🛑Phrases to reduce "Hallucinations" in LLMs

A "Hallucination" is a made-up or incorrect information in AI responses. Remember, Large Language models (LLMs) don't "know" facts - they know patterns of words based on huge amounts of text.

Also, they typically don't like to say they don't know because they are trained to help. And not giving an answer or saying "I don't know" it thinks is not helpful.

Below are some suggestions to add to prompts to reduce hallucinations. But they could still happen, so you should always fact check anything you don't know yourself.

1. "If you don't know, say so, I won't be cross. Do not make anything up"

   An explicit instruction on what to do when they don't know is a good start, coupled with a reminder to not make anything up.

2. "Use this source, and only this source. Do not include any other knowledge"

   If you are referring to a document or required source site, make it clear this is it's only source.

3. "Act as a XXXX. Only refer to ......"

   Define its role and its broad reference points. For example "Act as a compliance manager. Only refer to UK legislation"

4. "List your sources or explain how you arrived at the answer"

   Asking it to justify or verify itself will help keep it on track. You can also then check the sources.

5. "First do x, then do y. Do not assume anything"

   Hallucinations are more likely with complex, multiple instructions. Therefore be clear about the order of them.

6. "Stick to bullet points with facts only - do not embellish your answer or add extra narrative"

   This is particularly useful if its just being waffly rather than wrong. Sometimes, you just want an answer!

7. "If unsure, say so"

   Particularly useful if you are just getting used to prompting. Maybe you've not been explicit enough in your instructions. Vagueness can lead to hallucinations.

💡Pro tip: Add your phrase at the beginning and the end of your prompt. For example, ending with "Remember, don't make anything up" just helps keep that focus, particularly in longer prompts.

⚠️ Remember: Using these phrases should reduce the hallucinations, but cannot guarantee that you won't get inaccurate information back. You should always check your facts before acting on them.

Found this useful? Check out our AI Glossary for beginners

About the author

Connect with me on LinkedIn here

Create your own Scribe here

Check out my other Scribes here