Checking AI work before it goes out: a rule for staff

The rule can be one sentence. A person must read and take responsibility for anything AI helped make, before it goes to a customer, is published, or shapes a decision. The detail is in who that person is and what they look for.

Why the rule is needed

AI tools can sound sure of themselves while being wrong. Wikipedia defines an AI hallucination as a reply that states false or misleading information as if it were fact.¹ It adds that chatbots built on large language models can slip believable falsehoods into their writing, including invented citations.¹

The UK's cyber security centre lists the same flaw. It also says the tools can be biased and are often gullible when asked leading questions.² Australia's privacy regulator notes that output can look credible even when it's wrong.³

So smooth writing tells you nothing about whether it's right.

One reviewer or two?

A mix often works. Self-check for private drafts, second-person check for anything that leaves the building. Write down which jobs sit in which row.

What the reviewer checks

Give reviewers a short list. A long one gets skipped.

• Facts. Are names, dates, prices and claims right? Check against the original source, not by asking the same tool again.
• Sources. If the text cites a report, law or study, does it exist, and does it say that? Invented citations are a known problem.¹
• People. Is anything said about a real person? Is it true and fair?
• Tone. Does it sound like your business, and would you say it to this customer's face?
• Leftovers. Look for placeholder text, wrong names from a template, or instructions left in the draft.
• Data. Has the draft pulled in details it shouldn't contain? The paste rules are in this companion page.

For legal, tax, medical or safety content, the reviewer should be someone qualified in that field. A policy shouldn't ask a general reviewer to judge it. Point staff to the official source, such as a regulator's own page, instead of an AI summary.

Who counts as a reviewer

The National AI Centre's guidance says whoever oversees an AI system should know how to do that job properly.⁴ For this page, take that to mean: pick people who know the subject, and show them what a typical AI mistake looks like. Reading a few examples together is a good start.

Also decide what happens when the reviewer is away. Otherwise work either stalls or skips the step.

Keeping a light record

You don't need a heavy process. A line in a shared note or the job ticket is enough: who checked it, and when. If a mistake gets through, the record shows where the check failed, and you can fix the rule rather than blame a person.

Sample wording

"Anyone who uses an AI tool to help make work is responsible for it. Before it goes to a customer, is published, or is used to make a decision about a person, check every fact against a trusted source. Work that goes outside the business also needs a second person to read it. Don't send AI text you haven't read in full."

Add one example of a past slip in your own business, if you have one, and the date the rule starts.

Some work involves extra care when it reaches customers. When to say AI was used is covered under customer work.