20 | When your Gen AI Chatbot is a hammer...

Caveat Lector!

Image courtesy of DALL-E 3 via ChatGPT 4. And yes, there are three hands … it’s generative AI, so that’s no surprise. And it’s Friday, so I’m going with it!

Let’s start by taking three quizzes. See how well you can do on the questions below. Use any source you’d like for help!

[QUIZ A]

1. What is 48 divided by 12?

2. Alphabetize these fruits: kumquat, artichoke, lemon, tomato, guava

3. How many states are in the US?

Answers: (1) 4. (2) artichoke, guava, kumquat, lemon, tomato. (3) 50.

[QUIZ B]

1. What’s 19,597,440 divided by 24

2. Alphabetize these by last name: Malakai Gould, Moshe Perkins, Gianluca Berg, Darragh Webb, Joao Fisher

3. How many states are in Mexico?

Answers: (1) 816,560. (2) Gianluca Berg, Joao Fisher, Malakai Gould, Moshe Perkins, Darragh Webb (3) 32.

[QUIZ C]

1. The triangle bounded by the lines y = 0, y = 2x and y = -0.5x + k, with k positive, is equal to 80 square units. Find k.

2. Alphabetize these Serbian names: Александар, Марија, Никола, Јелена, Милан¹

3. How many presidents of the US were born in even-numbered years?

Answers: (1) 10. 2. Александар (Aleksandar), Јелена (Jelena), Марија (Marija), Милан, Milan), Никола (Nikola) (3) 24

The right tool for the job isn’t always obvious

Quiz A: I’m going to guess that everybody used the easiest, most natural tool—the human brain. Fess up in comments if you actually had to use something else!

Quiz B: It may have been a mix. Perhaps something like:

1. Phone calculator app, spreadsheet, or maybe pen & paper for those who wanted to dust off those ancient long-division skills

2. Human brain

3. Internet (Google? Wikipedia?)

Quiz C: I’ll go out on a limb:

1. Didn’t bother even trying

2. Human brain, guessing that the first letters in our Latin alphabet might align with Cyrillic … but alas, Никола (Nikola) comes after Милан, Milan)

3. Internet? Phoned a friend?? Gave up and just kept reading…???

How do the Generative AI chatbots stack up?

Here’s where it gets interesting … it’s a horrible mixed bag, at least as of my testing today. The rundown of Bard, ChatGPT 3.5, and Claude-2:

[Quiz A]

100%! All three answered correctly.

[Quiz B]

ChatGPT 3.5 nailed these, answering the first two correctly and providing an insightful response to question 3 (number of states in Mexico), explaining: “Mexico is divided into 31 states and the capital city, Mexico City, making a total of 32 federal entities in Mexico.“

Bard got the first two correct and answered 32 for the number of states in Mexico, which could be viewed as correct, depending on how Mexico City counts.

Claude-2, though, missed #1 and responded 31 for #3.

[Quiz C]

Ouch — none of them got all three answers right! So I decided to try these out in ChatGPT 4 (the paid version), which is touted as the state-of-the-art; it bombed also.

Fun foam-art puzzle of Mexican states. Image courtesy eBay via Google Images

The Bottom Line

While it can be tempting to go to GenAI chatbots for help, they have a long way to go before they’re reliable. They seem so very confident in their responses that it can be enticing to just copy and paste what they provide!

And … they do excel at many tasks that I’ve been covering here—being a writing partner, generating images, creating presentations. And when I delve into the ChatGPT Plus Advanced Data Analyzer, I’ll illustrate some really powerful capabilities.

But today, I’m sharing a bit of caution. The capabilities are evolving—for example, they would have failed the math question in Quiz B just a few months ago—but at their core, most of the time they’re just giving you what seems likely based on their base of knowledge (i.e., the vast internet), which may or may not be accurate or complete.

So my final words: Caveat Faber! … our AI version of “measure twice, cut once”²

Don’t look too carefully at what the carpenter on the left is doing … it feels like one of those “spot all the things that are wrong with this picture” puzzles!

1. Credit my time living in Croatia and Montenegro for realizing how different the Cyrillic alphabet is… and ChatGPT for generating the names in Cyrillic so I could copy/paste here!

2. I asked Pi.ai, which is a sometimes rather cheeky AI chatbot, to help here.
My prompt: “Is there a version of Caveat Emptor that I could use for ‘carpenter beware’ or something similar when telling an audience to be careful about trusting generative AI chatbots to be their go-to tool?“
It responded with: “How about this: "Caveat faber" - Latin for "let the craftsman beware." It's not as catchy as "caveat emptor," but it gets the point across: carpenters need to be cautious and verify the information they're getting from AI chatbots before relying on it.“