Jeff Kubina/Rog01/Daniela Hernandez

Last week, a bot learned to write short descriptions of images based on Taylor Swift lyrics its master had fed it. These were then set to music to produce a short Swiftian music video. The results weren't anything that passed for a chart-topping hit, but it raises an interesting question: Where is the line between robo-fan-art and plagiarism? Could the real Taylor sue for copyright infringement and financial damages?

It's a question that's likely to come up more frequently as we enter a new era of computational creativity. Art has historically been an entirely human enterprise, and so the law is set up to protect artists from human copycats, not robotic ones. But that is beginning to change as machines get smarter and start to generate what seem like autonomous works that, in some cases, look very much like what a human might produce.


"There is this giant new territory that's opening up in machine learning and computer creativity," Samim Winiger, a German AI artist, told me.

The Swift bot is just one example. There's also bots that tweet Bible-like verses, deliver their own TED talks, produce new Irish folk music, dream up magic card games, and write short stories based on photographs they're shown. In July, I taught a bot to write its own erotica.


Some artificially intelligent programs take images and generate new ones. The most famous among these is Google's hallucinogenic DeepDream and Facebook's EyeScream. In October, a group of researchers published a paper describing an AI that was able to "study" images and use its understanding of them to make its own illustrations. Some look like near replicas:

Lu et al., 2015

Others are more "original."

Lu et al., 2015

The question of copyright, Winiger says, is a hot topic of discussion among artists. He recently spent some time chatting with other machine-learning experts on what the future of copyright law might be and what it will mean for their art. The general consensus, he told me over Skype, is "😕."


Many of the projects I mention use data from public databases. When I trained my own Erotibot, I didn't want it (or me) to get in trouble for infringing on someone else's copyright. Zahr Said, an associate professor at the University of Washington School of Law who specializes in copyright law, told me my safest bet was to ask authors directly for permission or to stick to works released under a creative commons license, meaning I could use them free and clear. That's what I did.

Said mentioned my project, though, could fall under "fair use," as long as it differed enough from the original. As a judge wrote in a fair-use case, "the use must be productive and must employ the quoted matter in a different manner or for a different purpose from the original. . . . If . . . the secondary use adds value to the original – if the quoted matter is used as raw material, transformed in the creation of new information, new aesthetics, new insights and understandings – this is the very type of activity that the fair use doctrine intends to protect for the enrichment of society." That's why fanfic qualifies as fair use under copyright law. Plus, in the fanfic universe, companies have mostly chosen to turn a blind eye because it serves as free advertising.


Would my bot or other bots like it satisfy those criteria and enjoy the same protections? James Grimmelman, a professor of law at the University of Maryland who specializes in online copyright law, says yes.

Bots are "extensions of human creativity. They are like paintbrushes or an Etch-a-Sketch. They are tools people can use to create in interesting, new ways," he said.

As such, they "add something to the original work in a meaningful way," he added. "Most of us would agree that bots that generate poetry or erotica or DeepDream images based on a source would qualify as transformative." The technical details are just a "distraction from what's really an easy copyright issue."


Bot works fall clearly under fair use, even when the author of an original work doesn't give explicit permission for their work to be bot-transformed.

For one of his most recent AI experiments, Winiger used math to stylize a video released by YouTube celebrity JacksGap about a trip to New Orleans. Winiger used open-source programs that predict what color a given position in a frame should be, and then colors it in. It's been done with still images in the past, but Winiger wanted to apply it to video.

He told me he tried to reach JacksGap, but he was unable to. (Winiger does credit him at the bottom of the blogpost that describes his experiment.) The original and Winiger's version, which looks like a moving abstract cubist painting, are completely different. But without JacksGap, Winiger's wouldn't exist.

Winiger and others are mostly doing these experiments for kicks, but what happens if they start making money off of them?


Some experts think that setting up a system of micropayments for authors whose works a machine uses to learn to write, draw, or compose music would be the most fair way to go. The work of art could then be collectively profitable for all parties involved. But Grimmelmann says that will turn into "an administrative nightmare," requiring artists and engineers to figure out exactly how much an algorithmic work drew from each human artist. The simplest solution, he says, is to look at bot-generated art as fair use across the board, especially since humans are still at the helm.

Right now, the output of the machine-artists is still curated by humans. That's because, especially for text, the results are usually laughable. A human selected the released images we saw come out of Google's DeepDream. Of the words generated by the Taylor Swift bot and the robo-Ted talk writers Winiger coded, he's handpicked what he considered the best ones. I did the same with Erotibot.


But AIs will get smarter and one day, they'll be able to write their own stuff, based on the billions of pages of text they've scanned.

For instance, Winiger says, let's fast forward to the year 2020. "I drop a 'Disney bot' on some darknet server: it generates nonstop old and new Disney movies," he said. "Now, these are not simple copies, but generated [ones], including new fictional characters like Morphy Mouse."

These fully bot-generated works would still be protected under fair use, Grimmelman says. Even if some of its "creations" were pure copies, the bot would only be accused of copyright infringement if human eyes ever laid eyes on them. That's the whole premise upon which search engines work. Copies of text and images are created and indexed, but because it's for the purpose of helping you find information (and only software has access to these copies), it doesn't infringe on others' copyrights. When it comes to copyright, robot consumers don't matter.


Daniela Hernandez is a senior writer at Fusion. She likes science, robots, pugs, and coffee.