Control AI Hallucinations
Earlier posts, Using AI Content Generators for CIE Planning and Chess Defender: Mixed Feelings about AI, describe ways in which AI content generators serve as tools for teachers. Each contained a common caution: beware of “hallucinations”, that is the propensity of current content generators to make up their own facts.
So why don’t chatbots proof check their own output? For now it seems they need some human help. The good news is that there are several ways you and your students can mitigate the risk of hallucinations. ChatAI, a company that advertises access to multiple AI models in one place, offers several queries (1, 3, 4, and 5 below) to control hallucinations. If you are using an AI content generator/chatbot that allows for follow-up queries, try using these to proof check its response. While there are no guarantees these will work with every content generator, it may increase the confidence you can place in your generated text.
- Distinguish which facts and opinions are supported by reputable entities:
Example: “Give me a list of the fundamental facts on which your response relied.” - Use a search engine to check facts and citations from generated content. This isn’t full-proof. There have been instances reported where AI-generated hallucinations make it past the filtering algorithms used by search engines. This is most often happens when the topic content on the Internet is sparse.*
- Offer specific facts and have the content generator craft responses accordingly. The additional guidance limits the scope for inaccuracies:
Example: “Base your answer on these facts [insert the information you want to use].” - Ask the content generator to break down complex issues into smaller segments.
Make them easier to grasp and examine. At the end of a prompt, add an extra line asking your chatbot to “think step-by-step,” to break down the output. - Guide refinements to a questionable response by specifying updates or modifications:
Example: “Rewrite with these changes in mind [insert the needed changes].”
Using these prompts, you can unlock a new level of credibility for your AI-generated classroom content.
