AIs have no sense of humour

I've been exploring the limitations of AI, as my first course on it comes to an end. (It sold really, really well, so it won't be the last…)

We know some things about AI with a great deal of certainty now. For example:

• LLMs aren't reliable for facts
• They're good at synthesis and predicting patterns from data
• They can be reliable when given constrained data sets to work with (RAG approaches)

But their limitations are slowly becoming more apparent. And the more I experiment with image generative models, the more I become convinced that one thing they're going to be bad at is humour.

ChatGPT can't cartoon

I generated the above image in ChatGPT during a training course, from a prompt suggested by one of the attendees. And, you know what? It's done a pretty good job of meeting the prompt. (I couldn't test it against Midjourney or Google Gemini - both refused to accept the prompt, because of its political nature.) But it's done nothing more than that.

There's no sense of nuance here, nor a sense that editorial cartoons are fundamentally humorous in nature, even when making a serious point. An editorial cartoonist would create something from this that would prick the pomposity of Farage's claim that he would ride to the rescue of the farmers, not just create a literal rendering of it.

What ChatGPT has done is produce something that looks, frankly, like a piece of propaganda for Farage and Reform.

Even with some serious work on iterating on the image and the prompt, I was struggling to get anything worthwhile. Case in point:

Can AIs perform multiple skills?

Fundamentally, the skill of an editorial cartoonist is to examine a situation, find some absurdity in it that can be captured visually, and then execute on it. (I ran this past a cartoonist friend of mine, and he raised no objections to this description.)

The nature of imagine generative models is to split that process in two:

• Come up with the idea for the cartoon
• Prompt the AI to generate the image

But, in reality, it usually ends up as a three stage process:

1. Come up with the idea for the cartoon
2. Work hard at writing a prompt that generates what you want
3. Finally generate the finished image when you (after many interactions) create the prompt that works - or you give up and accept something subpar.

At what point does it just become easier and more time efficient to pick up the phone and commission a real, human cartoonist?

There's an issue here around scale, speed and cheapness, versus quality, time and cost. Do you want throw-away visual dressing for something else, or do you want something that adds meaning to the product?

I suspect that cartooning in current models will be hard to replicate. Humour is based on exaggeration, unpredictability and twists. AIs are based on probability and averages. These seems like uneasy bedfellows.

With AIs, expect the expected

Honestly, I'm not quite sure where I'm going with this; but it feels like there's something important in this insight. Perhaps it's this:

AIs are very, very good at doing the expected. That's what they been trained for: to create something based on existing works. They're not good at doing something that's meaningfully unexpected. Sure, you can set parameters for them to go a bit wilder, to be a bit more unpredictable. But that's just creative spaghetti-throwing. You're hoping that it will unexpectedly form a compelling pattern on the wall…

We are living in a world where AI technology exists, and the technology won't be uninvented. So we have to live with it. And to do that successfully, we have to:

1. Understand what AIs are good at
2. Then understand what's left for us to be good at

And, I suspect, humour will be part of that.