Home Hub How It Works Features Use Cases How-To Guides Help Docs Pricing Login
AI

What Are AI Hallucinations? A Field Guide With Examples

Radomir Basta • November 13, 2026 • 21 min read
Chess king at the head of a boardroom table, with two highlighted reports in the foreground

AI hallucinations are statements an AI model presents as fact that are false, fabricated, or unsupported by the source they are attributed to. Common examples are an invented statistic, a court case that never existed, and a summary that quietly changes a number in the document you uploaded. All of them arrive in the same calm, fluent prose as the correct parts of the answer, and that is the whole problem. Confident wording comes from the language. Evidence belongs to the claim, and you have to look for it separately.

Any AI answer raises two questions. Are its claims true? Are they faithful to the material they say they come from? This guide covers both. It uses annotated examples to show where support breaks down and explains why capable models still produce these errors. It then gives a practical way to decide which claims deserve your checking time before an answer goes into a report, a client deck or a decision.

In this guide:

What Does AI Hallucination Mean?

The term comes from human psychology, and it misleads a little. A language model does not perceive things that are not there, and it does not lie. Lying requires an intention to deceive, and a model has no private knowledge of the truth that it is hiding from you. “Hallucination” describes the output: text that asserts something the available evidence does not support.

NIST’s Generative AI Profile calls the same problem “confabulation”: confidently stated but erroneous or false content.

The definition is useful because it separates two things people tend to blur together:

  • Factual correctness: does the claim match the world? “The company was founded in 2011” is either true or false regardless of what document you gave the model.
  • Source faithfulness: does the claim match the material the answer says it is based on? A summary can state something true about the world and still be unfaithful, because the document it is summarizing never said it.

That gives three distinct ways an answer can hallucinate. It can contradict reality. It can misstate a document you supplied. Or it can attribute a claim to a source that does not make it, which covers both invented citations and real citations attached to the wrong claim.

The definition also has a boundary that matters in practice. A claim without a citation is not automatically a hallucination. Plenty of uncited statements are perfectly true. A claim nobody has checked yet sits in its own category, unverified, and should stay there until someone checks it. Labeling everything uncited as hallucinated makes people stop taking the word seriously. Treating everything uncited as fine is how fabrications get through.

Hallucinations are not limited to text. Transcription tools can insert sentences nobody spoke, and image generators produce details that were never in the brief. This guide stays with text, because that is where most professional research, summaries and recommendations live.

AI Hallucination Examples: What the Error Actually Looks Like

Abstract definitions only go so far. Here is a hypothetical source document and three summaries of it, one accurate and two not.

Hypothetical supplier note: Northvale Packaging has 40 employees and offices in Bristol and Leeds. It was founded in 2011 and supplies recycled corrugated packaging to food retailers. Its standard lead time is ten working days.

Accurate summary: Northvale Packaging is a 40-person supplier of recycled corrugated packaging for food retailers, founded in 2011, with offices in Bristol and Leeds and a standard lead time of ten working days.

Source-contradicting summary: Northvale Packaging is a 60-person supplier of recycled corrugated packaging, founded in 2011, with offices in Bristol and Manchester and a standard lead time of ten working days.

Summary with an invented result: Northvale Packaging is a 40-person supplier of recycled corrugated packaging for food retailers, founded in 2011, with offices in Bristol and Leeds. According to the note, it cut lead times by 25% last year after moving to a new warehouse system.

The second summary is easy to catch if you compare it with the note. The third is the dangerous one. Every detail you might spot-check first is correct: the headcount, the cities, the founding year. Once a few facts check out, most readers stop checking. The invented sentence borrows credibility from the true ones around it, and it even claims the note as its source.

The third summary also shows a distinction worth keeping. Maybe Northvale really did cut lead times last year. The claim might be true somewhere in the world. It is still a hallucination in this summary, because the source does not report it. “The source does not say this” and “this cannot be true anywhere” are different findings. A faithfulness check settles the first. Only outside evidence settles the second.

The same pattern turns up across the kinds of work professionals hand to AI. All five examples below are hypothetical.

