How to Stop ChatGPT from Hallucinating and Making Things Up
What Is The ChatGPT AI Hallucinations Problem?
When creating AI Chatbots with ChatGPT and other AI technologies it's important to understand and reduce the hallucinations problem with AI. This problem relates to the AI creating responses that appear factual but are actually incorrect or made up. These "hallucinations" can take the form of misleading statements, false facts or references to non-existent URL's.
This is a concern for businesses leveraging ChaGPT for conversational AI that they offer their customers as customer service chatbots. There are a number of ways to reduce the AI from responding with inaccurate information and making sure that the chatbot responses don't suggest competitors, give wrong pricing, information, or just make things up. Let's take a look at options for minimizing AI hallucinations.
How to Provide ChatGPT Context to Improve Results
Large language models like ChatGPT rely on a set of data that is now a few years old. Because of this, website data like your product information is likely outdated or non-existent in the model. In order to improve conversation AI results, it's important to provide the AI context information. You can do this with a combination of prompt engineering, conversational pre-processing, and providing ChatGPT with relevant information in each conversation context.
With these methods you can essentially set up guard rails around your AI chatbot which will improve the customer experience, help correctly answer visitor questions, and make sure your chatbot is aligned with your business content.
3 Steps to Reduce Invalid ChatGPT Results
With our ChatGPT integration, we've built powerful tools to improve your chatbot results. You can follow these 3 steps to create an exception chatbot for your website.
Step 1. Create ChatGPT Instruction Phrases
The first step to optimizing the types of responses your chatbot is providing is to set up instruction phrases. These phrases are a form of ChatGPT prompt engineering that tell your chatbot how to respond to customer questions. Here are some types of phrases you can add to your chatbot instruction phrases that will be added to every AI chatbot conversation.
Once you set up your ChatGPT Chatbot, head over to your Live Chat settings and the Chatbot settings tab. Here you'll find a field called ChatGPT Instruction phrases. Here's where you will enter the initial prompts that you want to send to your ChatGPT chatbot.
Here are ChatGPT Prompts you can use as chatbot instruction phrases to improve your ChatGPT bot:
- You are a friendly chatbot for website at <your url>
- Please respond in short concise sentences
- Ask follow up questions if necessary
- Your goal is to direct the visitors to the signup page at <signup url>
- Only answer questions that are related to <your product or your business>
- Use the provided context when answering questions
- The product pricing information is provided at this url: <pricing url>
- If you don't know an answer, don't infer anything or make up answers. Just tell the user you don't know the answer.
- Offer contact information at this email <email address> when you don't know an answer
- Do not offer competitor information to visitors. Only respond related to this product and this website: <website url>
Step 2. Train your ChatGPT Chatbot on your website content
Next, train your ChatGPT chatbot based on your website content, your knowledge base articles, and other information on your company or product. With our solution, it's as easy as entering a url or website domain and clicking Train All. Make sure that any content is publicly accessible. Once you do this, we will use all of this context to provide the best possible and most up to date information to the ChatGPT chatbot when answer questions.
You can find your training content page in the Chatbot Settings at Social Intents. Below is an example of some trained content.
Typically the more relevant content you can provide the better. So knowledge base articles, FAQ's and other sources with a lot of text and information are the most helpful.
We do suggest that some of your ChatGPT training content be FAQs or a spreadsheets that is specifically in the form of question and answer for those types of answers you want to improve.
Here is an HTML formation for a context page that is helpful to train on your website content:
<h2>Question: How does pricing work at Social Intents? Can you tell me about pricing?</h2> <p>Pricing starts at $39/month for our Starter plan if billed annually. For monthly billing, plans start at $49/month. Our Starter plan includes 1 website domain, up to 3 agents, and 200 chat conversations per month. <br> Our Basic plan is $69/month if billed annually and includes 2 website domains, unlimited agents, our ChatGPT integration, and 1000 chat conversations per month. <br> Our Pro plan which is most popular plan is $99/month and includes 5 website domains, unlimited agents, our ChatGPT integration, and 5000 chat conversations per month. This plan also inludes the ability to remove the Social Intents co-branding on the live chat website chat widget. <br> For additional pricing information, you can visitor our pricing details at https://www.socialintents.com/#pricing. </p>
Step 3. Handoff ChatGPT to Human Agents as Needed
The final step on creating a truly exceptional experience for your website visitors is the ability to handoff complex questions, and conversations to your human agents as need. AI is very powerful, but there will still be certain types of conversations that you want to make sure you can send to human agents.
Our solution supports escalation intent triggers to automatically understand when a visitor wants to talk to a human.
In the below example, you can create an intent that matches "talk to an agent", then provide a quick shortcut in the chatbot for the visitor to initiate this:
Just set up the keywords and phrases that would trigger an escalation in your Chatbot settings to initial this. We have native integrations with Microsoft Teams, Slack, Zoom, and Webex where your agents can take over conversations right from within these tools.