An AI hallucination is when a model states something false with total confidence and gives you no signal that it made it up. It is not a bug in the usual sense. The model is doing exactly what it was built to do: predict the next plausible token, not check whether the token is true. That gap is why a lawyer cited six court cases that never existed, why a support bot invented a refund policy its own company had to honor, and why your assistant confidently misremembers a fact you told it last week.
Below are real, documented AI hallucination examples, grouped by what actually caused each one. Once you can see the cause, you can predict where the next one will happen and cut most of them off before they reach a customer, a filing, or a decision.
Key takeaways
- A hallucination is a confident, false, unsignaled output. The model has no built-in "I do not know" reflex unless you give it one.
- Most real cases trace to one of five causes: missing context, stale context, ambiguous prompts, over-eager pattern completion, or the model inventing sources to satisfy a request.
- The single largest cause for personal and business use is missing or stale context. The model fills the gap with a plausible guess instead of your real facts.
- Fixes fall into two buckets: constrain the model (temperature, "say you do not know," ask for citations) and feed it real, current context so it does not have to guess.
- Retrieval-based grounding beats bigger models for factual accuracy. A smaller model with the right facts in front of it hallucinates less than a huge model guessing from memory.
---
What an AI hallucination actually is
A hallucination is any output the model presents as fact that is not supported by its input or by reality, delivered with the same fluent confidence as a correct answer. The important word is confident. The model does not hedge, does not flag uncertainty, and does not distinguish "I read this" from "this is statistically likely to be the next word."
There are two broad families:
- Intrinsic hallucination: the output contradicts the source you gave it. You paste a contract and it summarizes a clause that is not there.
- Extrinsic hallucination: the output cannot be verified from any source. You ask for research and it invents a citation, a statistic, or a quote that exists nowhere.
Both come from the same root: the model is a probability engine over text, not a database of verified facts. When the probable token and the true token diverge, and nothing forces the model to notice, you get a hallucination.
---
The famous ones: examples that made the news
These are the AI hallucination examples people actually cite, because they had consequences.
The fake legal citations (Mata v. Avianca, 2023). Two New York lawyers filed a brief containing six judicial decisions that did not exist. They had asked a chatbot for supporting case law, and it produced case names, docket numbers, and quotes, all fabricated. When opposing counsel could not find the cases, the lawyers asked the same bot to confirm the cases were real, and it said yes. The court sanctioned them. Cause: extrinsic invention. The prompt asked for cases that supported a position, and the model produced text shaped like real citations because that is what "supporting case law" looks like statistically.
The airline chatbot that invented a policy (Air Canada, 2024). A customer asked Air Canada's support bot about bereavement fares. The bot described a refund process that did not match the airline's actual policy. A tribunal ruled the airline was liable for what its bot said. Cause: missing context. The bot was not grounded in the current, correct policy, so it generated a plausible-sounding one.
The search summary that told people to eat rocks and glue. A widely covered AI search feature surfaced answers advising users to add glue to pizza and to eat a small rock per day, both traceable to satirical or joke sources the system treated as factual. Cause: bad source grounding. The model retrieved real text but could not weigh a joke forum against a medical authority.
Invented academic references. Researchers repeatedly document that when asked for citations, general-purpose models produce references with real-looking authors, journals, years, and DOIs that do not resolve to any actual paper. Cause: extrinsic invention plus over-eager completion. A citation is a highly patterned string, so the model can generate a perfect-looking fake.
---
The funny ones (that are still warnings)
Funny AI hallucination examples are useful because the same mechanism causes the dangerous ones.
- A model confidently explains the plot of a book that was never written, with character names and a three-act summary, because the title you invented sounds like a real book.
- Asked how many times the letter "r" appears in a word, models have famously miscounted, because they reason over tokens, not letters.
- A model gives a detailed, wrong biography of a real person, mixing two people with the same name into one profile.
- Ask for a "list of quotes by X" and you get elegant lines that X never said, because a quote is an easy pattern to fabricate.
