44.5 F
Wednesday, January 20, 2021
Home Blog Your new favorite author is a robot

Your new favorite author is a robot


Have you ever wondered if the article you’re reading was written by a human? I imagine you probably haven’t. After all, why would there be any reason to assume otherwise? What about art or music; are you certain that the works you see and hear are truly human creations?

AI is now smart enough to write like a human

If you assume that everything you read was written by a human, there’s a chance that you’re being tricked. Earlier this year, a blog post titled, “Feeling unproductive? Maybe you should stop overthinking.” reached the number 1 spot on the website Hacker News, garnering over 26,000 views on the day of its release. However, this article was not written by a person, it was written by a computer algorithm!

Liam Porr, a computer science student at UC Berkeley, used a deep learning algorithm called GPT-3 to write that article as part of a beta test launched by GPT-3 developer, Open AI. The GPT-3 artificial intelligence algorithm is the world’s most advanced language model, consisting of 175 billion “neurons” used to create text that can seem indistinguishable from human writing. The algorithm can be trained to write any kind of content by sampling tens of thousands of human written samples. It then figures out what patterns are present in those text samples and mimics them to generate text on its own. However, it’s important to note that the patterns the algorithm uses to generate text are part of a “black box”, meaning that these patterns are often not interpretable by humans.

While the text generated using this algorithm can be convincing, it doesn’t fool everyone. Some commenters on Porr’s Hacker News article suspected that the article was generated by the GPT-3 algorithm. When the algorithm is asked to create more complex forms of text, the results can struggle to feel convincing. While individual sentences that are grammatically correct and sensible are easy for the algorithm to create, complicated text that includes a narrative structure is much more difficult for an AI to produce. This has resulted in a number of hilariously bad, though still quite impressive, attempts at writing fiction, the most successful of which are curated by the humans working with the AI.

Can a robot artist find work in this era?

An AI generated image of a “Bookshop” I made using the program RunwayML with the BigGAN model. Image used with permission from the author.

The AI revolution isn’t just limited to creating text; art and music are fair game for machine generation as well. In fact, computer generated art has been around for quite a long time. In 1973, Harold Cohen developed AARON, a computer program that could create abstract paintings. If you didn’t know better, it would be almost impossible to suspect that AARON’s drawings were not created by a human. Perhaps it is because abstraction is widely accepted in the art world that works generated by AI seem convincing.

In 2018 the “Portrait of Edmond Belamy” was the first piece of AI generated art to go to auction, selling for $432,500 dollars. That portrait, while unique in style, was valuable because it was making history as the first AI-generated piece to go to auction, not because it was a master work of art. Much more affordable pieces are available for viewing and purchase from the Art AI Gallery and AIArtists.org, which showcases just how proficient algorithms have become at making art in the time since AARON.

AI image interpretations I created from a text prompt. Left: Image created from the prompt “Nana in the sunroom” (an inside joke suggested by a friend). Right: Image created from the prompt “Garden hose.” Images made in the program RunwayML using the AttnGAN model. Images used with permission from the author.

Step aside Gorillaz, there’s a new virtual band in town

Music is some of the most challenging material to make using artificial intelligence. The typical modern pop song has multiple layers of instrumentation, several passages within the music, and a human voice that sings, all of which makes a pop song incredibly difficult for an AI to make. That hasn’t stopped researchers from trying though. The Jukebox algorithm from Open AI uses a neural network that can create music from several different genres or even within the style of individual artists, including rudimentary singing.

The YouTube channel AI Radio – AI Generated Music has several examples of the algorithm at work. Some of my favorites are attempts at reimagining Bad Guy by Billie Eilish. In those examples, you can consistently hear that the more rhythmic sections of music, like the drum beats and bass lines, are done particularly well. However, I find that the Jukebox algorithm doesn’t seem to grasp the concept of melody and struggles to create the more non-rhythmic sections of music. The AI-generated songs feel directionless and aren’t very catchy. The AI singing is also fairly creepy at best and the mixing of instrumentation audio levels within each track is generally poor.

Like with visual art and text, the most successful examples of music made with AI are those that are made in collaboration with a human. In 2017 Taryn Southern, a long time creative on youtube, released her album I AM AI which was composed primarily using AI. The lead single Break Free has garnered over 2 million views on youtube to date. Taryn wrote the lyrics, composed the vocal melodies, and sang, all of which goes a long way to making the music feel human. With time, that strategy might be successful enough to get AI-generated music onto the radio and into the ears of people who don’t know that it wasn’t composed by a human.

Will our AI overlords make human creativity obsolete?

So what does this all mean for the creation and consumption of creative works like text, music, and art? Could it be that the use of AI for art devalues the talent and ingenuity of humans? I think there is space for both humans and AI to coexist in the art world, but like all forms of automation, some people may eventually be replaced by AI for the creation of works that don’t require a uniquely human perspective. The limits of this technology mean that this likely won’t happen any time soon, but when the technology advances to the point where it really can create something like this article without the need for human guidance, well then maybe you won’t be able to trust that what you see is ever truly human again.

About the Author

+ posts

Max Barnhart is a PhD studying the evolution of heat stress resistance in sunflowers at the University of Georgia. He is also the current Co-Editor-in-Chief of the Athens Science Observer. Growing up in Buffalo, NY he is a diehard fan of the Bills and Sabres and is also an avid practitioner of martial arts, holding a 2nd degree black belt in Taekwondo. You can contact Max at maxbarnhart@uga.edu or @MaxHBarnhart.

Must Read

%d bloggers like this: