AI mistakes

Before you trust a source an AI gives you

← Journal

Made by Lizvo · AI mistakes · Short read · Published

A convincing citation and a real one look the same on the screen. An AI tool can name a document, a section number, a past decision or a person’s comment, with confident detail, even when none of it exists. This is one of the more common kinds of AI mistake, and it is easy to miss because it reads like help.

Five checks, in the order that usually saves the most time:

1. Ask where it is

“Where exactly can I find that?” A real source has a location: a file name and folder, a link, a page, a message you can scroll back to. If the answer is vague, or changes when you ask again, mark the point as unverified.

2. Open it yourself

Don’t take a description of the source as the source. Open the file, follow the link, scroll to the message. Many invented sources fail at this step.

3. Find the sentence

If it says something specific, find the sentence. A real document can still be quoted wrongly, or stretched to support a conclusion it doesn’t support. If you can’t find the words, you have a paraphrase, not a source.

4. Be most careful with precise-looking details

Exact section numbers, dates, version names and quotation marks give a feeling of accuracy. They are also easy to produce without a source behind them. Precision is not evidence.

5. Keep a short list of what you actually checked

Two columns are enough: the claim, and where you saw the evidence. Anything without a second column entry is unverified, and you can say so when you pass the work on.

Asking the same AI “are you sure?” is not a check. It will usually say yes. The check is whether you can find the source.

Related

More from the journal

— Lizvo

General information only, not professional advice. See the disclaimers.