ChatGPT Wins Again...but what about Him?

Him.

Hayao Miyazaki, the 84-year old founder of Studio Ghibli, loves to draw with his hands.

His creative inspiration comes from human experiences — suffering, happiness — and perhaps it is this eccentric rootedness in emotions that makes his creations so relatable and comforting for people worldwide.

"I am utterly disgusted," he proclaimed when shown an early render of an AI-generated animation of a writhing humanoid zombie in 2016.

"I would never wish to incorporate this technology into my work at all."

Nearly a decade later, a trend of generating art that mimics Ghibli animations is taking over the internet — thanks to ChatGPT launching native image support with 4o model.

Now I am not quite AP News and an optimistic backer of AI. But I wanted to start with this criticism even when I don’t agree with Miyazaki’s words, because it highlights one valid concern — that of copyrights and compensating artists whose work is used to train AI.

Some are thinking that the Ghibli fad started spontaneously but that couldn’t be farther from the truth.

When Sam Altman and OpenAI team went live to announce the new image support in ChatGPT, they specifically displayed a Ghibli-style edited image of themselves, even though they didn’t use the name of the studio or artists in the prompt displayed.

Sam also confirmed this in a tweet later.

Why Ghibli?

Now, did you also wonder like me — why Ghibli? Why aren’t that many people interested in reimagining themselves as a Disney princess or a Marvel superhero?

Because — copyright.

You see, Studio Ghibli comes from Japan — a country very unique in its “progressive” laws that AI companies can train their data on copyrighted material for commercial purposes, without bothering about the involved artist’s consent.

Even as local anime artists in Japan protest, these laws mean they are a relatively hassle-free source for machine learning models to be trained on.

Disney and Marvel are protected from the same fate by much stronger copyright laws and deeper pockets (while also being based in a notoriously litigious country).

OpenAI and other LLM foundational model companies have signed big deals with U.S. corporations to make their data available for training.

It is unclear if OpenAI would have struck such a deal with Ghibli.

Based on ChatGPT refusing to generate images when directly asked to generate “Ghibli” images, combined with Miyazaki’s well-known views, it seems highly unlikely.

This wouldn’t be the first time for OpenAI to go through such a niggle. In fact, do you remember Her?

ChatGPT is alleged to have trained one of the voices on ChatGPT's audio conversational model on Scarlett Johansson's voice — despite her refusing to consent.

The “Black Widow” star has taken OpenAI to court over this.

I think the copyright battle is going to be a long-drawn one but I am resolute in my view that individual artists or organizations whose work is being used to be recreated by AI should be fairly compensated.

This matter is very different from the wider training on content available on the internet, where you could argue that AI is building the “memory” similar to humans rather than simply reproducing the work that counts as plagiarism.

Anyway, signing this one out to Jurisprudence.

The mind-boggling advancement in AI Images

On the optimistic side, the quality of the new natively-generated images by OpenAI’s 4o series mark an astonishing improvement over the previously-used Dall-E models in ChatGPT.

This is especially true on the front of graphic designing and image editing versus simple image generation.

Midjourney back in 2022 was the first model that impressed me enough and we used it to even generate feature images for articles at our publication at times.

Stable Diffusion made the images more realistic last year and Flux (I believe used in Grok) made images way better.

Now, Google DeepMind and OpenAI bringing native image generation to models is the next big thing.

While some critics were busy dunking on AI’s inability to generate hands and text successfully in images, the actual machine learning engineers were focused on building and shipping.

The results are here and will only get better.

Best from Around The Web

This guy made the entire trailer of the Lord of the Rings: Fellowship of the Ring partly using GPT-4o!