It's Funny Because It's True

Humor, Reality, Artificial Intelligence

Pure fiction can be very funny. Fantasy and science-fiction with only a minimal connection to our lives can still entertain; The Hitchhiker’s Guide to the Galaxy is able to make a great scene of a talking cow at the end of time. Jokes from distant cultures and languages can still entertain with the right mix of context and cultural universals: according to the Tamil stories of the Ramayana, Indra, king of the Gods, was caught with the wife of a great sage who cursed the Lord to have his body covered in 1000 vaginas. P.G. Wodehouse’s novels never pretend to be accountings of actual events, and the absurd situations that happen between completely unbelievable characters are some of the funniest that have been devised.

But reality adds an extra quality to humor. Wodehouse’s satire of the upper class of England in the early 20th century is made more funny by the fact that Wodehouse deeply knew those people and could accurately skewer them. Were it revealed that early 20th century England was simply an illusion, the stories would change. Perhaps even more interestingly, if Wodehouse was mistaken about the nature of the upper class fundamentally, if there were not a single idiot son purely concerned about his fortune and instead 100% of them were noble and dutiful patriots, it would be less funny. The ending of Blackadder Goes Forth is one of the great moments in comedy television, made much more funny and enjoyable by the underlying truth of its depiction of the “lions led by donkeys” theory of The Great War. When I learned that this theory is actually quite suspect according to more recent historical scholarship, some of the greatness of the ending dissipated.

There are some species of humor that rely more heavily on reality than others. standup comedy starts with the premise that the comic is telling true stories about their life, but with an agreed upon and acknowledged generic convention that the stories don’t have to be fully true, and that fictional elements can be inserted. But the stories are made funnier because they are true. There is an extra edge to confession, an extra violation of social tension that is processed as laughter. I am fairly confident that John Mulaney went to a doctor and had his prostate examined and felt awkward doing it. The rest of the bit may be false, and I would be shocked if it were all true (but I hope that “Batman” was real! That would be much funnier!). But if he has never had a prostate exam, the joke would be lessened. On the other hand, the late Norm Macdonald had very little of what he describes actually happen to him, or at least that is what he communicates to his audience through subtle signals, and his style of humor does not rely on reality.

—

Some things are only funny if they are true. It is very funny that the United States Navy had to move the USS John McCain out of the sight of President Donald Trump during a visit to Japan because Trump was known to be so vain by his staff that a mere reminder of his political enemy was too much for him to stomach. It is very funny that the most powerful man in the world was also so thin-skinned. The fact itself is funny. The situation as a hypothetical could be funny or not. “Imagine if Obama was so vain that he wanted to move a battleship” is not funny. If P.G. Wodehouse wrote a story about a fictional British dignitary who was so vain that he ordered a battleship moved at great expense and panic, it would be very funny. If P.G. Wodehouse wrote a fictional story about Trump doing this, it would also be very funny, partially because it draws from the otherwise confirmed reality of Trump’s personality. Even Wodehouse probably could not make the Obama variation funny, but he could certainly write other humor with its basis in Obama’s real or perceived personal characteristics. “Imagine if Trump were so vain that he wouldn’t climb a staircase named ‘McCain’” would be disappointing compared to an actual reported fact.

Fake headlines are an excellent example of this phenomenon. So many times people will read Onion headlines out loud and the immediate response amid laughter is: “wait is that real?” If the answer is no, the laughter lessens, it never increases, it only rarely stays the same. They’re still funny, but they’re not as funny as they could be.

—

AI can be very funny. AI humor has recently taken off thanks to a rapid increase in text-generation powers, but also a much greater cultural awareness of “AI.” Just as when the newspaper misprints a dirty word accidentally, it is very funny that these seemingly all-knowing systems can make things that are so silly, so wrong. The humor comes from the reveal that the AI is fallible, but also what it reveals about the subject matter it is failing to understand.

One could, but I will not, make a joke about an immigrant from a wartorn country who runs panting back to his apartment from a brawl between dueling militias, feeling that things are just as bad as the place he left. It could be revealed that what he actually witnessed was two fan clubs of rival sports teams. The joke would come from the immigrant’s silly misperception, but also from the surprising reveal of the actual but hidden violence of sports culture. AI jokes work much the same way. What’s funny is the truth that is revealed by the misperception.

Take this notable early example from an AI trained by the comedy group Botnik on the Harry Potter series:

“The only sounds drifting from Hagrid’s hut were the disdainful shrieks of his own furniture. Magic: it was something that Harry Potter thought was very good.”

Here are the characters we know, reconfigured. It is very funny that the AI has put out such nonsense, but has also managed to understand that everything is magical in this world. Later, after Ron appears:

“‘If you two can’t clump happily, I’m going to get aggressive,’ confessed the reasonable Hermione.”

