Balderdash

When I was growing up, my family used to play a game called Balderdash. The goal of this game was to write a made-up definition for some bizarre word no one had ever heard of and get other players to believe your definition was true. We were all masters at this game because we were great at sounding like we knew what we were talking about. We could mimic the style, tone, and language of a dictionary definition so well that it was hard to tell when something was balderdash (def. nonsense). This is what I immediately thought of the first time I encountered an AI hallucination.

An AI hallucination is when the computer responds with nonsense that sounds 100% believable. It is believable because AI is capable of mimicking the style, tone, language, and format of other similar responses. This has caught many people off guard and even caused outrage, but it really should not surprise us. We are ASKING generative AI to mimic our responses...so, why would we be surprised when it does just that?

I knew that hallucination responses were something to watch out for, but I was still shocked when it happened to me. I was shocked because it happened with a calculation. In my head, a calculation is 100% accurate, true, verifiable, and concrete. I was certain that a computer would always tell me that 2 + 2 = 4 (and not 5!). 

The issue happened when I was using an AI chatbot to help me check my work on a multi-step precalculus word problem that involved algebra and trigonometry. When it's response was different than my solution, I asked it to show me the steps it used. I carefully checked my steps trying to see where I went wrong. Then I ran a simple calculation multiple times (tan 15˚ ). I was still sure that I was wrong - maybe I was in the wrong mode (radians/degrees). So I even went to Wolfram Alpha to check what it said the calculation should be. I was dumbfounded when it agreed with me! Gasp! How could the computer have been wrong about a simple computation? So, I asked the chatbot to tell me what it believed tan 15 ˚  was, and it came back with a groveling apology about how it had made an error, and I was such an amazing person to have pointed it out.

The apology did not help. Nor did the attempt to flatter me. 

I was shocked. It was like hearing that your favorite person was actually not a nice person at all. Or that they had been lying to you. It made me wonder, "What else have you lied about?"  I got over it, but I learned a valuable lesson that we should all know: 

Use the tool, but check the results.

Until next time, happy balderdashing!

AI hallucinated with a calculation.