Cartoon Caption Contest Analysis | The New Yorker Style

The Data: A Catalog of Failed Dreams

Or: A comprehensive archive of 2.9 million people briefly convinced they are funny.

                384 Cartoons

                ~2.9 million Captions (Desperate attempts at humor)

                ~287 million Votes (Harsh truths)

Every week, the New Yorker throws one cartoon at the public. Thousands of captions follow. Only a handful survive. The rest are data.

The Ingredients

The Cartoons:
Surrealist drawings where the Grim Reaper is wearing a tracksuit, a snowman works retail, and a teddy bear is questioned on the witness stand. You know, high art.
References: Contest #541, Contest #595, Contest #732.
The Captions:
An endless scroll of puns. Some are gold. Most are just sentences that end with an exclamation point for no reason.
The Verdicts:
Crowdsourced judgment delivered with brutal efficiency:
- Unfunny (Roughly three out of four captions die here)
- Somewhat Funny (A polite nose exhale)
- Funny (Actual laughter, statistically rare)
The Metadata:
Detailed descriptions of the scene in the image, and what's uncanny about it. Because computers have no soul and need to be told that a dog sitting at a bar is unusual.

Why We Analyzed This

With millions of captions and votes, this dataset allows us to quantify humor, compare audience taste, and confirm a long-held suspicion: comedy is subjective, but not that subjective.

Lesson 1 : Structure, the skeleton of comedy

Brevity is the Soul of Wit.

You know that friend who takes ten minutes to get to the punchline ? Please, do not be that friend. Let's take a look at what science says : our structural analysis revealed a groundbreaking truth. Size doesn’t matter.

Average Word Count

From the graph you might say there is a peak at 10 words, but don’t be fooled: the cloudiness of the data means there is no correlation between the success metric and the amount of words used. In other words : We statistically proved that there is no link between our success metric, meaning our funny score, and the amount of words used.

In case you were wondering, we did the same study per theme, to see whether when talking about politics or food you ought to keep it short and sweet. And, plot twist, there is no secret recipe there either...

                        So please, believe us when we tell you to  trust your gut and not follow a single number,
                        just  do not do ramble on. Do not be that friend.                    
                    

Secondly, we hypothesized that the funniest people might employ a complex vocabulary. To test this, we calculated the Type-Token Ratio (TTR) to measure vocabulary richness across millions of entries.

Mathematically (and very scientifically, of course), the TTR is expressed as:

$$ \mathrm{TTR} = \frac{V}{N} $$

Where V is the number of unique words (types) and N is the total number of words (tokens) in the text. One is solely unique words, zero is solely repeated words.

Funny score in terms of TTR metric — Figure 1: Correlation between word count and success metric.

The Result: Everyone is Shakespeare. Or everyone is a five-year-old. The average TTR was 0.99 across the board. There is effectively no difference in vocabulary richness between the winners and the losers. To illustrate our point, look at the above figure, that red line looks pretty flat huh ?

"Using sesquipedalian verbiage does not augment comedic efficacy."

Translation: Big words don't make you funny, nor do simple words.

Lesson 2: Themes, Comedy's Comfort Zones

Comedy loves familiar territory. Relationships, family, social awkwardness, technology. These themes keep coming back for a reason: they’re relatable, emotionally charged, and easy to recognize.

In this lesson, you’ll learn which topics are comedy’s safe spaces.

This chart is basically a cheat sheet for what New Yorker captions keep talking about. Each big bubble is a popular theme, and the smaller bubbles inside it are the words that theme reaches for when it’s trying to make sense.

The rule is simple. Big bubbles mean lots of captions. Big little bubbles mean important words. Not “funniest,” just the ones doing the heavy lifting behind the scenes.

Click around. Poke the bubbles. You’ll quickly notice something comforting: even comedy has habits.

Wanna see some more ? Here is the same map, but taking into account how those themes are linked :

Most jokes come from the same ideas.
Once you know what repeats, you’re halfway to funny.

                So, if you thought you were going to be the first person to make a joke about Trump... Think Again.

Having identified the main themes in both images and captions, we can now look at how they interact.

To understand how image themes and caption themes interact, we turn to the alluvial graph below.

At first glance, the diagram looks like spaghetti. This is accurate. It’s showing many things at once. (To avoid complete chaos, only the strongest connections are shown.)

Image themes live on the left; caption themes live on the right. The flowing bands between them represent captions linking an image theme to a caption theme. Thicker bands mean this pairing happens more often.

Take the theme T68: restaurant as an example. This image theme appears in 44,525 captions in total. These captions then fan out across several caption themes. One particularly confident connection links T68: restaurant to T28: bakery, accounting for 8'289 interactions between the two. Hover on this connection, and look for yourself. We can also see that T28 bakery is a popular-ish caption theme, as it is connected to 15'901 images .

The plot also reveals a tendency that some themes are strongly related to themselves, thus self-pairing. This may not be groundbreaking, but it is reassuring: sometimes, a restaurant remains committed to being a restaurant.

To visualize this better, let's take a look at this heat map below :

See that dark grey diagonal line ? Those are all the self-pairing instances. The restaurant-restaurant thingy is on that diagonal.

The obvious next question is : Which of those combinations are actually funny ?

The obvious next answer is another heatmap.

