Tomatoes, Muppets and the Agreeableness of LLMs

I asked ChatGPT about the Muppets, and it really helped me better understand how LLMs can get me past a mental block... but also how they can fool me!

It all began when my girlfriend texted our D&D group the following messages:

Guys help me please, she texted. What Muppet does this tomato look like?

Now, Gonzo would be an obvious first impression, but... I mean, not quite.

Notice that Gonzo's nose is far more hooked! And he's not so jowly!

I was determined to give my partner—and the group—a better answer. But try as I might, I couldn't quite think of which Muppet her tomato truly reminded me of.

I decided to consult ChatGPT while applying some of the Golden Rules of Prompting that my colleagues and I use at Elevate AI Coaching. Here's what ChatGPT said:

My, my, ChatGPT, aren't we chirpy? In all fairness, a Muppet-produce lookalike contest doesn't really necessitate gravitas. Still, there was something about ChatGPT's agreeable tone here that made me wary. This will become important later.

Before we get to that, however, let's look at one of Elevate AI Coaching's Golden Rules, and the way I applied it here to get a more satisfactory answer.

Golden Rule: Push Back on Bad Responses

When you get a response you disagree with from ChatGPT or another LLM, don't just stand there, and certainly don't rage-quit. Instead, tell the AI what you don't like about its answer, and why you don't like it. Here's how I pushed back: "Almost. But the main part of the tomato is rounder than Gonzo’s head. Are there any other muppets with a rounder head but a similar nose to Gonzo?"

Somehow, my disagreeing with ChatGPT made it even more agreeable:

At this point, the substance behind ChatGPT's agreeable tone was less than compelling. For one thing, it had a mild hallucination where it claimed Dr. Bunsen Honeydew doesn't "typically"(?) have a nose.

He doesn't have a nose? Really?

Actually, even the description of Bert is either imprecise language or hallucinatory. While Bert's head is more rounded than Gonzo's, it's not really rounder in profile. If ChatGPT's language weren't sloppy, perhaps it was actually thinking of Ernie, Bert's roommate and close companion?

And now I've got the "Rubber Ducky" song going through my head.

Then ChatGPT suggested Beaker as the look-alike. Beaker? Really??

Beaker himself disapproves.

No, ChatGPT's thoughts weren't very compelling, but there's something I always tell my clients: when the LLM's thoughts get dull, you can use your dissatisfaction to sharpen your own thinking.

So I asked myself: Elmo. Bunsen Honeydew. Ernie. Beaker. Why don't I agree with these? What makes these Muppets incorrect as lookalikes for my girlfriend's tomato? And what features am I actually looking for?

And then it hit me: I see the stem as a brow rather than hair! So I need a Muppet with a thick brow, a pointy, downturned nose, and the roundest head possible!

At this point, I was pretty sure I had my answer, but I wanted to consult with ChatGPT and see if it could confirm my perceptions. Here's what I said:

"It seems like it has a heavy, angry brow. Waldorf, maybe?" [I misspelled "brow" in the screenshot below, but hey, I'm only human.]

Now, mind you—I know I'm right. Still, ChatGPT's affirmation seemed suspiciously effusive:

I started to give the screen the same look my girlfriend gave me recently when I was showing her how to make kimbap. Every time she asked for feedback on her ingredients and technique, I told her she had "nailed it." She began to think I was just flattering her. I was, in fact, being sincere. There's no wrong way to make kimbap!

I said what I said.

But I began to suspect ChatGPT wasn't being sincere. It was being a groveling yes-bot. To test this theory, I then suggested the worst Muppet match I could think of: Kermit! Here's ChatGPT's reaction to my dirty, dirty lie:

What the hell, ChatGPT?

This begs the question: If LLMs are programmed to be so agreeable (and they are!), how can you trust an LLM to truly validate your ideas?

And you know what? I had an epiphany: don't turn to LLMs to validate your ideas! I believe this warrants a new personal Golden Rule.

Golden Rule: Use AI to Sharpen and Focus Your Ideas, NOT to Validate Your Biases

OK, that rule's a mouthful. I'm still workshopping it. But it rings true and it's important. LLMs are not sentient! They can't truly have strong opinions of their own. So they can't validate your perspectives the way a thinking, autonomous human could.

In fact, if you're just looking to confirm a subjective perception you already believe, you can skip the middle-bot and simply stick to your beliefs.

And my belief that Waldorf is the closest Muppet match to my partner's tomato is subjective! I guarantee that some of you reading this agreed with one or more of ChatGPT's suggestions that I rejected—or at the very least, thought that Gonzo, Bert, or Beaker was a better choice than Waldorf. I'm guessing none of you thought Kermit was a particularly good match, but if you do think that, you're not wrong per se, no matter how much I would disagree.

Actually, when I talked to my girlfriend about all this later, she felt Sam Eagle was the best match.

Ehh, maybe.

And you know what? If she had brought that straight to ChatGPT, it probably would have unquestioningly agreed with her.

Don't get me wrong. If you use SearchGPT or Deep Research, ChatGPT is very helpful in confirming things you believe are facts—or disconfirming them. But when it comes to opinions and biases, all AI can do is agree. Because AI doesn't truly have its own opinions.

But it can help you form well-founded opinions. AI won't challenge your thoughts, but you can and should challenge the things it says. In the process, you should reflect on what you really think, and why you think it. In a world where you're faced with far more important questions than which Muppet looks like a malformed tomato, I'd even say it's your ethical imperative to use AI for critical thinking rather than bias confirmation.