AI Hallucination Examples and How to Catch Them

By Austin Little

Hallucinations are not "AI having a creative moment." They are confident fabrications: fake papers, tidy statistics nobody measured, court cases that never were, and product buttons that do not exist. If you publish, teach, advise clients, or send a note that could move money, catching them is part of the job. This page shows the shapes those errors take and a practical catching loop — free tools first, skepticism always.

AI disclosure. This page was drafted with AI assistance and edited for Apiary. We don't invent quotes, stats, people, or events. If something looks off, tell Austin — that's the point of a living hive.

What we mean by "hallucination"

In everyday use, an AI hallucination is output that sounds plausible and is wrong — especially invented facts, sources, or details presented as known.

It is different from:

• A typo ("teh")
• A tone mismatch (too corporate)
• An opinion you disagree with
• A summary that is incomplete but not false

Hallucinations hurt because they pass a quick skim. The grammar is fine. The layout is clean. The only problem is reality.

Related deep checklist: How to Fact-Check AI Writing Before You Publish. Disclosure wording: AI Disclosure on Articles.

Why models invent (without the mysticism)

You do not need a research paper to operate safely. Useful mental model:

1. Models predict helpful-looking text from patterns.
2. "Helpful-looking" often includes citations, numbers, and proper nouns.
3. When the true detail is missing, the model may still complete the shape of an answer.
4. Your prompt can accidentally reward invention ("add impressive stats," "cite sources," "be definitive").

Paid seats do not abolish this. Local small models do not abolish this. Free tiers do not abolish this. Catching is an editing practice.

Hallucination example patterns (shapes, not fake news reports)

Below are patterns. We are not inventing a fake court case and then "catching" it as theater. Each pattern includes what it looks like, why it fools people, and how to catch it.

1) The phantom citation

Looks like: A neat article title, author name, year, and journal — or a URL path that feels right.

Why it fools: Academic shape signals authority. Many readers stop at the shape.

Catch: Click or search the exact title in quotes. Check the author + year together. If the only hits are AI-written mirrors repeating the same blob, treat it as guilty until proven.

2) The real venue, fake paper

Looks like: "Published in Nature / JAMA / a known conference" with a paper that does not exist there.

Why it fools: Mixing a real prestigious venue with a fake title is more convincing than inventing a fake journal.

Catch: Go to the venue's search, not only a general web search. Confirm DOI when one is offered. No DOI + grandiose claim = slow down hard.

3) The confident statistic

Looks like: "63% of homeowners…" / "the average colony loses 47%…" with no source, or a source that does not contain the number.

Why it fools: Specific numbers feel measured.

Catch: Ask "who measured this, when, on what population?" If you (or the model) cannot point to a primary table, delete the number. Do not replace it with a different invented number.

4) The almost-real person

Looks like: A plausible expert name at a real university or company, sometimes with a fake quote.

Why it fools: You recognize the institution and stop.

Catch: Search the person. Confirm they exist and work on that topic. Never publish a quote you cannot source to a primary interview, talk, or document you actually opened.

5) The product UI mirage

Looks like: "Click Settings → Privacy → Neural Vault → Enable" for a product that has no such path.

Why it fools: Software writing loves menus. Models have seen millions of menu sentences.

Catch: Click through the real product or official docs the same day. Screenshot-level claims need human eyes. Mark churny UI with a recheck date.

6) The legal / policy phantom

Looks like: A statute number, case name, or "OSHA requires…" claim that is tidy and wrong.

Why it fools: Legal citations have a rhythmic shape models imitate well.

Catch: Primary sources only — official code sites, regulator pages, or a licensed human. Do not let a chat be your counsel.

7) The map / place error

Looks like: Wrong capital, wrong river, invented landmark hours, a store "two blocks from" somewhere it is not.

Why it fools: Travel and local tips are common training patterns.

Catch: Map apps and official pages. Call the business if hours matter for a reader who will drive.

8) The timeline mashup

Looks like: Events ordered wrong; a person "speaking in 2019" about a product launched in 2024; "since COVID" used as a vague era stamp for unrelated claims.

Why it fools: Chronology is hard; prose still flows.

Catch: Build a tiny timeline for any piece that depends on sequence. Verify each date against a primary page.

9) The dependency / API that never shipped

Looks like: Code samples importing a library method that does not exist, or "the endpoint returns X" when docs say otherwise.

Why it fools: Code blocks look official inside markdown.

Catch: Run the code in a scratch environment or read current docs. Do not paste sample secrets. For security topics, prefer official advisories over chat reconstructions.

10) The flattering customer story

Looks like: A detailed anecdote about "a shop in Ohio" that bought a product and saved "18 hours a week."

Why it fools: Narrative is sticky. Marketers love it. Models can improvise it forever.

Catch: If it did not happen to you or a documented public case, label it as hypothetical — or delete it.

A catching workflow you can finish before publish

Use this every time a draft touched a model:

Step A — Claims harvest

Skim the draft and list:

• Numbers
• Proper nouns (people, papers, cases, products)
• Causal claims ("X causes Y")
• "Always / never / required / banned"
• Quotes
• Links

If the list is long, that is good. You found the risk surface.