Quick Guide : Red = Funny combination. For example, Image theme: T69 : rotation VS Caption theme: T5:Acknowledgment is bright red. Hilarious right ? Let's double check that statistically :

What does that ~~Fox~~ Math say? Well, a few things:

For self-pairing themes, popularity does not equal comedy: the most frequent themes are not the funniest ones. Do not be a sheep describing the image of a sheep, it's not enough to win anyways.
Plot twist: even across different captions, some image themes stay consistently red. In other words, a good cartoon can do most of the work, regardless of the comedian’s talent. That T69 rotation-themed cartoon probably carried the joke on its own. Also, it’s number 69 ...
Final reality check: making the caption wildly unrelated to the image does not magically make it funnier, chaos is a very unreliable comedy strategy. That said, it will make you stand out, since almost no one else tries it. Is that a win? …ish.

            Final takeaway:  same-theme captions don’t add much, good images carry bad jokes, and unrelated captions mostly just make you memorable, not funnier. 
            

Lesson 3 : Semantic Analysis, Meaning Under Examination

Semantics is just what words mean.
Not how fancy they sound.
Not how they’re arranged.
Just: when you say something, what are you actually talking about?

To impose some order on the famously unruly universe of caption contests, we used Berger’s humor categories, borrowed from a man who looked at comedy, thought “this needs diagrams,” and wrote an entire book about it.

The categories:

Language: jokes that live by words alone : irony, puns, allusions, repartee, wit, and ridicule.
Logic: jokes powered by ideas rather than phrasing : absurdity, coincidence, conceptual surprise, and misunderstandings.
Identity: jokes that turn the lens on people and personas : parody, existential humor, condescension, self-deprecation, and other small human failures.
Action: jokes that skip language entirely, relying on physical or non-verbal cues like awkward bodies, odd expressions, clumsiness, or general chaos.

Using Ollama and the gpt-oss:120b model (Fancy words just to mean....AI), we analyzed the 25 best and 25 worst captions from each contest, a practical compromise given the computational cost. Captions could belong to multiple categories, with the model assigning percentages accordingly.

References: An Anatomy of Humor, Arthur Asa Berger (1993)

Our results point to a humbling truth: there is no cheat code for comedy. All humor categories perform about the same, which means being funny is less about what kind of joke you tell and more about not messing it up. This is visible by taking a look at those bar plots : They all yield similar results in terms of mean funniness.

Language gets a honorable mention for being the most popular humor category, with around 12'000 captions. That's a lot of, probably, mid puns.

Wondering how these categories interact ? Let's take a look at our map below :

One pattern does stand out on our second plot. Language and Logic don’t like sharing the spotlight, with a mediocre score of -0.87. What does this mean ? Either a joke is in the Logic category, or it is in the Language category. Jokes usually twist words or ideas, not both. Try to be clever in two ways at once and the meaning trips over itself. Comedy, like traffic, works best when everyone stays in their lane.

Also, look at the line or column about the Mean Funniness ! Notice anything ? Right, those numbers are tiny. Meaning, there is no clear winner for the Funniest-Semantic-Category Award, once again.

Further very scientific analysis has proven that Identity and Action mostly play supporting roles. A little self-awareness or physical awkwardness happens often, but rarely carries the whole joke unless someone is actively falling over.

 
            Long story short: Keep it clear, keep it simple—and remember: if the audience is busy decoding your joke, they’re not laughing at it. 
            

Lesson 4: Wait for it....Timing is Everything

They say comedy is about timing. In the case of the New Yorker Caption Contest, it's about being stuck at home with nothing else to do.

We plotted the number of captions submitted over time, and the results are about as unsurprising as you'd expect.

Notice that massive increase in yellow? Yes, that's circa 2020... that's the COVID-19 pandemic. Apparently, when the world stops turning, everyone decides they are a comedian.

Still, this doesn’t quite settle the case for timing. Popularity spikes, funniness drifts, but correlation remains hard to confindently establish. Maybe the real story isn’t when people submit captions, but what they’re writing about.

So let’s put the top 10 caption-themes in motion. Below, they race through time according to their mean funniness. Place your bets !

If you put your money on Justice, congratulations. Its sharp rise between 2018 and 2019 coincides with a period dominated by public reckonings, trials, and questions of accountability. The most immediate association is the #MeToo movement, sparked by Alyssa Milano’s call to action and the Harvey Weinstein case, although confirming a direct causal link would require a far richer external dataset.

Also, peep the COVID-19 theme in 2016 ! Surprisingly, the comedians did not acquire prophetic powers overnight. The so-called theme is merely a semantic umbrella for harmless words like sunscreen... or vaccine....and masks, terms that aged, unfortunately, very badly.

 
                    So… maybe next time there’s a global catastrophe or a societal wake-up call, you should submit a caption.                   
            

The Perfect Recipe

So, how does one win the New Yorker Caption Contest?

After millions of captions, hundreds of cartoons, and an unreasonable number of statistical tests, the answer turns out to be both disappointingly simple and mildly comforting.

Keep it short and simple. Fancy vocabulary won’t save a weak idea, and plain words won’t doom a strong one. Same foes for brevity : Size doesn't matter, but please don't be that friend that rambles.
Keep it comfortable. Don't overthink it : use the usual themes for your caption.
Keep it singular. Jokes work best when they commit to one clear idea instead of juggling multiple meanings.
Be lucky. Timing, taste, and the collective mood of thousands of strangers matter more than we’d like to admit.

And if all of this feels a bit unsatisfying, that’s because humor is stubbornly hard to quantify, which may be its greatest feature.

"Analysis kills spontaneity. That's why we did it."