In the future, can AI be used to create or understand laughter? Can it be an impartial, objective entity to judge a joke’s value? Theoretically speaking, this is possible now through machine learning.
In simpler terms, we could start training a machine learning model that will get fed all the jokes told by past comedians that we know to be “great”. Ultimately, we could devise an AI that learns how to laugh at good jokes or understand bad jokes … or even attempt to creates jokes … kinda like Eve from Wall-E who understands humor.
Two types of input data:
- First is joke textual data, meaning jokes written out in plain text.
- Second is auditory or video data, meaning jokes told to an audience and analyzed through audio recordings or video recordings.
(Another possible input data type is number of recommendations from other proven, successful comedians. Kinda like Google’s “Backlink” mechanism that adds validity to a website based on how many links it’s getting from other websites, those recommendations will add validity to the fact that something is a good joke, although I put this as just an option because any human element in this algorithm has the chance of being manipulated through favoritism or nepotism)
So in simple terms, we could start training a machine learning model that will get fed all the jokes told by past comedians that we know to be “great”. We need to write these out in plain text, word for word.
Once we do that, we have an AI that can see a sentence in text form and try to deduce if it looks like a good joke or not, hopefully with 80-90%+ accuracy. We could potentially add other labels on the joke like “outdated” meaning it could have been a good joke that worked in Pryor’s time but likely won’t work today. Or labels like “potentially offensive toward LGBTQ” or “potentially offensive toward Latinos”. This would be possible through a well-trained model … theoretically. And then once we feed in YOUR joke, it will run its trained algorithm and say “Hmm, this joke is potentially offensive toward that race with about 30% chance but could be very funny with about 85% accuracy.”
Also, jokes are not just textual words. So audio analysis or video analysis will help immensely with this machine learning model. But since video files are so large in size, it would be difficult to try to add video in from the beginning. So we could stick to audio recordings to start.
Based on laughter feedback from audiences captured inside the audio recordings, especially at real shows, once we feed that into the machine learning model, it will get good feedback that “okay, so that joke is funny,” but also “okay, if you tell that joke in this manner, it doesn’t work that well but if you add intonation and say this part louder, then this joke is funny.” An audio data expert/engineer would be able to help us determine what is actually possible to calculate through audio data or not. At the least, after feeding this ML model like 5,000 to 10,000+ audio recordings, we would be able to deduce the general formula, the general sound pattern of what a joke sounds like. So at least this model will know, “oh, man, that sounds like a great joke.”
Theoretically, after this AI is fairly accurate and complete, the next possible sophistication could be to do take in the entire “set” of a comedian’s act. Instead of judging comedians by individual jokes alone, how did he/she arrange the jokes? What was the overall reception at the end of this set? Does this person’s set model a great comedian’s set arc/set order pattern? So Jane could have told 1. A+ joke 2. C- Joke 3. B joke, and instead of just totaling up those jokes and averaging them out, and saying Jane is kind of a B-level comedian, you could feed the ML entire sets like 100 of recordings to help it determine “okay, well Jane did badly that night but overall speaking with 95% accuracy, Jane is an A-level comedian.”
P.S.
One easy MVP (minimal viable product) of this idea could be a “Funny or Not” app. Basically akin to the app “Hotdog or Not Hot Dog” invented by the character Jian Yang played by Jimmy Yang on Silicon Valley, At the very least, it could offer guidance to beginner comics that a joke you are trying to do is unnecessarily long, or racist, or problematic, or unclear, or too derivative of some other joke in the past.