The lesson is identical across all of them: the model optimizes for plausible, not true, and it will not warn you when those two part ways.
---
The real cause behind most everyday hallucinations
The headline cases are dramatic, but the hallucinations that hurt you day to day are quieter. Your assistant tells a client your old pricing. It references a feature you deprecated. It attributes a decision to the wrong project. None of these make the news, and all of them come from the same place: the model does not have your real, current context, so it guesses.
This is the "garbage in, generic out" problem stated precisely. Output quality is capped by input quality. A model with no access to your actual facts fills the gap with the average of everything it saw in training. That average is often plausible and almost always wrong for you specifically.
There are two sub-causes worth separating:
- Missing context: the fact was never given to the model. It never knew your pricing, so it invented one.
- Stale context: the fact was given once but has since changed. You told it your price in March, you raised it in June, and it still quotes March because the old note was never updated.
Stale context is the sneakier of the two, because the answer looks grounded. It cites a real fact that used to be true. Most personal knowledge setups make this worse, not better: a static note in Obsidian or Notion is correct the day you write it and silently rots the moment the underlying reality changes. Nothing tells the model the note went out of date.
---
AI hallucination examples by cause
| Example | Cause | What would have prevented it |
|---|---|---|
| Fake court cases in a legal brief | Extrinsic invention (asked for supporting cases) | Require verifiable citations; ground on a real case database |
| Airline bot invents a refund policy | Missing context | Feed the bot the current, correct policy |
| Search answer: eat a rock a day | Bad source grounding | Weight authoritative sources over joke forums |
| Invented academic references with fake DOIs | Over-eager pattern completion | Ask for sources it can quote from provided text only |
| Miscounting letters in a word | Token-level reasoning limit | Use a tool call, not the model's raw guess |
| Assistant quotes your old pricing | Stale context | Keep the fact current and superseded when it changes |
| Assistant cites a deprecated feature | Stale context | Update the source of truth automatically |
| Confident wrong bio of a real person | Entity confusion in training data | Provide the correct profile as grounding |
Read the middle column top to bottom and a pattern appears. A handful of causes produce nearly every hallucination, and most are about the context you gave the model, not the model's raw intelligence.
---
How to actually reduce hallucinations
There are two levers. Use both.
Lever 1: constrain the model
These reduce the odds the model wanders off, without changing what it knows.
- Tell it to say "I do not know" when it is unsure. Models will hedge far more often when explicitly permitted to.
- Ask for citations, then check them. If it cannot point to a source in the text you gave it, treat the claim as unverified.
- Restrict answers to provided material: "Answer only from the document below. If it is not there, say so."
- Lower randomness for factual work. High creativity settings raise hallucination rates.
- Use tools for anything mechanical (counting, math, dates, lookups) instead of trusting the model's guess.
A copyable instruction you can paste into a system prompt or custom instructions:
When you state a fact, cite where it came from in the material I provided. If a fact is not in that material and you are not certain, say "I am not sure" instead of guessing. Never invent citations, quotes, statistics, or sources. If I ask for something you cannot verify, tell me what you would need to verify it.
Lever 2: feed it real, current context
Constraining the model stops it from wandering. Grounding it in your actual facts stops it from needing to guess in the first place. This is the bigger lever for personal and business use, and it is the one most people skip.
Grounding means the model answers from your real material rather than from its training average. In practice that is retrieval: pull the relevant facts about you, your product, your customers, and your decisions, and put them in front of the model at answer time. A smaller model with the right facts in front of it beats a larger model guessing from memory. See the deeper argument in how to give your AI a memory that actually lasts.
The hard part is not retrieval. It is keeping the retrieved facts current. A grounding source that goes stale just launders old errors into confident-sounding answers. This is where the "stays current" mechanism matters: when a fact changes, the old version has to be marked superseded and the new one takes its place, with the history preserved so nothing silently contradicts itself.
