AI chatbots make things up because a language model doesn’t look facts up. It predicts the most plausible next words based on patterns in its training data. When it knows something well, the plausible answer is usually the true one. When it doesn’t, it still produces something that sounds right, with the same confident tone. These invented answers are called hallucinations.
Every language model does this, including ChatGPT. Smaller models, like the ones that run on your phone, do it more, because they hold less knowledge. You can’t eliminate it, but you can reduce it a lot and learn to spot it.
Why do language models hallucinate? #
A language model is trained to continue text in a way that looks like what it has seen before. It has no built-in sense of “I don’t know.” When you ask about something rare, recent or specific, a few things can go wrong:
- It never learned the fact. Obscure details, recent events and niche topics may not be in its training data.
- It learned it poorly. Something mentioned only a few times gets stored fuzzily.
- It blends similar things. Two people with similar names, two versions of a product, two laws from different years.
- It follows your lead. Ask “why did X happen?” about something that didn’t happen, and it may explain anyway.
- It’s rewarded for sounding helpful. Models are tuned to give complete, confident answers, which can crowd out “I’m not sure.”
Why do small models make more things up? #
A model’s parameters are, loosely, where its knowledge lives. A 4-billion-parameter model on your phone has a small fraction of the capacity of the huge models behind cloud chatbots. It can still write, summarize and reason well about text you give it, but its memory of facts is thinner. Ask it about a minor historical figure or a specific statistic and it’s more likely to fill the gap with invention.
Local models also can’t search the web, so they can’t check themselves against a live source. Anything after their training cutoff is unknown to them, which we cover in can offline AI know current events.
What do hallucinations look like? #
Watch for these patterns:
- Fake citations. Book titles, paper names, URLs and court cases that don’t exist.
- Precise-sounding numbers. “Studies show 73% of…” with no source.
- Wrong dates and names. Especially for less famous people and events.
- Invented features. Describing settings or menus in an app that aren’t there.
- Confident arithmetic errors. See can local AI do math.
- Made-up quotes. Attributed to real people who never said them.
- Details in images that aren’t there. Vision models can describe objects or text that don’t exist in the photo.
Seven ways to reduce AI hallucinations #
1. Give it the source #
The most effective fix. When the model answers from text you provide instead of from memory, it has far less to invent. Paste the article, or attach the document. In Personal LLM, attaching a PDF, text or Markdown file makes the model answer from the matching passages and show which ones it used, so you can check the answer against the source. See how to chat with a PDF offline.
2. Lower the temperature #
Temperature controls randomness. Higher values make answers more creative and more likely to wander. For facts, use a low setting. Personal LLM’s Precise preset uses a temperature of 0.3, compared with 0.7 for Balanced and 1.2 for Creative.
3. Turn on thinking mode for reasoning #
Writing out steps before answering helps the model catch logical and math errors. It doesn’t add knowledge the model lacks, so it helps with reasoning more than with facts.
4. Tell it that “I don’t know” is allowed #
Add this to your question or system prompt:
If you’re not sure or don’t know, say so. Don’t guess or invent sources.
It won’t make a model honest every time, but it shifts the default.
5. Ask narrower questions #
“What year was the Treaty of Versailles signed?” is safer than “Tell me everything about the Treaty of Versailles.” Long, open answers give more room for invented details.
6. Switch to a bigger model for hard questions #
A larger model knows more. In Personal LLM you can start on a fast model and switch to Qwen 3.5 9B from the chat header for the questions that matter, if your phone has the RAM.
7. Ask it to check itself #
“Review your answer. Which parts are you least sure about?” often surfaces the weak spots. Treat that as a hint, not a guarantee.
How to fact-check an AI answer #
- Check anything you’ll act on. Health, money, legal, safety: always confirm with an official source or a professional.
- Search for specific claims. Names, numbers and quotes are easy to verify.
- Don’t trust citations blindly. A small offline model can’t browse, so any URL or paper it gives you may be invented. Look them up.
- Compare against the document. If it’s answering about a file, read the passage it cites.
- Be most skeptical of oddly specific details. A precise statistic with no source is a classic hallucination.
Are cloud chatbots better? #
On facts, usually yes. Big cloud models know more, and many can search the web and cite real pages. They still hallucinate, especially on niche topics, and web search brings its own errors from bad sources.
| Small local model | Large cloud model | |
|---|---|---|
| General knowledge | Thinner | Broader |
| Recent events | None after cutoff | Can search the web |
| Answering from your document | Good | Good |
| Privacy | Stays on your phone | Sent to a server |
| Hallucination risk | Higher on facts | Lower, not zero |
The practical split: use local models for writing, summarizing, brainstorming and questions about your own text. Use a cloud model or a search engine for facts you need to be sure of. Setting clear expectations in a system prompt helps either way; our system prompt examples include a fact-careful one.
Frequently asked questions #
What is an AI hallucination? #
It’s when an AI states something false as if it were true, such as an invented fact, quote, citation or number. It happens because language models generate plausible text rather than retrieving verified facts.
Why do AI chatbots make up sources? #
A citation is just another pattern to a language model. It knows what a real citation looks like, so it can produce one that looks right but doesn’t exist. Always look up sources an AI gives you.
Do offline AI models hallucinate more than ChatGPT? #
Generally yes, on factual questions, because smaller models hold less knowledge and can’t search the web. When answering from a document you provide, the gap is much smaller.
Can you stop AI from hallucinating completely? #
No. You can reduce it by giving the model source material, lowering the temperature, asking narrower questions and letting it say “I don’t know.” Checking important answers is still necessary.
Does thinking mode prevent hallucinations? #
It reduces reasoning errors, like math slips and logic mistakes, by letting the model work step by step. It doesn’t give the model facts it never learned, so factual hallucinations can still happen.