Most AI benchmarks grade the same way. That is why, the authors write, models are "optimized to be good test-takers" and guess when unsure.

The paper opens with a small test. The authors asked a model, DeepSeek-V3, for Kalai’s birthday and told it to answer only if it knew. In three tries it gave three different wrong dates.

The paper also explains why some facts are harder than others. A fact that shows up only once in training data gives the model little to learn from. The authors put a number on it. If 20% of birthday facts appear only once in pretraining data, base models should miss at least 20%.

That fits what our readers see. Models do well on famous facts and poorly on local ones. Think of a small nonprofit’s founding year, a board member’s title or a grant deadline. Those are the facts nonprofit communications depend on, so that is where AI hallucination does the most harm.

Why AI hallucination is not a random glitch

OpenAI’s summary of the paper, published September 5, 2025, makes three points. The causes are understood: statistical pressure in training, reinforced by scoring. It is not inevitable, because a model can decline when unsure. As the post puts it, "errors are worse than abstentions." And 100% accuracy is out of reach. Some questions cannot be answered from any data the model has.

The authors want benchmarks to score a wrong answer worse than a blank. That is the vendors’ job. Yours is to apply the same rule to your own prompts and reviews.

What the numbers show, as of 2025 testing

This is the one dated section of this piece. OpenAI’s summary reprints figures from its GPT-5 system card comparing two models on SimpleQA, one of OpenAI’s accuracy tests.

Both models got about the same share right. The one that declined more than half the time gave far fewer wrong answers. On a leaderboard that counts only right answers, the older model looks slightly better.

These are OpenAI’s own figures on its own test, and both models are older than current ones. Treat the table as a picture of the trade-off, not a ranking.

Grounding: give the model the source

In client work, AI hallucination often starts with asking a model for facts it does not have. The fix is grounding: put the source in front of the model and have it answer from that.

• Attach the document. Anthropic’s guide to reducing hallucinations suggests telling the model to use only the documents provided, not its general knowledge.
• Retrieval (RAG). A search step finds the right passages in your content and adds them to the prompt. See What Is RAG? and our pgvector on Supabase build.
• Search grounding. Google’s Grounding with Google Search connects Gemini to live web results. Google lists this benefit: "Reduce model hallucinations by basing responses on real-world information."

Grounding can go wrong too. OpenAI’s guide to optimizing LLM accuracy has a warning. Retrieval can "supply too much irrelevant context, which drowns out the real information and causes hallucinations," it says. In one example there, adding RAG lowered accuracy by four points.

Feed the model the pages that answer the question, not the whole site. Our post on context windows covers why long prompts lose detail.

What citations check, and what they leave to you

Asking for sources turns a claim you would have to trust into a claim you can check.

Anthropic’s guide suggests having the model cite a supporting quote for each claim. Claims without one get retracted. Anthropic’s API also has a Citations feature for documents you supply. Anthropic says its citations "are guaranteed to contain valid pointers to the provided documents."

That guarantee covers the pointer. It goes to a real passage in your document. A reviewer still has to read the passage and confirm it supports the sentence. Web search citations need the same check.

Prompts and settings that lower AI hallucination

These steps work on any vendor’s model. They come from Anthropic’s guide unless noted.

• Allow "I don’t know." Tell the model it may say it does not know. Anthropic says this simple step can sharply reduce false information.
• Quotes first for long documents. Over about 20,000 tokens, have the model pull exact quotes first. Then it answers only from them.
• Set a confidence bar. Kalai and colleagues suggest stating the stakes in the prompt. For example: answer only if you are more than 75% sure. A wrong answer costs two points, and "I don’t know" costs nothing.
• Run it more than once. Anthropic suggests running the same prompt several times and comparing. Answers that change between runs deserve a closer look.
• Ask for a claims list. This one is our suggestion. Have the model list every name, number, date and quote in its draft. That list becomes the reviewer’s checklist.

These techniques reduce hallucinations but "don’t eliminate them entirely," Anthropic says.

Match the use to the stakes

Some uses carry little AI hallucination risk because they add no new facts: rewriting your own paragraph, tightening a headline, outlining a page, brainstorming event names. A bad idea just gets thrown out.

Grant reports, donor letters, impact numbers, member help pages and anything legal, financial or medical are the opposite. They depend on facts being right.

OpenAI’s accuracy guide shows how to decide. In one support example, it says 85% accuracy "could be ok" if the misses were mostly early escalations. Where mistakes cost more, the team decided "the human would drive and the AI would function as an assistant," per the guide.

Ask the same two questions about each task. What does a wrong answer cost? Who catches it before it ships? Our post on AI implementation covers building that review step into a rollout.

AI hallucination checklist for nonprofit communications

Use this before any AI-assisted draft goes to donors, members, funders or the press.

1. Every number traces to a source. Impact figures, budgets, dates and percentages match your own records.
2. Every name and title is right. Board, staff, partners, funders and honorees. Google’s fake author warning shows invented credentials also hurt search.
3. Every quote is real. The person said it and approved it.
4. Every link supports its sentence. Read the passage, not just the page title.
5. Program details are current. Check eligibility, deadlines, prices and contacts against the live source.
6. Nothing was added. Cut any claim your source document does not contain.
7. A named person signs off. Someone who knows the subject reads the final version.

AI hallucination checklist for agency deliverables

Agencies carry a second risk: the client trusts that you checked.

1. Client facts come from the client. Dates, services, bios and stats come from the client’s own materials.
2. Case studies are real. Proposals and RFP responses never include an invented project, client or result.
3. SEO copy makes no new claims. Titles and meta descriptions say only what the page says. See writing for AI citation.
4. Code references exist. Check every function, hook, package and API field the model names against current docs.
5. Migrated content matches the original. Spot-check figures when a model rewrites old pages.
6. The review is recorded. Note who checked what.

Frequently Asked Questions