AI at WorkCheck AI answers: hallucinations and a six-step test
AI tools can state something false in the same confident voice as something true. The invented part is often the detail: a number, a name, a source.
Why it happens
A language model writes what fits, not what it has verified. When it does not know something, it tends to produce something plausible. People call this a "hallucination". It appears most in precise details: legal section numbers, case names, quotations, web addresses, book and paper titles, statistics, prices and dates. It is also more likely when you ask about something rare, very recent, or very specific to your own situation.
What always needs checking
| Kind of answer | Why check |
|---|---|
| Laws, sections, rules, deadlines, rates | They change, and a wrong one can cost money. Use the official site |
| Numbers and calculations | Models can slip in arithmetic. Recalculate |
| Quotations and sources | They may not exist. Open the link, find the page |
| Names, titles, addresses | Plausible but invented ones are common |
| Medical, legal and financial advice | Ask a qualified person. AI is not one |
| Anything about a recent event | The model may not know it |
The six-step test
- Source. Ask: "Which source supports this?" Then open the source yourself. If the tool gives a link, check that the page says what the tool says it says. If it gives a name but no link, search for it.
- Numbers. Redo the arithmetic in a spreadsheet or calculator. Check the units and the period.
- Names and places. Check the spelling and that the person, company or law exists, on an official or reliable page.
- Dates. Check that a rule applies now, not only in the past. State "as on" dates in your own document.
- Second opinion. Ask the same question in a different way, or in a different tool. If the answers disagree, one is wrong, and you must find out which.
- Reader check. Give the draft to a colleague who knows the subject. If they hesitate at a sentence, check that sentence.
Ways to reduce errors
- Provide the facts yourself. Paste the rule or the table and ask it to work only from that: "Use only the text below. If the answer is not in it, say so."
- Ask for uncertainty: "Mark anything you are not sure about."
- Keep the task small. Long outputs hide small errors.
- Use tools that cite sources, and still open the sources.
A rule for your work
Never send, file, print or publish something that you could not explain line by line. The tool writes the draft; you are responsible for the result. If you work with clients or students, say how you use AI where it matters to them, and never put a made-up source in front of them.
An example of a check
Suppose a tool writes: "Under section X of the Act, the deadline is the 20th of each month." You open the official website of the department, search for that section, and read the current text. If the section is different, or the date has changed, you correct your draft and keep the official page address in your notes.
Can I just ask the AI whether its answer is true?
It may say yes to a false answer. Use real sources instead.
Do paid tools make fewer mistakes?
Often fewer, never none. The check is still yours.
This is an independent guide. It is not connected with OpenAI, Anthropic, Google, Microsoft or any other AI company; product names belong to their owners. Statements about companies are taken from their own help pages as read on 13 and 14 November 2026. The prompts are our own wording and were not run against every tool. Not legal, financial or professional advice.