- The five kinds of content people should check, and why
- How to check without just rubber-stamping
- How to decide who checks what, and against which source
The short answer
When you use answers from generative AI at work, not everything needs the same level of checking. People should check the things that are hard to undo if they are wrong: facts and figures, anything sent outside, decisions about money, contracts or HR, personal information, and actions that cannot be reversed. And rather than assuming "someone will look at it," decide in advance who checks what, and against which source.
AI answers can look plausible and still contain errors. Decide who checks what, against what.
Five kinds of work people should check
AI can write things that sound plausible but are not true, which is called a hallucination. IPA, Japan's Information-technology Promotion Agency, also lists false and inaccurate information, and leaks of personal information and trade secrets, among the challenges around AI. Keep a human check on these five.
| Kind | Examples | Why a person checks |
|---|---|---|
| Facts, figures, dates | Amounts, deadlines, names of laws, proper names | Errors are hard to spot just by reading the text |
| Anything sent outside | Emails to clients, text you publish | Once sent, it cannot be taken back |
| Money, contract and HR decisions | Payments, contract terms, evaluations | The impact on people and the company is large |
| Personal information | Contact lists, addresses, health matters | Mishandling input or output can become a leak |
| Irreversible actions | Deleting, sending, overwriting | Some cannot be undone |
For example, suppose you ask AI to draft an apology email to a client. Even if the wording is polite, a wrong delivery date or recipient causes trouble for the other side, much like an email sent to the wrong person.
Do not just rubber-stamp it
Even when someone is assigned to check, they may simply trust the AI's answer and let it through. Japan's AI Guidelines for Business, published by the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry, call excessive trust in or reliance on automated systems "automation bias" and urge caution about it. One countermeasure they mention is that when a person approves an AI judgment, they should first think through the reasons and evidence for approving it on their own.
In practice, make the checking concrete.
- Open the original material and compare numbers and dates one by one
- Ask whether you could point to the supporting source yourself, rather than relying on "the AI said so"
- If in doubt, do not use it; send it back to the person in charge
How to assign checking roles
Assign checking roles task by task, using four items. The template below can be copied as is.
Task:
Who checks:
What they check:
Checked against:
Where the check is recorded:As a fictional example, take a monthly internal report drafted with AI. You might decide: "Task: drafting the monthly report", "Who: the accounting staff member", "What: sales figures and the direction of comparisons", "Against: totals in the accounting software", "Record: the checker's name and date in the report file."
Once the roles are set, add them to your organization's AI usage rules so the practice continues even when the person in charge changes.
In summary
Checklist
- Did you identify the parts of the AI answer that fall into the five kinds?
- Did you compare numbers, dates and proper names against the original material?
- Does a person check before anything goes outside and before irreversible actions?
- When approving, did you confirm you could point to the evidence yourself?
- Did you decide who checks what, and against which source?
One thing to check today
Pick one task where you use AI answers, and decide one thing: who checks what before the answer is used.