General Intelligence
Deepfake Music Is So Good It Might Be Illegal
And other fascinating A.I. research from this week
Welcome to General Intelligence, OneZero’s weekly dive into the A.I. news and research that matters.
A.I.-generated music is finally good.
This week, a programmer mimicked Jay-Z’s voice and cadence so effectively that his agency, Roc Nation, got SoundCloud to take the track down. Then, yesterday, OpenAI released Jukebox, an algorithmic system able to generate music — complete with lyrics — in the style of famous musicians like Elvis and Frank Sinatra.
Essentially, these algorithms analyze large collections of an artist’s songs, find patterns in the audio data that humans would correlate with hallmarks of music style, and then use those patterns to generate new audio.
Some of these generated songs are incredibly good. One sample, found embedded on OpenAI’s website, is a Christmas song about being in a hot tub, sung in the style of Frank Sinatra.
Here’s a sampling of the lyrics:
It’s Christmas time, and you know what that means,
Ohh, it’s hot tub time!
As I light the tree, this year we’ll be in a tub,
Ohh, it’s hot tub time!
Looks like there’s a new Sultan of Swoon.
But these algorithmic reimaginings of artists’ styles raise complex legal questions. Amanda Levendowski, a professor at Georgetown Law and founding director of its Intellectual Property and Information Policy Clinic, says that while, at first look, the Jay-Z situation may look like a copyright issue regarding the data used to create the voice, it’s actually closer in spirit to prior legal cases of human impersonators.
Levendowski pointed to one case from 1988, when the Ford Motor Company made an advertisement that mimicked the voice of singer and actress Bette Midler, using a former backup singer to create the “realfake.” Midler sued, and the Ninth Circuit Court of Appeals found that Midler’s voice was protected from recreation as a property right. In his decision, Judge John T. Noonan wrote that the First Amendment protects the recreation of a person’s voice, but if that recreation does nothing but exploits a person’s identity, it’s not protected under the First Amendment.
There’s another example: In 1991, Samsung dressed up a robot like Vanna White, again for an advertisement. In response, White sued the company under the same grounds in California, referred to as “right of publicity.” She won the case.
Right of publicity laws aren’t federal; they’re enforced state by state. For instance, Tennessee, as the home of Elvis, Dollywood, and a thriving music capital, has very favorable right of publicity laws for artists.
“You can really see why it might be tempting to go deep on some of these copyright questions. The right of publicity is actually way more wild and strange,” Levendowski says. “Copyright has been administered federally for a very long time… Some states just disagree foundationally on why we have right of publicity.”
So, is A.I.-generated music that impersonates a human singer legal in the United States? The answer is a satisfying and emphatic maybe.
Here are some other interesting pieces of A.I. research from the week:
Addressing Artificial Intelligence Bias in Retinal Disease Diagnostics
New research from Johns Hopkins University and the Brazilian Center of Vision Eye Hospital investigates the potential for bias in retinal diagnostic algorithms. The researchers found that accuracy in detecting diabetic retinopathy was nearly 15% higher for white patients than black patients. They believe this is due to an imbalance in training data.
MakeItTalk: Speaker-Aware Talking Head Animation
Using only an audio file of speech, this algorithm is able to animate nearly any digital image of a face to mouth along to the words being spoken. It works on photos, as well as sketches and animated characters.
Assessing Car Damage Using Mask R-CNN
Finally, someone in the A.I. ecosystem is thinking about car rental companies. These researchers built an algorithmic system able to detect scratches, broken glass, and other kinds of damage to cars.
