Cost
Fixing a published mistake is its own cost. This page does not estimate it.
Home · Start here · 07 · When not to use AI
Lesson 7 of 7.
Lesson · No account
When not to use AI
Do not use a chatbot when you cannot check the result and a person is accountable for it.
1 What it is
Some tasks need a judgement you are willing to sign. A filing, a medical decision, a hiring decision, a grade, a number that will be published, a message that harms someone if it is wrong. A chatbot can draft words. It does not take the responsibility. If you cannot check the draft against a source or your own expertise, do not hand the task to a chat.
NIST's AI Risk Management Framework treats clear responsibility and human oversight as part of managing AI risk. It is voluntary guidance. It does not list these tasks, and this card is not a compliance checklist or legal advice. The rule here is simpler: if you cannot check it, you do not send it.
Example · the figure is made up
A drafted sentence says the return rate last quarter was 4%. You cannot find 4% in the spreadsheet, and the number will be published.
Do not publish it. Check the spreadsheet, or leave the chat out of this task. The 4% was written for this card. It is not a measured rate.
2 Why it matters
Quality
A draft you cannot check is not a finished answer.
Risk
The person who sends it is the person who answers for it.
3 How to apply · try today
Look at the last thing you sent that a chatbot drafted. If you did not check it, and someone could rely on it, decide whether you can still check it. If you cannot, do not send the next one the same way.
Common mistake
Using the chat because you are short on time, on the task where being wrong is the problem.
Self-check
You cannot tell whether a drafted number is right, and the number will be published. What do you do?
List one task you were about to give a chatbot this week. Write who would rely on the result, and what you would open to check it. If you cannot name that source or that expertise, do the task without the chat.
OpenAI's safety best practices recommend having a human review outputs before they are used, especially in high-stakes domains, with access to the information needed to verify them. The Safety card is the builder version of the limit: what the product blocks, and what the provider's policies block. This lesson stays with the person who is about to send the reply.
Go further
The Safety card is the limit on harmful or disallowed use, in a product you are building and in the provider's policies.
Questions
- Can a chatbot draft it if I check every line?
- Yes. The draft is not the problem. Sending it without a check is.
- Is this legal advice?
- No. It is a working rule: if you cannot check the result, and someone is accountable, do not hand the task to a chat.