This is still random, still the AI messing up: Hermione confessing makes no sense, let alone “clump happily.” But it is much funnier that Hermione is criticizing Ron and Harry, remains confrontational as she is in the real books, and furthermore the AI has learned just how “reasonable” she is. An alternative “’Hi I like magic, would you like to smell Hagrid last Tuesday’ screamed the hungry Hermione” would not be as funny.

None of the above text would be funny if a person wrote it. It would reveal nothing about Harry Potter, nor anything about the writer. It would be random nonsense. And at least for me, all this is made much less funny by the fact that an AI only sort of wrote it. The RNN that Botnik trained did next word prediction: every time the AI produces a word, it looks at what it has previously written and gives several suggestions for what a good word to follow would be. It is a human who chooses which of the several suggestions. If the RNN were allowed to run on its own, likely the text produced would be not be grammatically coherent. While interested that “Hermione” was a suggested successor to “reasonable,” it doesn’t quite pack the same punch if there was a chance it also could have been a non-sequitur.

Most people did not know this about AI though. The Botnik humor AIs went viral and made millions of people laugh. It was clear that people found it very funny that AIs could produce such nonsense, and also reveal the absurdity of certain source material.

Take this example:

Here we have the familiar character Trump speaking hilarious nonsense, but nonsense that bears a real, slightly uncertain, connection to reality. Trump would never say these words, but these words are insane parody of words he might say. Replacing Obama’s name with Trump’s in the script would not be funny. And once again the tweet is made less funny by the fact that it was actually written by a human, not an AI, although this deception was not disclosed in an accessible way to the people who spread it. It would be foolish to train an AI on footage in order to produce a script, switching domains is technically hard and would require a whole set of advanced AI functionalities to even set-up the system.

It was much easier for many on the internet to use and abuse the wide misunderstanding of AI’s actual capabilities for their own benefit (and to successfully make millions of people laugh). There was not any sort of shared understanding that AI couldn’t actually do these things and it was humans doing it all. The humor was based on deception.

There were some who were honest. Janelle Shane accurately communicates the power and functionality of AI in her work, and makes legitimately very funny results. But such things rarely go as viral as the faked. Honest AI humor is possible, but AI’s often do not live up to the humorous vision of reality we demand.

—

AI is much better now. OpenAI’s recent GPT-Chat can make humorous things upon request, and sometimes unintentionally as well. Take this example:

I gave the AI an absurd prompt, and in a technological marvel it produced a highly coherent story. And it is funny that an AI produced it, funny because it has been trained on the combined text of millions of internet users and knows these cultural tropes and characters so well. If I wrote this, rather than an avatar of combined contemporary knowledge, it would not be as funny.

But this would be funnier if it were unintentionally funny, like the real events it is mimicking. This response is ultimately a copy of many internet users writing absurd stories about Trump. This revelation, that it is based on the reality of people making fun of Trump, rather than Trump himself makes it less funny. When you ask for similar prompts without adjectives like “funny,” it instead produces serious or inspiring stories which just fall completely flat. It would be much funnier if the AI actually thought this about Trump, but it no longer makes the same kind of mistakes as the early RNNs.

GPT-Chat still makes errors that are unintentionally funny, but they are fewer, and the days of state of the art AIs making humorous blunders are probably dwindling. The humor that comes from an AI producing nonsense might soon be obsolete.

I predict that the temptation to lie will be too great though. The general public’s understanding of AI and its capabilities and nuances is no better than it was when Botnik first emerged, and the humorous possibility of revealing reality, falsely, will be too much. If I swapped out the prompt and posted this to social media with an introduction indicating that this is what an all-knowing AI knows about Donald Trump, it would produce much more laughter than its current form. We will see people post ridiculous statements by AIs which were in fact written by the poster with no intervention from AI. This will continue either until the general public changes its understanding of AI, or the lying loses its charm.

—

There is an unexamined ethical question here. So what if jokes are lies if they make people laugh? If every standup comedian was lying the whole time, the combined laughter of the world would certainly outweigh any outrage at the revelation of deceit. It’s more fun when comedy is funny, and not finding things funny is more commonly a defect than a virtue.

The only response I have, and I don’t want to get into the underlying metaphysics because I haven’t figured them out, is that Truth is more important than enjoyment and humor. It is better to enjoy things than to not enjoy them, it is even better to know properly what is true and what is false, and it is best of all to enjoy things that are true. More than we want people to enjoy things, we should want them to know things. When people harmlessly laugh at jokes based on lies, there is a tragedy of ignorance and imperfection alongside the mirth. It would be best to laugh together at what is real.