Grounding Your Gullible AI

The other day, ChatGPT told me that “The phrase ‘don't hard code a coder’ is a clever, metaphorical play on programming language, and it can be interpreted in both technical and philosophical ways.” Which is interesting, because I just made it up on the spot. Though I appreciate the compliment. The AI also thought my phrase “goes together like turkey and almonds” was "unusual" but still proceeded to break down the interpretation for me. Once again, I had just made it up.

Your AI can’t do drugs, but it does hallucinate.

This has been making headlines of late, the fact that AIs are generating outputs that are factually incorrect, nonsensical, or misleading, and yet presenting them with the confidence as if they were accurate. It’s becoming common enough that the phenomenon has a name (hallucinating), as well as a name for the strategy of fixing it (grounding).

We’ve gone from Gen X parents telling their kids to not believe everything they read on the Internet to large portions of the population taking what an AI is saying as 100% fact. And this is an issue that companies have to address. Because AI is in its gullible stage.

Of course, there are the viral incidents of telling lawyers false case results or customers the wrong loyalty information. But even if the mistakes aren’t that drastic, they still have major ramifications on the customer experience.

The Challenge of Advanced Thinking

For decades, computer programming involved purely mathematical, step-by-step instructions where the computer did exactly what the developer said and nothing more. I’ve taught computer science classes where we make peanut butter and jelly sandwiches with the students giving us the step-by-step instructions. I bet you can imagine how that went. Their instruction to “smooth the peanut butter on the bread” led us to roll jars of peanut butter on a loaf of bread.

The lesson we wanted to illustrate was that a computer doesn’t know how to do anything until you write the step in terms of what it knows how to do. You have to break down the problem and be precise.

The benefit of this new wave of machine intelligence is that these large learning models make it so you don’t have to speak its language. It speaks yours. You don’t have to worry so much about literal interpretations. But now there’s a new challenge that’s on the other end of the spectrum, where you have to add phrases to make sure it doesn’t fill in the blanks with nonsense.

Some tools, like Notebook.LLM for example, only ground their responses in the data you provide. But that’s not the case for most. ChatGPT and Gemini pull from everything and anything. Which is why grounding your AI tools is critical. 

The Key to Grounding

As I mentioned above, there are plenty of use cases that won’t make headlines, but are still problems for brands whose AI isn’t grounded. If you want an AI to support your online store, it has to be grounded in information only from your store. Providing information from competitors doesn’t help you. Similarly, data versioning and permissions are an issue. How do you ensure that the AI isn’t pulling an old document that has since been updated three times or is only for certain employees?

The answer is a RAG (retrieval augmented generation) system. Since AIs are only as good as what they have been trained on, a RAG allows the AI to access certain external information and distill it back to the user. Major tools like Salesforce’s Agentforce utilize a RAG to ensure the response is more up to date and accurate.

A RAG is the way to honor role permissions, understand which document is a new source of truth (while forgetting older docs), and overall be a line of defense in ensuring accurate information. It provides you with an authoritative data source that allows the AI to retrieve the facts it needs to give the right answer.

Arm Your AI

The more you leverage an AI, not only is it continuing to learn, but you’re continuing to expand the quantity of information. The AI itself and the pool of data it pulls from will change on a daily basis. Take steps to keep the information it's pulling is up to date and accurate to reduce hallucinations.

The last thing you need is your AI giving out your CEO’s salary information or details around a loyalty program that no longer exists. I’ll be diving deeper into more of the specifics and nuances of RAGs in upcoming posts, so make sure to check back on the blog. And if you want to learn more about grounding and enhancing your AI’s performance and experience, reach out to us.