Step B — Sort by harm

Step C — Verify or delete

For each high/medium claim: confirm, rewrite with a sourced phrase, or remove. "I couldn't verify this" is an acceptable publish state. "Sounded right in chat" is not.

Step D — Link check

Open every link. Confirm the page supports the sentence that cites it. Watch for 404s and for pages that exist but say something else.

Step E — Read aloud once

Your ear catches timeline nonsense and too-perfect statistics better than tired eyes.

Prompts that reduce invention (they do not eliminate it)

Helpful constraints:

• "Use only details present in the notes below. If a detail is missing, write UNKNOWN."
• "Do not add statistics, citations, or quotes."
• "List assumptions separately from facts."
• "Propose questions I should verify, not answers to invent."

Harmful prompts:

• "Add citations to make this credible."
• "Make it sound researched."
• "Fill any gaps so it reads finished."
• "Invent a case study."

Even with good prompts, run the catching workflow. Prompts are seatbelts, not teleportation.

Free-tier first toolkit for catching

You do not need a paid "AI detector" subscription to catch fabrications. Detectors that claim to spot AI writing are a different (and often shaky) product category anyway — this page is about factual invention, not authorship policing.

Free-first tools:

1. Your browser + official documentation sites
2. Library search / Google Scholar-style search for papers (still verify full text)
3. Government and standards sites for legal/regulatory claims
4. The product itself for UI claims
5. A second human when harm is high
6. Local model as a devil's advocate ("list every factual claim in this draft") — then you verify; do not let model B rubber-stamp model A

Local privacy-friendly drafting still needs this catch loop: Free AI for Writing Blog Posts Locally, How to Use AI Without Paying.

Field guide: what catching looks like in real editing

Example A — Newsletter draft Model adds: "Studies show 80% of readers prefer morning emails." Catch: No study linked. Delete. Replace with what you actually know: "We send Tuesday mornings because our members said that slot works in last year's survey" — only if true.

Example B — How-to software post Model adds a menu path. Catch: Open the app. Path wrong. Rewrite from what you clicked. Add a "UI changes; recheck on publish week" note.

Example C — Beekeeping sidebar Model suggests a treatment dose. Catch: Hard stop. Verify with extension guidance / mentor / label. If not verified, remove. AI is not your applicator license. See Can AI Help Me Learn Beekeeping?.

Example D — Business FAQ Model invents a refund statute citation. Catch: Replace with your actual written policy and a link to the regulator page you truly mean — or say "ask our office" instead of fake law.

Team habits that prevent repeat embarrassment

• Keep a shared "banned inventions" note: no fake stats, no fake quotes, no fake customers.
• Require a claims list for any AI-assisted public post.
• Separate drafting roles from verification roles when you can.
• Disclose AI assistance honestly near the top when you used it.

What "AI detectors" do not solve

If your goal is catching hallucinations, an "is this AI-written?" score is the wrong instrument. Human-written text can be false. AI-assisted text can be true. Focus on claims, not vibes about authorship. If a school or workplace requires detection tools, follow their policy — and still fact-check.

Local vs hosted: same disease, different privacy

Do not buy a subscription believing you purchased immunity from invention. Buy one only if other checklist items say so — Should You Pay for ChatGPT Plus? A No-Hype Checklist.

Teaching skeptics without creating panic

People new to AI sometimes swing from "it knows everything" to "it always lies." Neither is operationally useful.

Better teaching line:

It is a fast drafting assistant that will sometimes invent details. We assume invention risk on every factual claim we did not supply. We verify or cut.

That sentence works for students, volunteers, and grandparents. Pair with dignified setup help: AI for Grandparents, Best Free AI for Students.

A 15-minute pre-publish catch (solo)

1. 2 min: Export or open the final draft. Turn on a claims mindset.
2. 5 min: Highlight every number, name, quote, and "required/banned."
3. 5 min: Verify the top three highest-harm items against primary sources.
4. 3 min: Delete or hedge anything still soft. Click links once.

If you cannot finish verification, narrow the article. A shorter true piece beats a long invented one.

Joe-Google test

People type:

• "AI hallucination examples"
• "ChatGPT made up a source"
• "how to catch AI hallucinations"
• "does ChatGPT invent facts"
• "AI fabricated citation"

They want recognition ("oh, that happened to me") and a method. Give both. Skip cosmic essays about consciousness.

Common excuses that publish falsehoods

• "The model is usually right."
• "I only used it for a first draft" (then forgot to check the shiny parts).
• "Readers won't click the citation anyway."
• "It's close enough for a blog."
• "We'll fix it if someone complains."

If someone could act on your words — spend money, take a medicine-adjacent action, climb a ladder wrong, treat a hive wrong — "close enough" is not a standard.

Sources / further reading

• Product docs for any UI you describe — rechecked on publish day

More patterns worth recognizing

11) The "everyone knows" weasel that hardens into a fact

Looks like: Starts as "many experts say," then later paragraphs treat it as settled measurement.

