Guide
When AI makes things up
Please note: This site is only a guide to help people understand. Consult your IT, security and/or privacy department for clarification.
What it is
People call it a hallucination when an AI tool gives an answer that sounds sure but is false or made up.
These tools work by predicting likely words. A smooth answer is not always a true one, and a tool can write something false in the same calm voice it uses for something true.
The tool is not trying to trick you. It is finishing a sentence that sounds right. That is why you, not the tool, have to check.
Where it can hurt a nonprofit
These are the places where a made-up detail can damage the trust of donors, funders and clients.
- Invented statistics and research citations, with a title, author and year that look real but are not.
- Made-up quotes put in the mouths of real donors, officials or clients.
- Invented grant requirements, deadlines, eligibility rules or page limits.
- Wrong legal or tax rules.
- Web addresses that do not exist.
How to catch it
"Review for accuracy" is too vague to act on. Do these steps instead.
- Check every number against your own records, such as your database, your last report or your budget.
- Open every source before you cite it. Does the page exist? Does it say what the draft claims? If you cannot open it, do not use it.
- Check each quote with the person quoted, or against your own notes. If you cannot, cut the quote.
- Take grant rules, deadlines and page limits only from the funder's own guidelines, and check that they are the current version. Never from the AI's memory.
- Whether your organization qualifies for a grant is your call, made by reading the funder's own text, not the AI's.
- Check any legal or tax rule against the official government page or a qualified professional before you act on it.
- Search for each name, title and web address. If you cannot find it, treat it as made up.
- Ask the tool to quote only from text you pasted in, then compare its quotes to your original word for word.
How to reduce it
You cannot remove the problem, but you can make it less likely. Put the facts in your prompt so the tool does not have to guess.
Then tell it what to do when something is missing. Here is a line you can copy:
"Use only the information below. If something is missing, write MISSING INFO instead of guessing."
You can also ask, "List anything in your answer that you are unsure about." That does not catch everything, so keep checking.
Tools that search the web
Some tools search the web and show links. They still make mistakes. They can misread a page or link to one that does not say what the answer claims.
Open the pages they cite and read the part that matters.
If something wrong goes out
It happens. Correct it quickly and say so plainly, for example, "We made an error in our earlier post. Here is the right number."
Then find out which step let it through, and add a check there.