Hypothetical output Type of problem Evidence that resolves it
“The contract summary confirms a 60-day termination notice period.” Source-contradicting summary. The contract says 30 days. The clause itself, read in the original document.
“72% of mid-sized manufacturers now use AI in procurement.” Invented statistic. Precise, plausible, and with no traceable origin. The named survey, its publisher, sample and question wording. No traceable survey means the number is unusable.
“See Harlow & Mendes (2021), Journal of Supply Chain Resilience, vol. 14.” Fabricated citation. Authors, journal and volume form a convincing pattern that points to nothing. A search of the publisher’s site or a library database for the exact paper.
“Remote teams are more productive, as a major consultancy’s 2023 workplace report shows.” Real source, unsupported claim. The report exists but measures adoption rates, not productivity. The cited passage. Existence of the source proves nothing until you read what it says.
“You can sync directly to SAP using the CRM’s built-in connector.” Nonexistent software capability. The model has generalized from similar products. The vendor’s current documentation or release notes.

Look at the right-hand column. Every resolution requires going outside the answer itself. Rereading the answer, or asking the model whether it is sure, does not appear anywhere in that column.

These errors are not hypothetical in the real world either. In the 2023 case Mata v. Avianca, a US federal judge sanctioned lawyers who had filed a brief citing court decisions generated by ChatGPT that did not exist. The citations had the right format, plausible case names and internal quotations. They looked like legal research, and nobody had checked them against the record. We cover the safeguards in our guide to fabricated citations in legal work.

Why Does AI Hallucinate?

Chess knight beside an AI-written supplier summary marked up in red against the source document

Large language models learn statistical patterns from enormous amounts of training data: how words, facts, arguments and document structures tend to fit together. When you ask a question, the model generates a response that continues your prompt in a way consistent with those patterns and with whatever context it has been given. Modern models do much more than simple autocomplete. They can reason through multi-step problems, call search tools, run code and work across long documents. But at the moment a sentence is produced, being fluent and plausible is what the process aims for. Truth gets checked only if something in the process actually checks it.

That is why a model can write “Harlow & Mendes (2021)” with the same ease as a real citation. It has seen thousands of citations, and it knows what one looks like. Producing a plausible one is a pattern task. Confirming that a specific paper exists is a different task, and it only happens if something in the workflow performs it.

What pushes a model toward invention

  • Missing knowledge. The answer is not in the training data or the supplied context, often because the topic is obscure, recent or highly specific. The model fills the gap with something shaped like an answer.
  • Ambiguous requests. “Summarize the supplier’s performance” leaves the model to decide what performance means and how much detail you expect. The gaps get filled with plausible material.
  • False premises. Ask “Why did the regulator fine Northvale in 2022?” and many answers will explain a fine that never happened, because the question presupposes it.
  • Pressure to be complete. Models are trained and evaluated in ways that tend to reward a full answer over “I don’t know.” OpenAI’s 2025 research paper Why Language Models Hallucinate argues exactly this: common evaluation methods reward guessing over acknowledging uncertainty, in the same way that a multiple-choice exam rewards a student who guesses on every question.

When the source is right there and the answer still drifts

Supplying the right document helps a great deal, but it does not close the gap. Source-use failures have their own recognizable shapes:

  • Blending passages: a figure from section 2 attached to the subject of section 5.
  • Dropping qualifiers: “in the pilot region, under ideal conditions” becomes a general claim.
  • Misattribution: a statement from one of three uploaded files credited to another.
  • Overreaching conclusions: the document reports two facts, and the answer presents an inference from them as something the document states.

Context windows and retrieval

A model’s context window is the amount of text it can take into account at once: your prompt, the conversation so far and any documents loaded into it. Material outside that window plays no part in the answer. Material inside it can still be skimmed, weighted poorly or blended with patterns from training. Uploading a file means the model can draw on it. It does not mean every sentence the model writes is backed by that file.

Retrieval-augmented generation (RAG) is the technique behind most “chat with your documents” and AI search features. Rather than relying on memory, the system first searches a document store or the web for relevant passages, then hands those passages to the model to write from. It is one of the most effective forms of source grounding available, and it adds new places for things to go wrong:

  • The search retrieves the wrong passage, such as last year’s pricing page instead of this year’s.
  • The retrieved material is stale, so the answer is faithful to an outdated source.
  • The right passage is retrieved and then misread, summarized loosely or combined with model memory.

