Bronze sculpture of melancholy Laputian robot at Ghibli museum

22 rules of generative AI, 2 years on, Ghibli intermission

This update comes at the peak of the Ghiblification fad. There’s nothing new here and yet I found the release of and response to GPT 4o image generation, including mechanistically crushed and artificially reconstituted Ghibli, particularly shocking.

So here’s a retrospective case study for the recent review of rule #10 (labelling ingredients) and a longer intro to the upcoming review of rule #21 (supply chain governance)!

There’s nothing new in the fundamentals:

  • Image-to-image style transfer has been around for years (to the extent I thought the novelty had worn off),
  • It’s long been established that GenAI producers have opaque and contentious data accumulation practices*,
  • And that tech companies have a tendency to ask respond to questions later when appropriating items of cultural significance

This release also has ingredients consistent with priors from OpenAI:

  • Leveraging a recognisable cultural reference point without engaging original creators,
  • Taking another incremental step (which is totally valid) rather than a giant leap (that would better justify the hype) to the moving targets of GPT 5 or AGI, and
  • An attendant hypercumulus cloud of AI influencer noise, including regrettable takes like this.

And yet something about this remains particularly disquieting for me. Perhaps it is that:

  • Miyazaki himself is on the record for criticising the use of AI generated animation and drawing, saying it knows nothing of human experience, and musing we’re losing confidence in ourselves in the face of AI capability. Being based on strong values, I don’t believe his position would have changed.
  • A fundamental theme throughout Studio Ghibli’s work is celebrating and preserving the uniquely human and natural world in the face of technological encroachment and destruction. AI slop is informational pollution to rival the environmental pollutants in Ghibli’s worlds and, beyond the inappropriate application, I can only imagine Miyazaki’s despair at the current and potential future environmental costs of generative AI.
  • While generic anime style, and mimicry of many artistic and writing styles has been possible for years, the degree of precision in the replication of Ghibli’s style (and others) at scale feels viscerally wrong, at a minimum inviting fresh legal challenges based on copyright or moral rights, or though extension aptly illustrating the semantic apocalypse, eventually draining all content of meaning.
  • Or maybe it’s just a vulgar display of power?

I’ll come back soon to a more prosaic assessment of how my back-of-a-napkin GenAI rules from February 2023 have stood up over the last 2 years.

* Including my copyrighted work – search Libgen for the authors of Effective Machine Learning Teams or the C4 dataset for safetydave.net.


Posted

in

, , ,

by