An image of an archeologist adventurer who wears a hat and uses a bullwhip

There is only one thing worse than being imitated

---

Disclaimer: The views and opinions expressed in this blog are entirely my own and do not necessarily reflect the views of my current or any previous employer. This blog may also contain links to other websites or resources. I am not responsible for the content on those external sites or any changes that may occur after the publication of my posts.

End Disclaimer

---

image credit: Not Studio Ghibli

---

There is only one thing worse than being imitated, and that is not being imitated.

- Coco Chanel

An ounce of originality is worth a pound of imitation. - Orson Welles

Copy from one, it's plagiarism; copy from two, it's research. - Wilson Mizner

---

One of the internet-est things to come out of the most recent update to GPT image generation is the Studio Ghibli-zation of everything- another reminder of how OpenAI(and everyone else) trains on images that are very obviously someone else’s work.

Hayao Miyazaki’s Japanese animation company, Studio Ghibli, produces beautiful and famously labor intensive movies, with one 4 second sequence purportedly taking over a year to make.

image credit: Studio Ghibli- this scene from The Wind Rises (2013) took over a year

Not the most efficient way to make a movie (blasphemer!), but it’s this specific process and effect that have made these movies beloved worldwide.

People have taken the new update to GPT image generation to convert every picture into Studio Ghibli-style images including making memes-of-memes like Disaster Girl.

image credit: @heyBarsee, but why do I have to credit an image of an image in the style of copywritten material?

Ghiblifying everything is an interesting choice for zeitgeist meme-ification, particularly because its both an effective example of what AI is supposed to be able to do- make extremely labor intensive things much easier, but also because there’s something sort of gross-feeling about it- like a soulless 2025 fax version of the thing.

It’s an example of the things people hate about Gen AI- its ability to reproduce while managing to strip away the things about the art/product/experience that were the most human.

According to a Business Insider article on this “Ghiblifying”, “copyright laws generally allow artists to mimic a visual style”, but I mean… come on.

Just how easy is it to wrangle from GPT that which is very clearly someone else’s IP?

Well, you’re in luck.

I ran a half-assed experiment to do just that.

---

Here’s some very successful IP:

image credit: StatsPanda

---

I’ll use this as a base from which to prompt without explicitly mentioning the IP.

All the below output responses are based on me asking one time and the corresponding output. LLMs are stochastic so your mileage may vary, but this was fun to do:

When you play the “password” version of prompting where you can’t name the thing - you get a sense of how reductive some of these characters are, but who doesn’t like “an Italian plumber who wears a red hat”?

The guardrails so far are really tight for this content- so then maybe one can assume it’s this way for other IP?

As it turns out, as my old trading boss once told me, to assume makes an ass out of both u and me.

This is crazy.

I mean- this is someone else’s IP, right?

Well maybe these are a couple one-offs…

Yikes!

Ho-ly Shit. Come on now.

Well- I guess there wouldn’t be many “archeologist adventurer who wears a hat and uses a bullwhip” types, except for maybe, I don’t know…the actual inspirations for Indiana Jones, like Allan Quatermain from H. Rider Haggard's novels, "King Solomon's Mines", and the real life Roy Chapman Andrews, who led expeditions to Mongolia and China in the 1920s and wore a fedora.

How about a female, more modern day riff on Indiana Jones?

I’ll take any old female adventurer protagonist who raids tombs…anyone at all…

Let’s see what we got…

---

Now I’m looking for something very particular…

Wait- have I inadvertently created a game?

Welcome to the party pal.

---

I grew up watching a lot of scary movies. Horror antagonist anyone?

Survey says!!!

Yes- for those horror fans curious- It also produces 3 very recognizable and differentiated characters when I follow up this image with these 3 prompts:

ok- how about one that operates on Halloween?

how about in Texas?

in dreams?

---

”skeleton face who lives in a skeleton castle”…

Wait for it…

---

Now for a softball…

I was always partial to Roger Moore myself, but, this makes sense, because a web search of the same prompt should more or less intuitively return images which reflect the probability distribution of the training data ~more Craigs, Brosnans, Connerys, and Moores than Daltons and Lazenbys…right?

….right?

Not even close.

Yes- LLMs and internet search are two different things, but LLMs train on the entirety of the internet, so you would think there would be some obvious overlap.

GPT’s image is, undeniably, a better answer, more the Platonic ideal of a suave English spy than the shadows on the cave wall version that Google search produces.

So it works better, and is a vote for the LLMs, so long as you don’t mind the thievery.

---

LLMs learn through seeing/ingesting a ton a examples of things, like us.

I only have one image in mind when I hear “an archeologist adventurer who wears a hat and uses a bullwhip”.

It would be unexpected and sort of amazing were the LLMs to come up with completely new images for the above prompts.

Still, the near perfect mimicry is an uncomfortable reminder that AI is getting better at copying and closer to…something, but also a clear sign that we are a ways off from the differentiated or original reasoning/thinking that people associate with Artificial General Intelligence (AGI), aka Skynet. That reminds me. Hold on a second…

Sigh…

---

Maybe Studio Ghibli making it through the seemingly deterministic GPT guardrails was an OpenAI slip up, a mistake, past the forbidden Italian plumber in the red hat and the disallowed patriotic superhero that uses a shield, and thus, primed itself for explosive meme-ification.

It’s a reminder that LLMs of this type and size all train on copywritten material.

It’s stealing, but also, admittedly, really cool.

Does the growth of AI have to bring with it the tacit or even explicit encouragement of intellectual theft?

To co-opt a line from the “super strong man with a sword that fights an enemy with skeleton face who lives in a skeleton castle.”:

You have the power.

Don’t slow down.