GlossaryAI search
What is an AI hallucination?
Also called: hallucination, LLM hallucination, confabulation
Definition
An AI hallucination is a plausible-sounding statement from a language model that is false, such as an invented feature, price, quote or source, delivered with the same confidence as a correct answer.
AI hallucination, explained
A 2025 paper from OpenAI researchers, "Why Language Models Hallucinate" by Kalai, Nachum, Vempala and Zhang, gives a clear explanation. Like students facing hard exam questions, models sometimes guess when uncertain instead of admitting it. The authors argue this happens because training and evaluation procedures reward guessing over acknowledging uncertainty: most benchmarks score an honest "I don't know" the same as a wrong answer, so confident guessing wins.
In practice, hallucinations cluster around specifics: prices, dates, version numbers, integrations, people's roles, and citations. They're more likely when the model has little reliable information, which describes most young startups. A model may know your category well and still invent your pricing tier or confuse you with a similarly named product.
Retrieval helps. When an assistant searches the web and grounds its answer in real pages, it has something concrete to quote. But retrieval can surface outdated or wrong pages too, and the model can still misread or blend sources. Grounded answers are better, not guaranteed.
What you can do about hallucinations about your own product is mostly unglamorous. Publish the facts plainly on pages crawlers can reach: pricing, plans, integrations, what you don't do. Keep them consistent across your site, directory listings and profiles. Fix or remove outdated pages that contradict the current truth. And check what the main assistants actually say about you on a schedule, so you catch errors early.
If you ship AI features, design for hallucination too: show sources, let users verify, and tune prompts so the model is allowed to say it doesn't know.
Why it matters for founders
An assistant that invents your price or a missing feature can cost you a sale you'll never see. Clear, consistent, crawlable facts are the main defence you control.
Example
An assistant tells a buyer that a startup's tool has no free plan. The startup's pricing page mentions the free tier only inside an image. They add it as text, and later answers get it right.
Common mistakes
- Hiding key facts like prices in images or scripts that crawlers don't read.
- Leaving old pricing pages and outdated listings live.
- Assuming citations mean an answer is correct. Check what the source actually says.
Sources
Checked
Related terms
- Large language model (LLM)A large language model (LLM) is an AI model trained on very large amounts of text to predict the next piece of text. That one skill lets it answer questions, summarize, write and reason in plain language.
- GroundingGrounding is giving an AI model real source content, such as search results or documents, at the moment it answers, so the answer is based on that content instead of only what the model memorized in training.
- Knowledge cutoffA knowledge cutoff is the point in time after which a language model has no training data. Anything that happened or launched later is unknown to the model unless it's retrieved at answer time.
- Retrieval-augmented generation (RAG)Retrieval-augmented generation (RAG) is a technique where a system first retrieves relevant documents, then passes them to a language model so it can answer from that material instead of memory alone.