Reasoning and tool use reduce some hallucinations and change the shape of others. A model that searches the web can still cite a page that does not support its sentence. A model that reasons step by step can still build a careful argument on a premise it invented two steps earlier. No single capability makes every claim reliable, which is why checking stays part of the job.

Is Every Wrong AI Answer a Hallucination?

No. Lumping every error together makes it harder to fix any of them. The neighboring categories each call for a different response:

  • Outdated information. A model describes a regulation as it stood before an amendment. The information was true once. It becomes a false claim when presented as current, but age alone does not make it fabricated. The fix is a current source.
  • Arithmetic and reasoning mistakes. A margin calculated wrongly from correct inputs is an error in the working. It can be found by redoing the calculation, with no outside research needed.
  • Opinion and judgment. “This supplier is the better choice” is a recommendation. It can be poorly reasoned without containing a single fabricated fact.
  • Bias. An answer can be skewed by what the training data overrepresents while each individual statement in it is defensible.
  • Explicit uncertainty. “I couldn’t find evidence of this” or “this is my best estimate” is the model doing its job well.
  • Requested fiction. Ask for a sample customer story and you get an invented one. That is the requested work, not a hallucination, until someone pastes it into a case study as if it were real.

These categories overlap. An unsupported recommendation can rest on a faulty assumption, such as “Northvale can absorb your volume” when nothing in the note addresses capacity. There is no fabricated fact in that sentence, but the gap is just as real, and the check is the same: what is this resting on?

Two more points belong here. First, confident wording is not a probability. A model saying “definitely” has not computed a high chance of being right, and a hedge is not a measured low one. Second, disagreement between two answers is a signal to investigate. It does not prove either answer wrong, and it is often where the most useful information is.

Why Hallucination Rates Vary So Much

Hallucination rates depend on the task, the topic, how much relevant evidence the model can access, whether it uses search or retrieval tools, how it is configured, and above all how errors are counted. Summarizing a short supplied document is a very different test from answering obscure factual questions from memory, and a model can look excellent on one and weak on the other.

Accuracy and hallucination rate are different measures, and the gap between them can be wide. On AA-Omniscience, a benchmark of hard knowledge questions, Claude Fable 5.1 answers more questions correctly than any other model on our board. On the questions it does not get right, it still gives a wrong answer instead of declining 70.0% of the time. The lowest rate among public models that answer at least 45% correctly is 4.3%, from Gemini 4 Argon. Our page on which AI hallucinates the least tracks who holds that spot.

Counting method matters more than most headlines admit. An answer with twenty claims and one invented figure is one hallucinated claim out of twenty. Scored by whole answers, it counts as a failed answer. Both measures are legitimate, and they produce very different-looking rates. That is also why a benchmark result cannot predict whether the specific answer in front of you is right. It describes behavior across a test set, not the response you are about to paste into a report.

We keep a running collection of the public research at AI hallucination rates and benchmarks. It brings the major public hallucination benchmarks together with our own data, the Multi-Model AI Divergence Index, which tracks how often frontier models give materially different answers to the same question.

Keep one distinction in mind when reading any of it: divergence measures disagreement, not error. When two models disagree, at least one of them is incomplete or wrong, or the question has more than one reasonable answer. When they agree, they may both be right or both share the same mistake. Divergence tells you where to look. Evidence tells you what you found.

Why AI Hallucinations Matter in Everyday Work

Chess knight beside a desk of annotated reports under a brass lamp

The same fabricated figure can be trivial or serious depending on where it ends up. Three hypothetical situations show the range.

Watch this video about what are ai hallucinations:

Video: Why do AI models hallucinate?

An internal research summary. An analyst asks for a quick overview of a market before a team discussion, and it includes an invented growth rate. If the summary only shapes which questions the team asks, the cost of the error is low and the discussion will probably expose it.

A client-facing report. The same growth rate goes into a deck sent to a client, credited to “industry research.” Now the error carries your firm’s name, can be quoted onward, and costs credibility if the client checks it.