Catch: Search for the softened claim and the hardened claim separately. Force the draft to keep hedge language unless you have a source.

12) The version number fanfic

Looks like: "In Python 4.2…" / "since iOS 19…" / "GPT-4.7o-mini-ultra says…" with confident release notes.

Catch: Check official release channels the same day. Software versions are hallucination candy. Prefer "current stable as of [date]" with a link over theatrical version theater.

13) The misplaced precision price

Looks like: "$19.99/month includes unlimited X, Y, and Z" for a plan that changed, differs by region, or never included Z.

Catch: Open the vendor pricing page in a private window. Do not let last quarter's blog become this quarter's lie.

14) The bibliography padding

Looks like: Five sources at the bottom; two are real, three are remixes of titles the model has seen.

Catch: Verify each entry independently. One fake source poisons trust in the real ones.

15) The "translated" quote that was never said

Looks like: A famous person "explaining" your topic in suspiciously on-message language.

Catch: Find the original interview or speech. If you cannot, cut. Paraphrase without quote marks only when you are summarizing a source you actually read — and even then, be careful.

Classroom, church, and shop floor: catching in public institutions

Volunteer editors and teachers need a lightweight ritual, not a 40-page policy.

Classroom: Students may use AI for outline help where allowed. They still must show sources for factual claims. Invented citations are an academic integrity issue whether a human or a model typed them.

Church / community newsletter: Appoint one "claims buddy" who only checks names, dates, times, and addresses. Tone can be warm; directions to the potluck cannot be imaginary. See Free AI for a Church or Community Newsletter.

Shop / small business: If a model drafts a how-to for customers, a licensed or experienced tech confirms any safety step. A wrong torque spec or chemical tip is not "content." It is liability-shaped.

Building a personal "red flag lexicon"

Keep a note file of phrases that trigger verification:

• "Studies show"
• "Experts agree"
• "It is required by law that"
• "Clinical trials prove"
• "As documented in"
• "According to internal data" (especially when you have no internal data)
• "Always" / "never" on empirical topics
• Exact percentages
• Court case style names
• Menu paths with nested arrows

When you see one, slow down. The lexicon trains your eye the same way bees teach you to spot mites — pattern first, panic never.

Using two models without creating a hallucination circle

A common mistake: Model A invents a citation; Model B says "looks good"; you publish.

Safer two-model use:

1. Model A drafts.
2. Model B extracts claims as a bullet list (no judgments).
3. You verify bullets against the world.
4. Optionally Model B helps rewrite after you supply corrected facts.

Never ask Model B "is Model A correct?" as your only check. That is vibes laundering.

When cutting is braver than hedging

Hedges help ("approximately," "in many U.S. states," "check your label"). Sometimes they become a fig leaf for a claim you should not make at all.

Cut when:

• You cannot name a source even vaguely
• Harm is high
• The sentence exists only to sound smart
• You are tempted to "leave it for now"

Readers forgive short. They do not forgive fabricated certainty.

Incident response if you already published a hallucination

1. Confirm the error with a primary source.
2. Fix the live page.
3. Add a short correction note if people may have acted on it.
4. If you emailed the falsehood, send a plain correction — no theatrical blame on "the AI."
5. Update your checklist so that class of claim gets caught earlier next time.

Blame-shifting to the model reads as immature. You shipped it.

A longer practice lab (30–40 minutes)

Take any AI-assisted draft you have (or generate a disposable one from public notes).

1. Print or copy into a file marked claims-lab.md.
2. Highlight every proper noun and number in a loud color.
3. Build a table: Claim | Source needed | Status (verified / cut / hedged).
4. Verify five claims fully.
5. Time yourself. Notice which claim types ate the clock — those are your personal risk magnets.
6. Save the table template for next week.

Do this twice and your default skim changes forever.

Apiary house stance (plain)

We use AI to draft and we disclose it. We do not invent quotes, stats, people, or events. That is not bureaucracy. That is how a living hive stays trustworthy.

FAQ

Do hallucinations mean the tool is useless? No. They mean unverified factual prose is dangerous. Drafting and rewriting still help when you catch claims.

Are hallucinations rarer on bigger paid models? Sometimes error rates improve on some tasks; that is not a promise of truth. Always verify high-harm claims.

Can I ask the model to verify itself? You can ask it to list uncertain points. You still need external primary sources. Self-audit is not a court of law.

What if I only use AI for outlines? Lower risk, not zero — watch for invented section claims that sneak into later drafts.

Is a wrong opinion a hallucination? Usually no. Call those disagreements. Save "hallucination" for fabricated facts presented as known.

Should I tell readers I used AI? Apiary's practice is clear disclosure when AI assisted. See the disclosure guide linked above.

Does local AI hallucinate less because it is on my machine? Privacy changes; truthfulness does not automatically.

What is the fastest red flag? A precise statistic or citation you did not provide in the prompt, appearing in a polished paragraph.