---
Comparing the two levers
| Approach | What it fixes | What it misses |
|---|---|---|
| Prompting tricks (say "I do not know," ask for citations) | Reduces confident guessing on the spot | The model still has no access to your real facts |
| Lower temperature | Cuts creative drift on factual tasks | Does nothing about missing or stale context |
| Bigger or newer model | Slightly better general recall | Still guesses about you specifically; can hallucinate more fluently |
| Retrieval grounding on current context | Removes the need to guess about your world | Only as good as how current your source is |
The pattern: prompting and model choice help at the margins, but they cannot supply facts the model never had. Grounding on current context is the only lever that addresses the largest cause, missing and stale facts. Curated fact bundles help here too. A Memory Pack lets you inject a validated set of facts and playbooks about a domain so the model answers from vetted material instead of improvising.
---
Where a living context layer fits
If most of your hallucinations come from missing or stale context, the fix is a context layer that (1) is built from what you already produce, (2) keeps itself current when facts change, and (3) is available to your AI tools at answer time.
That is what Locul does. It builds a searchable second brain from your own material, distills it into memories tagged as facts, decisions, and preferences with confidence scores, and marks old facts superseded when they change so your AI stops quoting last quarter's price. It serves that brain to your AI tools over MCP, and it runs local-first, so you can pair it with a local model and keep everything on your machine. It is free to start, 500 memories, local AI, no credit card. Pricing tiers are on the pricing page.
Grounding does not make hallucinations impossible. Nothing does. But it removes the single largest source of them for real work: the model guessing about facts it should not have had to guess about.
---
FAQ
What is the most common cause of AI hallucinations?
For everyday and business use, the most common cause is missing or stale context. The model was never given the correct fact, or was given it once and the fact later changed, so it fills the gap with a plausible guess. This is different from the flashy news cases, which are usually extrinsic invention (the model fabricating sources to satisfy a request). Both trace back to the same root: the model predicts likely text rather than verified fact. Grounding it in current context, as covered in how to give your AI a memory that actually lasts, removes most of them.
Are AI hallucinations getting better with newer models?
Newer models reduce some categories of error, but they do not eliminate hallucination, and a more fluent model can make a wrong answer harder to spot. No model can supply a fact it never had access to. That is why factual accuracy tracks the quality and freshness of the context you feed the model more closely than it tracks model size. A smaller model grounded in your real, current facts often beats a larger one guessing from memory.
Can you fully stop an AI from hallucinating?
No. Because the model generates probable text rather than checking truth, you cannot reduce the rate to zero. You can reduce it substantially by combining two levers: constrain the model (permit "I do not know," require citations, restrict answers to provided material, lower randomness) and ground it in real, current context so it does not need to guess. Treat any factual claim without a checkable source as unverified until you confirm it.
Why do AI models invent fake citations and sources?
A citation is a highly patterned string: author, year, journal, page, DOI. That structure is easy for a model to reproduce even when no real paper exists, because the model is matching the shape of a citation, not retrieving a real one. When a prompt asks for "sources that support X," the model produces text that looks like supporting sources. The fix is to make it cite only from material you provided, and to verify every reference before you rely on it.
What is the difference between a hallucination and a factual error?
A plain factual error is any wrong statement. A hallucination is a specific kind: a confident, fluent, false statement delivered with no signal that the model is uncertain or that it made the information up. The danger is the missing uncertainty signal. A model that said "I am not sure" would be far less harmful than one that invents six court cases and then insists they are real when asked to double-check.
Does keeping my notes updated actually reduce hallucinations?
Yes, if the model can read those notes at answer time. Grounding on your own current material means the model answers from your real facts instead of its training average. But there is a catch: a note that has gone stale grounds the model in an old fact and produces a confident, wrong answer that looks sourced. That is why the freshness of the grounding source matters as much as its existence. A context layer that marks facts superseded when they change keeps the grounding honest, so you can also inject a curated Memory Pack of vetted facts for a domain.
FAQ