LAB31

Try this GPT (requires ChatGPT+ subscription)

Everyone has their favorite holiday. One of ours is Public Domain day, when all kinds of fresh goodies are freed from the shackles of IP protection and open for gleeful experimentation.

We’re also super interested in how AI has raised big questions about what counts as “creativity.” Is it a matter of inventing something wholly new? Or is creativity really an exercise in lateral thinking, banging divergent things into something greater than the sum of its parts?

Not sure this GPT will answer any of those big questions, but it sure is fun to play with.

Hit “do your thing” and it will randomly select two public domain works, whether film, art, photography, or literature. It analyzes each piece then uses the outcome of that analyze to remix the two works into something as unique as GPT is capable of.

Two big challenges we had to work through with this one:

1) Genius may steal, but outright theft is boring. It took a lot of fine tuning to get the model to not just mash the subject from one piece into the visual style of the other. And even with a ton of work, it’s not always perfect. This is especially apparent with visuals that are over-represented in the model’s training data.

See what I mean?

We really had to dig deep and use every tool in our prompting kit to get the model to output something that felt like it was more than a straight mashup. And tbh, we would happily spend three more days refining the instructions even more. But alas, that is not the nature of this experiment.

2) OpenAI’s content policy is a harsh mistress. Even though these are public domain works, the guardrails on image generation are so strict that even a whiff of a hint of something can trigger the content filter. It took a ton of work to get GPT to generate a Dall-E prompt that didn’t say “take the X from VERY SPECIFIC REFERENCE TO WORK A and mix it with the VERY SPECIFIC ELEMENT FROM WORK B.” We basically had to teach it how an artist transmutes inspiration in the creative process.

But the content filter goes beyond copyright to get REAL annoying sometimes. Here’s an example of what I mean: “This prompt did not align with the content policy, which may be due to the inclusion of the “Nosferatu” character, a figure that might be considered too dark or frightening for image generation.”

This points to another of the fundamental issues with AI alignment. The more you work to keep people from doing bad stuff, the more you lock down the system in general. It’s a tricky balance and nobody’s really figured it out.

Side note: Our original name for this GPT was “Barely Legal Art Lab” and, well, OpenAI did not like that. Get your minds out of the gutter, guys. we’re talking about copyright law here!