L

Hallucination vs Ordinary error

Both leave you holding a wrong answer, delivered with the same fluency. They are worth telling apart because the repair is completely different.

Two wrong answers. Different repairs.

Both are confidently wrong about the same contract. What you do next is not the same.

Wrong answer A

“Clause 12.4 sets the termination penalty at £5,000.” Clause 12.4 exists. It says £10,000.

An ordinary error. A real thing was read and read wrongly. Open clause 12.4 and it is settled in ten seconds.
Wrong answer B

“Clause 19.2 sets a 60-day cure period.” The contract has fourteen clauses. There is no clause 19.

A hallucination. There is nothing to open. You cannot correct it by looking — you have to stop trusting the whole answer.
Both are wrong, both are fluent, both name a clause. One misread something real. The other invented the thing it was reading.

What each one actually is

One misuses a real source. The other has no source at all. That single fact decides what fixes it.

The difference is not severity — it is repairability. An ordinary error has a real source sitting behind it, so opening that source both finds the mistake and gives you the right answer.

A hallucination has nothing behind it. There is no page to turn to, so it cannot be corrected by looking closer — and its presence tells you something about the whole answer, not just that line.

So the diagnostic question is not “is this right?”. It is does the thing it names exist? If yes, you are checking a figure. If no, you are re-doing the task.

Diagnose six answers, then see the fix

Tap an answer, then tap which kind of wrong it is. Then see the diagnosis and what actually repairs each one.

The same wrong, different desk

Pick a situation. Notice how the repair changes the moment you know which kind you have.

Every one of these is somebody trying to correct a hallucination by reading more carefully.

The citation that pointed nowhere

Author, year, journal, page. Correct in every respect except existing.

Test the distinction

Five questions. Nothing is scored.

Question 1 of 5
Multiple choice

The words this pair actually contains

Five terms, not two. Tap one.

Questions people ask

Open all 21 questions
What is a hallucination in AI-generated answers?

A hallucination is when a language model creates content—such as a fact, quote, or source—that has no support in any real document, database, or conversation. The model makes it up entirely, rather than pulling from something that exists.

How is a hallucination different from an ordinary error?

A hallucination is made-up content with no real source. An ordinary error is a mistake with something real, like misquoting a date or misreading a number. Both are wrong, but only hallucination invents from nothing.

Can an AI hallucinate a source or document?

Yes. A language model can invent a source, document, or citation that does not exist. If you search for it, you will find nothing. This is a classic hallucination.

What are examples of ordinary errors from language models?
  • Getting a date wrong from a real document.
  • Misquoting a figure from a spreadsheet.
  • Summarising a report but missing a key point.
  • Mixing up two real people’s names.
  • Using the wrong section number from a policy.
How do I check if an answer is a hallucination or an ordinary error?
  1. Find the claim or fact you want to check.
  2. Look for the source: a document, database, or person.
  3. If you can't find any support for it, it's a hallucination.
  4. If the source exists but the detail is wrong, it's an ordinary error.
Why do language models hallucinate?

Language models are trained to generate fluent, plausible text. Sometimes, they fill gaps by inventing details that sound right but have no real backing. The model does not know whether these details exist.

Can better prompts stop hallucinations?

Better prompts can reduce the chance of hallucination, but cannot remove it entirely. The model may still invent details if it lacks information or context. Checking claims outside the model is always needed.

What should I do if I suspect a hallucination?
  1. Identify the claim that seems unsupported.
  2. Search for it in your documents, databases, or ask a colleague.
  3. If no evidence exists, treat it as a hallucination.
  4. Flag or correct the answer before using it.
Do hallucinations only happen with facts and figures?

No. Hallucinations can happen with names, quotes, policies, or even entire documents. Any content invented by the model with no real backing counts as a hallucination.

How can I reduce ordinary errors from a language model?
  • Give clear, specific instructions.
  • Provide the relevant documents or data.
  • Double-check numbers and dates.
  • Review summaries for missing points.
  • Ask for sources where possible.
Is it possible for an answer to contain both a hallucination and an ordinary error?

Yes. An answer can invent a source (hallucination) and then misquote a figure from it (ordinary error). Both types of mistakes can appear together in a single response.

What is the risk of using hallucinated content in work documents?

Using hallucinated content can lead to decisions based on things that do not exist. This can damage credibility, waste time, and even create legal or financial problems if acted upon.

How do I explain hallucination to a colleague who is not technical?

Say the model sometimes makes things up—like a name, rule, or report that nobody can find. If you cannot point to where the claim came from, it might be a hallucination.

What are the signs that an answer might be a hallucination?
  • A rule or figure you have never seen before.
  • A citation you cannot find anywhere.
  • A document that does not exist in your records.
  • A quote that does not match any real source.
Can hallucinations be caught by reading the answer more carefully?

No. Careful reading can spot contradictions or clumsy writing, but hallucinations look as fluent as true statements. Only checking outside the answer—against real sources—can reveal them.

Does asking the model to double-check its answer prevent hallucinations?

No. The model may agree, change its answer, or repeat the same claim. Since both the original and the check come from the same place, this does not count as real verification.

What is the best way to verify a claim from a language model?
  1. Identify the claim you would act on.
  2. Find where that claim would live if it were true—a document, register, or person.
  3. Check that source yourself.
  4. If you cannot find it, treat it as a hallucination.
How do hallucinations and ordinary errors affect decision-making?

Both can lead to mistakes, but hallucinations are riskier because they introduce things that do not exist. Acting on hallucinated content can cause bigger problems than a simple misquote or miscalculation.

Can hallucinations happen even with up-to-date training data?

Yes. Even with current data, models can still invent details when they lack information or context. Hallucination is a property of how models generate text, not only of outdated data.

What kinds of claims need checking for hallucinations?
Type of claimExampleCheck for hallucination?
Policy rule‘Section 12A requires approval’Yes—verify the section exists.
Summary of a real report‘The 2023 report says X’Yes—find the report and check.
General advice‘It's best to review quarterly’No—advice is not checkable for hallucination.
Known contact name‘Contact Priya Sharma in HR’Yes—make sure the person exists.
Can hallucinations be completely avoided in AI answers?

No. While better prompts and more context help, hallucinations cannot be fully avoided. Always check important claims against real sources before acting on them.

One kind of wrong points at a real thing and misreads it. The other points at nothing at all. Only one of them can be corrected by looking again.

Copyright © Pawan Nayar · LLOS.ai · 2026 — Hallucination vs Ordinary error: no source at all, versus a real source used wrongly.Original pedagogy, voice, and design — all rights reserved.