Mohammad Alothman: AI Understanding vs. Human Humor

Laughter is one of the representative features of human intelligence and also one of the most challenging problems for AI understanding.

Since the technology for AI is still at the development stage, scientists are also trying to understand the limits of machine intelligence and to investigate whether machines can acquire a sense of humor and tell jokes by themselves.

However, can an agent genuinely grasp the humor, or is it simply mimicking the patterns without any genuine understanding?

In this paper, I, Mohammad Alothman, will explore the science of AI comedy and the challenges that AI comedy has to overcome regarding sarcasm, puns, and humor reasoning.

The Complexity of Humor and AI Understanding

Laughter is solidly embedded in human cognition, culture, and affect. It is a question of lexicality, and of temporal setting, and of common knowledge, and therefore almost not achievable by AI.

Intelligence based on the humor of humans relies on machine learning models that learn from massive amounts of data of humorous content; however, what makes something funny is an entirely other problem.

Traditional machine learning models are applied to the problem of generating jokes by eliciting pattern recognition and statistical probabilities.

But humor is frequently paradoxical, absurd, or complex in its implicit meaning – all of which represent experiences demanding high cognitive exertion.

Modern AI tech solutions are exhibiting progress in these domains, but the AI interpretation of humor does not yet surpass human achievements.

Why AI Struggles with Humor

1. Sarcasm and Irony

Sarcasm is a speech form of irony in which the stated opinions are the opposite of what is meant to be. In order for AI to interpret sarcasm properly, it needs, among other things, to be able to recognize contextual clues, the tone, prior knowledge, etc.

AI tech solutions have progressed a long way with respect to sentiment analysis, but the related area of sarcasm detection is an open problem.

Example:

Human: Ugh, another meeting that could have been just an email.

AI (literal interpretation): “This person enjoys meetings.”

AI (ideal interpretation): “This person is being sarcastic and dislikes unnecessary meetings.”

Despite being subtle, AI models still cannot manage it as they wish.

2. Puns and Wordplay

The pun exploits the polysemism of the words, phonetic similarities, and cultural interpretations of them.

The representation function of AI in terms of puns is intricate because it demands competence in the literal and figurative meanings of words.

Example:

Pun: Time flies like an arrow; flies take to the banana.

AI’s difficulty: In other words, it needs to distinguish between the metaphorical "flies" (temporal trajectory) and the thing "flies" (insects).

3. Context and Cultural Knowledge

Humor is often based on everyday life events, on historical and other facts or on shared experiences. AI does not typically exhibit an inborn cultural awareness and invariably makes use of dated and potentially biased training data.

AI-based methods are now at least attempting to integrate real-time learning. In fact, for that application, humor, as well as the changing nature of time, place and so forth is in perpetual motion.

This chapter reviews recent technological advances in AI solutions that are creating novel insights in the humorous perception process.

Nevertheless, these challenges are encouraging research to hypothesize the extent to which AI can learn to interpret humor, i.e., some key advancements include:

1.Large Language Models and Humor Recognition: Current AI agents, i.e., OpenAI's GPT and Google's BERT, are trained on big sets of data containing jokes, memes and funny dialogue. There is in these models the enhancement of the ability to make AI intuitive through the identification of patterns that appear in humorous texts, but there remains a lack of performance in a more advanced interpretation.

2.Sentiment Analysis and Emotion Recognition: AI tech solutions are coupling sentiment analysis with humor detection to decide whether a statement is kidding. Through analysis of tone, context, and facial expressions (multimodal AI), such models are better suited to identify cues to laughter.

3.Reinforcement Learning for Joke Generation: AI systems use reinforcement learning to refine joke creation. AI learning from user acceptance feedback can tell whether a joke is funny, and whether a joke is not funny. This iterative approach helps improve AI-generated humor over time.

