A couple weeks ago, I had a thought: It would be incredible if ChatGPT is creative enough to win the New Yorker cartoon caption contest.
It turns out of course that I’m not the first person to think of using ChatGPT to write cartoon captions, but I couldn’t find evidence that anyone actually submitted them as entries to the weekly contest.
So I typed up a description of that week’s cartoon (depicting a duck dressed as a pilot speaking to a passenger preparing to board a plane) and asked GPT-4 to write a bunch of captions, using Bing Chat’s “Creative” mode. Some of them made no sense. But among them was one that I deemed funny enough to enter. It tapped into something everyone can relate to — what a headache air travel is these days — and ends with a dark twist. This was my entry, using Bing’s suggestion verbatim:
It did not get selected as a finalist.1
I only planned to write about this experiment if I actually got one of my A.I.-written captions selected. But I have a strong suspicion that someone else beat me to it and already did make it to the finals.
I present my evidence.
This past weekend, I decided to go back to some earlier cartoons from the contest and ask Bing to write captions to see what it comes up with.
This is the cartoon from the contest three weeks ago:
So I wrote to Bing:
Write ten captions for a single-panel comic. It depicts two scientists looking at mice or rats in a maze. But instead of an ordinary maze, the rodents are in a rat-size art museum. They are walking on two legs, looking at the paintings. One scientist is holding a clipboard. The other scientist is speaking. The caption should be whatever that scientist is saying. Think abstractly.
Here’s what Bing came up with:
Most of those are pretty awful. But that one second-to-last isn’t bad. “They’re not looking for the exit, they’re looking for meaning” feels like a good New Yorker caption. If I were entering the contest based on these suggestions, that’s the one I would pick for my entry.
I wondered what the finalists actually were. Let’s take a look, shall we?
Well, well, well, William Shapiro of Davis, California. That caption seems awfully familiar.
But wait a minute. There’s some randomness in Bing’s answers. How possible is it that William Shapiro and I both asked Bing for a caption and it gave us both the same suggestion? Was this really repeatable?
Only one way to find out. I pasted my original prompt into a fresh chat window several times. At first, I didn’t get the exact same answer. But I did realize it was giving me a lot of variations on “They’re not [x], they’re [y]” including:
They’re not just rats, they’re critics.
They’re not just looking at the paintings, they’re studying them.
They’re not mice. They’re artists in disguise.
They’re not interested in the paintings. They’re looking for the exit.
And so on.
Finally, after a few more attempts, it gave me the same “They’re not looking for the exit, they’re looking for meaning” caption again. Ah-ha! It was repeatable!
But Bing has access to the internet. Is it possible that based on my description it was able to find the original New Yorker cartoon and the finalists, and used William Shapiro’s entry as the basis for its suggestion?
I doubted that for several reasons. First, I never mentioned The New Yorker or that it was an existing comic. I was asking Bing to be creative, not do research. Second, Bing lets you know when it’s looking things up online and cites its sources, neither of which it did in this case. And third, it only gave me that one finalist as a suggestion. It didn’t give me any versions of the other finalists, which it would have also found.
But just in case, I went back to the original chat and asked it:
Still, maybe it was lying. It’s been known to do that. So I repeated this experiment with several other recent New Yorker caption contest cartoons. And not once did I get a response that was at all like any of the finalists. So I doubt it was looking them up.
So this leads me to conclude one of two likely possibilities:
William Shapiro of Davis, California did the same thing I did, giving Bing or ChatGPT a description of the cartoon and going through suggested captions until he found one funny enough to submit, which then actually got selected as a finalist by the New Yorker cartoon editors.
All on its own, Bing came up with the same caption a human came up with that was good enough to be selected as a finalist by the New Yorker cartoon editors.
Either one of those scenarios blows my mind.
As you ponder that, I leave you with this recent New Yorker cartoon:
And that’s it for another newsletter! I voted for Doug Finkelstein’s caption “Nothing has sold yet, but we’ve gotten a few nibbles” which is definitely the funniest of the three finalists. The winner will be announced on the New Yorker website next week.
If you missed the announcement in the last issue, the big news is that Ironic Sans is now a video series! You can catch the first episode on YouTube or Nebula. I’m working on the second episode now, but it probably won’t be finished until after I return from a trip next week. Subscribe to the channel so you don’t miss it!
Thanks as always for reading. See you next time!
David
It’s against both the spirit and letter of the cartoon caption contest rules to use A.I. in your submission (“Entrants represent and warrant that their Submission is their original work”), and I feel kinda bad about it now. I was looking at it as a fun test of A.I. creativity, but it’s not cool to use A.I. against people using actual creativity to win a contest.
First, are you *sure* creative Bing doesn’t look stuff up? I’m pretty sure that the searches I’ve done on Bing aren’t listing all of the sources it used.
Second, *why* isn’t it “cool to use A.I. against people using actual creativity,” to win a contest or otherwise? Isn’t that now pretty much what AI is for? Not saying I like it (I’m actually petrified, unless our robot overlords are listening in—in which case I ACTUALLY LOVE IT HerrGPT!), but I do think “cool” in creativity is slipping away from humanity’s grasp.
I enjoyed reading your comment. It was great to know others may be using ChatGPT. I have submitted one that ChatGPT wrote and planned to tell the New Yorker that Chat was the author of the caption if I won. I thought it was a great caption and it was within the last few months. I cannot recall which one it was, however, as I submit almost weekly. I also ask Chat about the NPR Sunday Puzzle for suggested solutions. Chat is not that good at puzzles like that. It did solve one a few weeks ago that I never would have gotten, though. I just had to reword the puzzle to get the correct answer.