This has already happened at the top of the market. In 2025, Deloitte Australia agreed to partially refund the A$440,000 the Australian government paid for a report that, according to the Associated Press, contained a fabricated quote from a federal court judgment and references to research papers that do not exist. The revised report disclosed that a generative AI system had been used to write it.

A legal or investment question. A model summarizes the change-of-control clauses across a target company’s contracts and misses a qualifier in one of them. That error can shape a deal price or a negotiating position, and it may not surface until it is expensive to undo.

The harder risk is propagation. A fabricated number gets copied from the AI summary into a briefing note, from the note into a board pack, and from the board pack into next quarter’s plan. With each copy it loses its original context, and nobody remembers that it came from an unverified chat answer. By the third document it carries the authority of the earlier ones, and questioning it starts to feel like questioning the whole team’s work.

The stakes also change how often an answer gets challenged. Our Multi-Model AI Divergence Index put 17,885 real questions to GPT, Claude, Gemini, Grok and Perplexity, with a blind AI judge checking the answers. Corrections came on 45% of high-stakes questions, against 28% of low-stakes ones.

So the right amount of checking is proportionate, not uniform. Three questions set it:

  • Impact: what does it cost if this specific claim is wrong?
  • Reach: how many people or documents will rely on it?
  • Reversibility: can the decision be undone cheaply once the error is found?

Low on all three, a quick read is enough. High on any of them, the claim needs real evidence. For consequential legal, medical or financial work, part of that evidence is a qualified specialist reading the output. Good AI output makes their review faster. It does not replace their judgment.

How to Reduce AI Hallucination Risk

Hallucinations cannot be prevented, only mitigated to an acceptable level. What counts as acceptable depends on the use case. A brainstorm can tolerate a few invented details. A regulatory filing cannot tolerate any. Better prompts and newer models reduce the rate, but no current approach removes the risk, so a working method has three stages: constrain the task before the model writes, examine the claims after it writes, and verify the ones that matter before anyone uses them. Our hallucination mitigation guide covers the wider approach.

1. Constrain the task before generation

Give the model the relevant source material instead of asking it to recall it. Ask it to keep source-backed facts separate from its own inferences, and explicitly permit it to say that evidence is missing. That permission matters. Without it, the pressure to give a complete answer works against you. Narrow, specific questions also help, because a vague request leaves the model more gaps to fill.

A prompt you can reuse:

“Using only the attached material, answer in four parts. 1) Facts the source states directly, each with the passage it comes from. 2) Inferences you are drawing that the source does not state. 3) Information I asked for that the source does not contain. 4) Any claims I should verify against an external source before using them. If the material does not answer the question, say so.”

This prompt organizes the answer. It does not guarantee correctness, because the model can still misquote a passage or put an inference in the facts column. What it does is make each claim’s claimed support visible, so checking takes minutes instead of an afternoon.

2. Examine the claims in the answer

Read the answer as a list of claims rather than as prose. Precise numbers, named sources, dates, quotations and statements about what a product or law allows are the most expensive kinds to get wrong, so give them the closest look. Automated hallucination detection tools can help sort a long answer, but a flag from one is a reason to check, not a verdict.

For citations, apply two tests: the source must exist, and it must support the specific claim attached to it. Plenty of hallucinated citations pass the first test and fail the second. Asking the same model “Are you sure?” is not a substitute for examining evidence. It tends to produce either a confident reassurance or an apologetic reversal, and neither one tells you which version is correct.

A second model is useful here for a specific reason. Models trained differently often fail differently, so a second reading can expose an omission, a contradiction or a claim that one model simply cannot reproduce. Models also share training data and blind spots, so they can agree on the same error. Treat agreement as useful context and disagreement as a lead to follow. Neither one counts as verification.

3. Verify consequential claims before use

Run this checklist before an AI answer goes into anything that matters:

  1. Isolate the consequential claim. Find the one or two statements the decision actually depends on, not every sentence in the answer.
  2. Inspect its evidence. Open the cited passage or the supplied document and confirm it says what the answer says it says.
  3. Check the primary source where needed. Go to the regulator, the filing, the vendor documentation or the original dataset rather than to someone else’s summary of it.
  4. Retain unresolved uncertainty. If a claim cannot be confirmed, carry it forward labeled as unverified rather than quietly dropping the label.

