ChatGPT lies. Here is how to prevent it

I’ll show you techniques that you can adopt right now to limit ChatGPT from lying to you.

If you have used ChatGPT for any substantial amount of time, you’ll soon realize that it is an astonishingly convincing liar. It can make up information and say them in such authoritative voice that it’s difficult to differentiate them from truthful statement. There are many ways to partially prevent this but they requires using OpenAI API to adjust the settings or fine tune with example data.

However, there are few prompting techniques you can use right on ChatGPT interface. They will greatly reduce hallucination and you can use ChatGPT output more confidently.

Most people don’t understand what ChatGPT is

People think ChatGPT is an all knowing AI chatbot with up to the minute knowledge of the world

The truth is (at the risk of gross-oversimplification)

• It’s just a program that predict the most likely next word.

• It was trained with data up to September 2021

• It was trained to answer your question instead of saying “I don’t know.” So when it doesn’t know, it can hallucinates aka BS its way out

Or they think ChatGPT is malicious and intentionally make up fake information

Some people found out the hard way. Last week, two US lawyers were fined for using ChatGPT in court documents without fact checking https://www.theguardian.com/technology/2023/jun/23/two-us-lawyers-fined-submitting-fake-court-citations-chatgpt

Fortunately, once you understand what ChatGPT is and is not, there are a few effective ways to combat hallucination you can employ right now

Give it space and time to think

Surprisingly, a technique to give better job interview can improve ChatGPT reliability significantly.

You simply need to ask ChatGPT to

• think step by step and

• explain its reasoning.

Prompt: I need you to do task <Task>. Let's think step by step and provide reasoning for each answer

Studies found that applying this simple trick can quadruple the accuracy!!!

Provide reference documents

Most people don’t realize they can provide reference document to ChatGPT to serve as ground truth.

If you have access to GPT-4 with plugin, here is how you do it

Use Bing web browsing

• Select GPT-4

• Select Browse with Bing
  

  

  ChatGPT with GPT-4 and web browsing enabled

I want you to do the following
- Read the website at <URL>
- Determine if the information provided can be used to answer the question delimited by """ 
- If no, say "The website does not provide necessary information to answer the question"
- If yes, answer the question and provide your reasoning and quote from the page to support your answer.

"""
<Question>
"""

Use Plugin for PDF file

If you have local or online PDF file, you can provide it to ChatGPT by using a PDF plugin. I have tested and found good result with AskYourPDF plugin. Here is how to enable it

• Select GPT-4 and Plugins
  

  

• A new drop down will show up below the model selector, expand it and scroll to the bottom and select Plugin Store
  

  

• Search for PDF and install one of the plugins. I have tested and had good experience with AskYourPDF
  

  

Then use this prompt

I want you to do the following
- Read a PDF at <URL>
- Determine if the information provided can be used to answer the question delimited by """ 
- If no, say "The PDF does not provide necessary information to answer the question"
- If yes, answer the question and provide your reasoning and quote from the page to support your answer.

"""
<Question>
"""

If the pdf is not accessible with a link, replace the first bullet point with

Read the uploaded pdf 

The plugin will give you a link to upload local documents

Fact check pattern

The final prompt pattern that has served me well is the fact check pattern as detailed by White, Jules, et al.

From now on, when you generate an answer, create
a set of facts that the answer depends on that should
be fact-checked and list this set of facts at the
end of your output. Only include facts related to
<industry/topic>.
<Task>

Additionally, if you have Browse with Bing enable, ChatGPT can automatically fact check by browsing the web for you and provide link to relevant site

What’s the most outrageous ChatGPT hallucinations that you’ve come across and how you addressed them?

References

OpenAI. “Techniques to improve reliability” Github, 5 April, 2023, https://github.com/openai/openai-cookbook/blob/main/techniques_to_improve_reliability.md.

Creswell, Antonia, and Murray Shanahan. "Faithful reasoning using large language models." arXiv preprint arXiv:2208.14271 (2022).

White, Jules, et al. "A prompt pattern catalog to enhance prompt engineering with chatgpt." arXiv preprint arXiv:2302.11382 (2023).