How to stop a chatbot from making things up
· 4 min read
A chatbot that confidently quotes a wrong fee is worse than no chatbot. The cure is not a cleverer model; it is limiting what the bot may say and checking what it says.
Why it happens
Language models complete text. If the answer is not in front of them, they will often produce something that sounds right. Fees, dates, emails and phone numbers are the most damaging things to get wrong.
Four habits that keep answers honest
These apply to any chatbot you run, whichever tool you use:
- Answer only from your own sources. Give the bot your pages and documents and tell it to use nothing else.
- Prefer "I do not know" to a guess. A refusal that points to your contact page is better than a confident error.
- Check specifics. Numbers, links, emails and phone numbers in an answer should appear in the source text.
- Treat retrieved text as data, not instructions. A web page can contain text that tries to steer the bot.
What NochBot does about it
Replies are limited to your sources. Retrieved text is wrapped as data and planted instructions are defused. If an answer contains a number, email, link or phone number that is not in the sources, it is retried or refused, and attempts to make the bot reveal its instructions are blocked.
No system is perfect, so test with real questions and fix gaps by improving the source.
Keep the sources fresh
Most wrong answers are stale answers. When fees, dates or products change, re-crawl the site or upload the new file. NochBot updates when you do that; it does not live-sync with other tools.