Here's Why Your AI Has Bad Humor

Have you ever asked an AI model to tell you a joke and received something painfully unfunny? Something like:

Why do programmers prefer dark mode? Because light attracts bugs!

\ Groundbreaking. Truly.

\ More often than not, these so-called jokes generated by AI are weird and extremely predictable. The reason? It’s a machine! They do not have a human context— all it sees are patterns in the data they have been trained on, and when we ask them to tell a joke, they stitch together words based on probabilities rather than actual wit or comedic timing.

\ This is not rocket science for someone who works with AI, but it’s interesting to see what makes it the way it is. With that understanding, you might also figure out how to “ask“ AI models for better jokes.

AI “learns“ what a joke looks like

An AI model (say, GPT or a joke-telling bot) isn't born with a sense of humor. Instead, it learns from thousands of joke examples scraped from books, websites, and conversations. But instead of understanding humor like we do, it recognizes statistical patterns.

\ Here's an oversimplified look at how an AI model might process joke structures (completely different code, just for your understanding):

\

import random setup = ["Why did the chicken cross the road?", "What do you get when you mix AI with a comedian?", "Why don't programmers like nature?"] punchline = ["To get to the other side!", "A neural network with bad timing!", "Too many bugs."] print(random.choice(setup) + " " + random.choice(punchline)) # NOTE: This is oversimplified for easier understanding. # Real AI models don't just randomly select from pre-written jokes. # They use neural networks with billions of parameters to predict # the most likely next words based on patterns in their training data.

\ See the problem? It recognizes the pattern of a joke: first comes the set-up, and second comes the punchlines; it doesn't understand why a joke is funny.

\

\

AI predicts the next words, not the funniest ones

Outputs of AI models, whether humorous or not, solely operate on statistical probability. Large language models like GPT often work by predicting the most statistically likely next word(s). That's great for autocomplete and basic text generation, but terrible for original humor.

\ For example, if the AI generates the setup: “Why did the chicken cross the road?”

\ It predicts common responses for the punchline. Since "To get to the other side" is the most statistically probable answer, it will often generate that— even though it's the least funny.

from collections import Counter jokes_dataset = ["To get to the other side!", "Because it was programmed to!", "To escape the AI overlords!"] probabilities = Counter(jokes_dataset) # chooses the most frequent joke print(probabilities.most_common(1)[0][0])

That’s why they are really predictable and repetitive.

AI lacks context and world knowledge

Humor, unlike usual text generation, is deeply tied to emotions, human experiences, and cultures. Sad to say, they are the exact things AI fundamentally doesn’t grasp. Even with models trained on vast amounts of data, AI struggles when it comes to:

\

• Sarcasm: "Oh great, another AI-generated joke!"
• Wordplay & Puns: "An AI walks into a bar… and fails the Turing test."
• Social Awareness: "Why did the AI go to therapy? Because it couldn't process its feelings."

\ Even though AI can recognize common joke formats, it doesn't get why something is funny. As a consequence, they try too hard to be funny.

\

attempts = ["I'm hilarious, trust me!", "Why did the algorithm break up? It lost its spark!", "What's an AI's favorite drink? Java."] print(random.choice(attempts))

\ And then again, who to blame? These are machines trained on not-so-lawfully obtained data. The last thing AI engineers would want to happen is being sued for training AI models on data extracted from stand-up comedians.

Fixing AI’s humor

To get better jokes from an AI, you’ve got to be specific. Tell it the style you want (maybe like Louis CK’s dark humor or Chandler Bing’s sarcasm), set the scene, and let it know what kind of humor you’re after. It wouldn’t make it 50 times better, but at least it wouldn’t be so plain and stiff.

\ That’ll give AI a framework within which to work. Because right now, without context and an understanding of human emotions, all it can do is guess the punchline— usually in the most predictable, robotic way possible.

\ That’s just 700 words, no summary is needed. See you next time!