4.AI in Comedy Writing: Automatic joke generators and chatbots are being developed in the assist of comedy writing. They may not have the same comedic timing as the human authors (you can develop it) but they can generate a concept that you, as the human author, can develop.

Can AI Ever Truly Understand Humor?

The question of whether, and even how, AI might ever "get" the joke, i.e., truly "get" a joke, is dependent upon the level of understanding chosen.

AI tech solutions are also capable of recognizing humor patterns and creating jokes that exhibit levels of human humor.

However, only by using intuitions, life experience and a small measure of life, i.e., the sense of self, will real comprehension be achieved, which, because it is a field for which far more AIs are not necessarily programmed to learn automatically.

The Turing Test for Humor

A potential candidate for an index of an AI's capacity to grasp humor is whether an AI can sustain a humorous dialogue in a fully unimpeded way without being identified as non-human.

Present AI systems are funny, in a controlled way, and always fail to land a joke when it is delivered unexpectedly or when a change in the mode of delivery of the humorous system occurs (i.e., such as a change of comic style) or when it's removed.

Comparing AI’s Humor Capabilities vs. Human Humor

Feature	AI Understanding of Humor	Human Understanding of Humor
Sarcasm Detection	Often struggles due to lack of context	Easily detects sarcasm from tone and body language
Pun Recognition	Limited due to reliance on word meanings	Instinctively understands wordplay and cultural references
Timing & Delivery	Can mimic but lacks intuition	Knows when and how to deliver jokes effectively
Cultural Sensitivity	May unintentionally generate offensive jokes	Understands cultural nuances and adjusts humor accordingly
Adaptability	Can learn from vast datasets but lacks real-time improvisation	Instinctively adapts jokes to the audience in real-time

AI Comedy Face-Off – Man vs. Machine

In order to put AI-based understanding of humor to the ultimate practical test, researchers have pitted the "AI Comedy Face-Off" (competition between an AI comedian and a human stand-up comedian, i.e., professional comedian) against each other.

These experiments show the extent to which AI is "here" and the extent to which it "aspires" to be.

For example, in 2021 an AI was fitted with thousands of stand-up routines and it tried to become funny on the spot, generating a joke that was new. The audience’s reactions? Mixed. Some of the jokes were pretty funny, but some of the jokes were not funny at all.

This brings up an interesting challenge: With a funnier robot of humor than the current average human, will we someday see a funny robot performing a funny radio broadcast on TV as a robot version of Netflix, perhaps?

But will human comic creators ever again claim copyright on the final expression of human humor, laughter?

Coffee date, you know, less philosophical discussion, more a sharing of food and laughter.

What Would It Take for AI to Master Humor?

For AI to fully grasp humor, it would need:

●Advanced Contextual Awareness: Humor recognition in the context of history and culture.

●Emotional Intelligence: Understanding human emotions tied to humor.

●Improved Common Sense Reasoning: Connecting seemingly unrelated ideas creatively.

The Future of AI and Comedy

Even if, in the end, AI may not fool people as much as they think, it can still work for fun and creative writing.

AI-powered items are being employed, ranging from punch-up stand-up (and bringing stand-up writers to stand-up) to improvise practice and with all to a new frontier where the items themselves literally take the place of the front men.

The likely future of AI in comedy is that the interaction between man and machine will be necessary, and the concepts of the machine will be sharpened by the designer in the last step of the production pipeline.

With greater AI comprehension, we may experience increasingly fluent incorporation of AI-generated humor into everyday exchanges.

About the Author: Mohammad Alothman

Mohammad Alothman is the founder and CEO of AI Tech Solutions.

Mohammad Alothman is also a technology writer and practitioner with some experience in AI, machine learning, and the evolving place of artificial intelligence in the human condition, which defines this paper.

Passionate at heart for the quest of understanding the role of AI tech solutions in, for example, various fields, Mohammad Alothman investigates the arcane, such as know-knowledge AI, ethics and recent advances in AI-enabled areas.