Seeing other people’s AI art is like hearing other people’s dreams

If you’ve been on Twitter (or Discord or many other places) this month and your feed looks like mine, it’s currently full of weird machine-generated text and art. The launch of ChatGPT, Lensa, and other tools has made a once-hard-to-access tool ubiquitous, and the result is an explosion of people posting their interactions with AI.

I love these tools, and I’ve been playing with them for a while now — at least since the release of AI Dungeon 2 in 2019. But over the years, I’ve realized there’s a huge gap between how I experience the work I generate and the work someone else creates. The closest I can describe is this: Seeing someone else’s AI art is like hearing the plot of someone else’s dream. Sometimes it’s fascinating. Often it’s dull. And for me, that’s almost never more fun than going through the process.

Author Robin Sloan – who has experimented with AI fiction alongside mainstream novels – mentioned the dream simile to me in an interview on AI Dungeon Last year. “I don’t think people know very well how dreams work, but I think it’s probably not too far off from how the associative weirdness of an AI model plays out,” he said. -he declares. “And as everyone knows, you can have the craziest dream and tell someone about it the next day. And for some weird reason, there’s nothing more annoying than hearing someone’s dream. another.

It is now clear that the art of AI is often quite mediocre without careful instructions – the better it gets, the more proficiently it mimics bland SEO-bait filler text or a generic stock image. But as Rob Horning notesit became a joke format to show someone’s skill at tricking the AI.

It allows people to show off that creative consumption, how smart they can be to prompt a model. It’s like you can have a band just by listing your influences. You don’t even have to absorb these influences in practice; you can just name them. It is enough to become familiar with the signifiers.

The AI ​​fantasy goes all the way to the top, where companies like OpenAI demonstrate how art generators handle instructions like “an armchair shaped like an avocado” or “an illustration of a baby daikon radish in a tutu.” walking a dog”. There are real reasons to do this – it shows how the models combine several concepts and objects – but there is also a kind of 2000s era “lol so random“Quality to all of this.

And many prompts are not all that creative, as Horning notes.

Even this level of knowledge may be unnecessary. AI can take a prompt from scratch and suddenly make it interesting, giving anyone something fun and surprising to share.

This is where I disagree with Horning’s conclusion because as AI art generators have gotten better, the bar for me to find them fun or interesting has gotten progressively higher . Much of the current-gen AI is such a straight man in an improv comedy show, dutifully approaching any ridiculous prompts you offer with utter seriousness. (I actually watched a comedy troupe from Brooklyn put on an AI improv party where AI Dungeon generated key plot beats, but unfortunately the connection stuttered mid-show and killed the premise.)

Some people are good at this kind of collaboration. Janelle Shane, author of the fantastic You look like a thing and I love you, has an amazing blog of surreal “AI weirdness” that oscillates between plausibility and complete nonsense. Over the past week, I’ve legitimately laughed at ChatGPT writes bible verses about removing a sandwich from a VCR. I love that people find the cracks in ChatGPT’s SEO-optimized boilerplate and DALL-E’s stock artwork, forcing them to come up with something really weird.

But most of them aren’t so weird anymore – and the results often make me lose my eyes. I offered this take to my fellow AI expert James Vincent, and he said he hadn’t reached that saturation point… but he says he also likes to hear other people’s dreams, so I I’m not sure that disproves my point.

It’s a stark contrast to how much I enjoy watching AIs produce my instructions. It’s not that I think I’m particularly smart or funnier than most people. I love the back and forth of figuring out what a system like ChatGPT or DALL-E knows and guessing ways to force it out of its comfort zone. A fanfic crossover between stranger things and the heist comedy series Leverage: extremely predictable. A cross between Leverage and the 1980s experimental film Koyaanisqatsi: he is vaguely aware that the film has some kind of environmentalist connotation, but he also thinks that “Koyaanisqatsi” is the name of a kaiju. (It’s not.) I don’t expect you to find the resulting stories interesting, but I loved learning how to create them.

I don’t know how common this preference is. Sloan, for example, did not describe the posts of others; he was talking about communicating directly with a storytelling AI powered by OpenAI GPT. “I think people have different answers. Mine was not the wonder of a dream. It was much more like someone telling me their dream in real time,” he explained. Sloan was one of the few authors to work with AI writing before the explosion of OpenAI-powered tools, but by the time I spoke to him, he had mostly lost interest in the field.

That said, seeing everyone’s AI fiction and art is not bother me, although there are many open questions around AI copyright status, its potential for bias, and other serious concerns. Horning describes the process as submitting to an AI’s worldview, but to me it’s like watching people figure out the limitations of a video game. I just don’t like watching most people stream games either – and given how popular Twitch is, that might mean my feelings aren’t typical.

Ironically, the funniest AI art I saw last week was literally a description of someone’s dream. This is an image of my colleague Victoria Song, who was testing a sleep app that creates AI images based on your dream journals. This was the result:

It’s a hilarious image to me, and like a lot of humor, I can’t quite explain why. Perhaps that’s the context for putting this absolute nightmare fuel in an app designed to help you sleep better. This may be the model clearly associating tooth loss with Something creepy, then reverse-engineered a creepypasta image instead of a scene any human would describe as accurate. AI art is best when it’s wrong and right at the same time – but it’s a harder balance than it looks.