Will AI Take Creative Jobs? Judging by These Paint Names, Probably Not Today

Coloring in shades of 'Turdly'

Colors can affect our feelings and behavior in ways so profound that studies have been conducted on how we can better use them in life, and in society, to “hack” culture. This covers subjects from what colors to wear to a job interview … to the ideal shade for prisons.

Because of this odd relationship we have with color, the colors we choose, and what we name them, become meaningful shorthands for much bigger stories. Every year, Pantone selects a “Color of the Year” that thematically puts us on the right track for the next 12 months. We project as much onto Pantone’s choices as the brand seeks to project onto us. (This year’s color was “Greenery.”)

With all this in mind, research scientist and neural network geek Janelle Shane decided to see how well artificial intelligence fares at both selecting colors and naming them. A writer at Ars Technica calls her results “the greatest work of artificial intelligence I’ve seen to date.”

On Tumblr, Shane describes both the terms of the experiment and its output.

“I gave the neural network a list of about 7,700 Sherwin-Williams paint colors along with their RGB values (RGB = red, green, and blue color values),” she writes. “Could the neural network learn to invent new paint colors and give them attractive names?”

Shane selected a neural network algorithm called char-rnn, which predicts the next character in a sequence. The algorithm’s task was to come up with sequences of letters to develop color names, then sequences of numbers that tie those names to an RGB value.

At the first checkpoint, she observed the algorithm could create colors before it could properly name them.

“These are colors, all right, and you could technically paint your walls with them,” Shane allows. “It’s a little farther behind the curve on the names, although it does seem to be attempting a combination of the colors brown, blue, and gray.”

By her second check, the neural network could properly spell “green” and “gray.”

“It doesn’t seem to actually know what color they are, however,” she admits.

A more “creative” setting yielded word results that were even more surreal. We do think “Dondarf” shows promise, though we ourselves might have gone with a name more along the lines of “Bridesmaid’s Lilac.” (Just because we’ve worn it in that capacity too many times to count.)

The longer the dataset was processed, the closer the neural network got to developing meaningful color names … though the results are both surprising and laughable. Consider “Gray Pubic,” the name it so gamely gives an otherwise lovely turquoise hue.

Before we come down to the algo’s final results, here are a few Sherwin Williams color names from its “Energetic Brights” family, for reference:

And here’s the Shane algo’s final results. A few of our favorites include “Snowbonk” (as good a description as any for a dirty beige), “Dorkwood,” “Stoner Blue,” “Stanky Bean,” “Turdly” and “Dope.” (Simple and direct. Far from dope, though.)

Shane writes:

In fact, looking at the neural network’s output as a whole, it is evident that:

  1. The neural network really likes brown, beige, and grey.
  2. The neural network has really really bad ideas for paint names.

Whole books have been written on how colors (and names) affect us, including Drunk Tank Pink, named for the shade that supposedly soothes both aggressive prisoners and rowdy schoolchildren. What we find here, though, is that both the ability to chart shades in agreeable families, and give them names that make us dream, are fairly critical to making colors “pop.”

The bottom line is, brains vary from person to person, but our inputs come from multiple senses that can be combined to form whole constellations of values. The toughest thing for an algorithm to understand may well be that words and colors mean different things to different people at different times, based on a diversity of references that are as much personal as universal.