Can’t think of a Halloween costume? Let artificial intelligence do it for you.


Forget Sexy Doctor, Batman or whatever other generic costumes your friends are wearing this Halloween. Why not try something new, like Skeleton Ginsburg, Space Pants or “Magic an of the Foand Tood-computer”?

No, that last one was not a typo.

If you choose one of those less-conventional costumes, you’ll have Janelle Shane to thank. She’s a research scientist who programmed a neural network to generate Halloween costume ideas.

Shane, who has also programmed neural networks to create hilarious names for metal bands and craft beers, started by crowdsourcing more than 4,500 examples of Halloween costumes from her readers. These functioned as the neural network’s building blocks, the ingredients it mixed and matched to generate its own ideas, like Spartan Gandalf, Dancing Bellyfish, the Game of Nightmare Lightbare and Petty Fairy.

Janelle Shane/aiweirdness

“The most popular submitted costumes are the classics — 42 witches, 32 ghosts, 30 pirates, 22 Batmans, 21 cats (30 including sexy cats), 19 vampires and 17 each of pumpkins and sexy nurses,” Shane’s blog post reads.

Shane then dumped these submissions into a neural network that “learns words from scratch, letter by letter, with no knowledge of their meaning,” a choice Shane said made things a bit “weirder” than if she had picked one that actually understood the meanings of words.

And “weird” is certainly no understatement. Here’s the first batch of costumes her network spit out:

Janelle Shane/aiweirdness

It was a start, but clearly it needed to get a bit smarter.

“By the time I checked on the neural network again, it was not only better, but astoundingly good,” Shane wrote. “I hadn’t expected this. But the neural network had found its niche: costume mashups. These are actually comprehensible, if a bit hard to explain.”

Examples from that batch include Martian Devil, Panda Clam and the Shark Knight, all of which would be guaranteed hits at the office Halloween party.

Shane said the neural network also generated hundreds of “sexy” costumes — not unlike the actual Halloween costume market — but “it never quite got the hang of it.” That is, unless, the idea of dressing up as Sexy Conchpaper or Sexy Gumb Man doesn’t sound particularly appealing.

Then again, if this all just seems too overwhelming, you could always decide dressing up just isn’t for you — but then who else would get the chance to represent Lady Garbage or Sexy Gargles?

What a shame it would be to let those ideas go to waste.