Keywordshow to catch ai hallucinationsBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searcheshow to catch ai hallucinationschatgpt hallucination examples at workstop ai making up factsai fluent falsehoods at workdetect chatgpt made up sourceshow to use how to catch ai hallucinations
A hallucination is a fluent falsehood, not a clumsy sentence
An AI hallucination at work is not a glitch in the interface. It is a fluent falsehood: a policy clause that does not exist, a SKU the catalogue never carried, a colleague who supposedly confirmed a date. The sentence is grammatical. The tone is calm. That is why it passes a skim. Catching hallucinations is not a vibe check. It is a habit of opening the underlying file, mapping claims to sources, and noticing the gaps where the model wrote something because a gap looked ugly. Fluency is the camouflage. Your job is to look for the missing fact, not for a clumsy phrase.
The usual workplace example is a made-up pointer. ChatGPT cites policy section 4.2. You are busy, so you assume 4.2 exists. Section 4.2 is about something else, or the document jumps from 4.1 to 5. The model completed a pattern that policies have numbered clauses. Completing a pattern is not reading. Another example is a person. The draft says Priya in operations confirmed Friday. Priya was on leave. Nobody confirmed Friday. The name made the logistics feel owned. Owned is not the same as true. You catch both by opening the policy and the calendar instead of admiring the sentence.
This essay is a catcher's method for ordinary artefacts: notes, emails, slide titles, and answers you are about to paste into a ticket. It is sceptical of fluency on purpose. The model is not trying to trick you. It is trying to finish the page. Finishing the page is how fiction appears in a professional register. A named human still has to open the file. If you cannot open the file, you cannot catch the hallucination, and you should not send the line. Open the file before you send it.
If a sentence would collapse without a fact you have not opened, treat it as a hallucination until the file says otherwise. Do not send it in the meantime.
Source maps, missing-fact gaps, and the file on the desk
Three checks catch most workplace hallucinations. First, a source map: every name, number, clause, and date gets a pointer or a gap. A gap is a finding, not a prompt to let the model try again with more confidence. Second, a missing-fact scan: where would a complete answer need a file you did not provide? Those are the places invention hides. Third, open the underlying file. Not a summary of the file. The file. If the draft mentions a SKU, search the catalogue. If it mentions a clause, scroll to it. If it mentions a confirmation, find the mail. The check is physical. Skimming the chat for a nervous word will miss the calm lie.
Ask the model to mark [missing] instead of guessing. That instruction reduces some invention and does not remove the need to open the file. Models still invent citations that look like marks. They still promote a nearby fact into the one you needed. Your catch is correspondence: does this sentence survive contact with the object? If you only have consumer ChatGPT and a restricted file, do not paste the file to catch the error. Catch it with your eyes and the system of record. The method is the opening, not the brand of the assistant that drafted the page.
| Hallucination shape | What it looks like | How you catch it |
|---|---|---|
| Invented clause | Policy 4.2 says we must | Open the policy. Read the section. |
| Invented SKU or ID | A code that sounds in-family | Search the catalogue or ticket system |
| Invented confirmer | Priya confirmed Friday | Find the mail, the minutes, or cut it |
| Invented source | A paper or page that fits | Open it. If it is missing, delete the claim |
A catch pass before the draft leaves your machine
On a copy of the draft, highlight every pointer: sections, names, codes, dates, and according-to lines. For each, open the object or mark [missing]. Then look at the unhighlighted sentences that still assert a fact. Those are the quiet hallucinations. Delete them or demote them to a question you will ask a human. Do not ask the same model whether the draft contains hallucinations and then treat a no as clearance. It will often say the text is consistent. Consistency is what fluent fiction already has. You need the file, the catalogue, and the calendar. Stop when the remaining page only contains what you opened.
Keep examples. One invented clause, one invented person, one invented figure, written down after you catch them. Share the list with the people who use the same briefs. Catching hallucinations becomes faster when the local shapes are named: our policies are not numbered that way, our SKUs have a prefix, our confirmations live in mail not in chat. A generic warning to be careful does not change behaviour. A named miss from last Tuesday does. The log is the training. The tool will not remember your last embarrassment unless you write it into the brief as a ban.
Role: You are a reviewer hunting fluent falsehoods in a draft I will sign.
Task: Extract every claim that depends on a file, a person, a code, or a date.
Context: I will paste the draft and name the underlying systems I actually have.
Constraints:
- Do not invent a source to make a claim look grounded.
- If a claim has no support in the paste, mark [missing].
- Do not say the draft looks consistent. Consistency is not the test.
Output: A list of claims with [missing] or a pointer I supplied, plus the three claims most likely to be invented.
Quality checks: What would collapse if the file, catalogue, or calendar disagreed.Take the three claims marked as most likely invented and open the objects yourself. Search the catalogue for the code. Scroll to the clause. Find the mail. If any object is absent, delete the sentence. Do not ask the model for a replacement fact. Write a shorter true page, or send a question to the person who owns the file. The catch is complete only when the remaining lines survive that opening.
Self-review, pretty citations, and filling the ugly gap
The model is a poor sole reviewer of its own charm. A second pass that asks is anything made up will often produce a polite reassurance and a minor caveat. That caveat is not a catch. Pretty citations are a related trap: a real-looking title, a plausible year, a section number. You open nothing because the shape of a source is there. Open it. Filling the ugly gap is how the rest appear. You asked for a complete answer. The policy was silent. Silence became a clause. Completeness was the instruction that created the fiction. Prefer an honest hole to a finished paragraph you cannot defend in a complaint file.
Do not catch hallucinations by pasting more secrets. If the draft might be wrong about a customer, open the ticket in the helpdesk. Do not dump the ticket into consumer ChatGPT to see whether the model now agrees with itself. Agreement between two chats is still not the file. Also watch role-play. Asking the model to be the policy will produce policy-shaped text with no instrument behind it. That is hallucination on request. Use the approved copy of the policy. If you need an explainer, ground it in the paragraphs you paste and keep the gaps visible.
| Mistake | What it looks like | What to do instead |
|---|---|---|
| Skim for clumsy tone | It reads smoothly, so it is fine | Open the file, catalogue, or mail |
| Trust the citation shape | A section number in the sentence | Scroll to that section or delete it |
| Self-clearance | Ask the model if it invented anything | You open the object. The model does not. |
| Fill the silence | Complete answer from a silent policy | Write [missing] or ask the owner |
A calm sentence with a section number is not a sourced sentence. If you have not opened the clause, you have not caught the hallucination. Do not send it.
Related reading on StudyGrid
Read next: Where AI Falls Short How to Fact-Check ChatGPT What Not to Paste into ChatGPT. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.
What to do this week
Take one draft you almost sent. Use a draft from this week, not a training example. Highlight every pointer. Open each object. Search one catalogue code, one clause, and one named confirmer. Delete what you cannot open. Send the thinner page. Write the miss in a shared note so the next brief bans that shape. That is how you catch AI hallucinations as a habit rather than as a scare story.