How Suprmind Fights AI Hallucinations

Chess rook behind a review table with an open report, a checklist and a pen

Suprmind is a multi-AI decision intelligence chat platform that brings GPT, Claude, Gemini, Grok and Perplexity into one conversation thread. We built it on one idea: disagreement is the feature. A single model answering alone has no one to contradict it. Five models in the same thread do.

In Sequential mode, the default, each AI reads every previous answer before writing its own, then adds to it, critiques it or corrects it. In practice, if the first model credits the supplier note with a lead-time improvement it never mentions, the next model has the note and that claim in front of it and can point out that the source says no such thing. Unsupported assumptions get challenged the same way, while they are still on the screen instead of after they have reached a client deck. Where the models diverge, our Disagreement/Correction Index flags it inline, so you can see exactly which claims were contested.

This is the first layer of AI Anti-Hallucinogen, our two-layer hallucination mitigation system. Layer one is the multi-AI conversation itself: five models reading and correcting each other. Layer two is True North, an independent AI fact checker.

The two layers do different jobs, and that is deliberate. Cross-model challenge tests the reasoning: it surfaces contradictions, gaps and claims that do not hold up under a second reading. But as the previous section explained, several models can still agree on the same error. Independent fact checking examines the claims themselves rather than whether the models agree about them. Challenged reasoning catches what consensus misses, and evidence checking catches what consensus gets wrong together.

We fight hallucinations. We don’t pretend they’re gone: mitigation, not prevention. Your judgment on the consequential claims stays in the loop, and our job is to make sure those claims reach you already contested and checked.

Frequently Asked Questions

If the response has citations, why might it still be wrong?

A citation can fail in two ways. It can point to a source that does not exist, or it can point to a real source that does not support the sentence it is attached to. Both look identical on the page. Only opening the source and reading the relevant passage tells you which you have.

If I upload the correct document, can the summary still invent details?

Yes. Grounding the model in your document reduces invention a great deal, but the model can still blend passages, drop qualifiers or add a plausible detail from its training, as the Northvale example shows. Ask for each claim with its supporting passage, and spot-check the ones that matter.

If several models agree, does that make the answer true?

No. Agreement is useful context, because independent models converging is more reassuring than one model’s answer. But models share training data and can repeat the same error. Treat agreement as raising your confidence, not as proof, and still verify consequential claims against evidence.

Can better prompts or newer models eliminate the problem?

They reduce it. Clear constraints, supplied sources and permission to say “I don’t know” all help, and models keep improving. No current approach removes hallucination risk entirely, so the realistic goal is mitigation to a level your use case can tolerate.

Three Questions to Ask of Every AI Answer

A fluent answer, a true claim and a faithful summary are three different things, and AI hallucinations live in the gaps between them. You do not need machine-learning expertise to catch most of them. You need three questions, asked about the claims that matter:

  • What is being claimed?
  • What supports it?
  • What happens if it is wrong?

If you want to see how we apply those questions across five models and an independent fact checker, the AI Anti-Hallucinogen page walks through how Suprmind mitigates hallucinations, layer by layer.

author avatar
Radomir Basta CEO & Founder
Radomir Basta builds tools that turn messy thinking into clear decisions. He is the co founder and CEO of Four Dots, and he created Suprmind.ai, a multi AI decision validation platform where disagreement is the feature. Suprmind runs multiple frontier models in the same thread, keeps a shared Context Fabric, and fuses competing answers into a usable synthesis. He also builds SEO and marketing SaaS products including Base.me, Reportz.io, Dibz.me, and TheTrustmaker.com. Radomir lectures SEO in Belgrade, speaks at industry events, and writes about building products that actually ship.

Suprmind AI Anti-Hallucinogen

Suprmind's proprietary
two-layer system for fighting AI hallucinations.

Suprmind's AIs challenge and correct each other in real time. True North adds a second layer, independently verifying claims and correcting detected errors in Context Fabric before they affect your work.

Learn More Now →