Is Your AI Hallucinating?

Is Your AI hallucinating?

The term “AI hallucination” is new to me. Is the AI on acid? Perhaps it is to some of you.

Hallucinating is when the AI creates inaccurate information. Sometimes spectacularly inaccurate information! Is it eating shrooms?

Case in point. We engaged an AI company to generate meetings with leaders of companies that might find our services useful.

It’s been amazingly successful, resulting in about fifty truly good meetings and one paid engagement in a couple of months’ time, with lots of follow up to come.

But yesterday I had a message from a woman who was angry about the email she received from my company, generated by the AI. Here’s one part of her critique of the email:

“I do not work for a company that the email references as “in the building and construction industry.”  It then goes on to reference my work history at two . . . companies that apparently ‘contributed to significant cash windfalls reported by food & beverage firms and medical manufacturers.’  WHAT???????”

The rest of the email was almost as crazy. The ai was high!

When I reported her complaint to my vendor company, they said their hallucination rate was 1%. I have since learned that number is highly unlikely.

According to an article in Silicon Angle, Vectara rated GPT-4 LLM the best with a hallucination rate at 3% and Google’s PaLM 2 for Chat the worst at 27%. They added that most major LLMs run between 5% to 10% hallucinations.

I have seen hallucinations in responses produced by ai sources. For example, footnoting data that was clearly invented, as well as made-up footnotes.

What does it mean to sales leaders and their teams?

1. When you choose an ai LLM product, find out its hallucination rate before you buy.
2. Don’t trust the ai to give a true or real or legitimate answer, knowing that some percentage of the time it’s on a high! Review everything that’s going public. Don’t believe data or footnotes. Check them out.
3. If the public is using ai that you provide, provide a disclaimer about errors.

Has your AI gone off the rails?  Please share it in the comments.