Artificial Imitation

Last year, the good folks at Bottled Imagination asked me to dig into a deceptively simple question:
Which artists are being most copied by A.I. - and who’s losing the most potential income because of it?

The financial impact of A.I.-generated covers was impressive, sure. But I found the real gold lay in the finer details.

People aren't just using A.I. to generate deepfake Taylor Swift covers of Billie Eilish songs. They're turning anything and everything into a pop star. Who knew the entire SpongeBob SquarePants cast could drop a gangster rap album? Or that an electric toothbrush covering Eminem's Lose Yourself could rack up millions of views?

Defining the Idea

The client’s brief was really three questions in a trench coat:

1. Which songs are being covered the most with A.I.?

2. Whose voices are being used most often?

3. How much money might those artists be losing because of it?

And underneath those? Three challenges:

• Identifying and counting A.I. covers online

• Correctly linking each video to the right song and artist

• Estimating potential revenue loss based on views

The Basics

After a scout around, YouTube turned out to be the wild west of A.I. covers — tons of genuine content, only a little moderation. So that’s where I started. But unlike some search engines, YouTube doesn’t tell you how many results your search returns.

To get around that, I used Google’s search operators to return only YouTube results. I ran queries like site:youtube.com "A.I. cover" and tracked how many results came back. Then I scraped a couple thousand video titles and descriptions from each search and de-duplicated them.

This approach meant that some title formats or phrasing would slip through my search, but it gave me the best estimation possible at a genuine number of A.I. covers online.

Separating the chaff from the wheat

Amongst the A.I. “art” I had scraped were some genuine human content which I wasn’t interested in - real musicians performing real covers, people reviewing A.I. content, and much more.

To remove any human talent from my dataset, I had to be pretty heavy handed. Using some pretty stiff Topic Modelling parameters, I trimmed out any videos where the title or description didn’t insinuate the video was made using A.I., making sure my dataset was entirely soulless.

Title Finding

Now I had a big pile of video titles. I needed to translate that mess into something meaningful. Step one: figure out what songs were being covered.

So I grabbed a bunch of “most popular song” lists — based on Spotify streams, record sales, meme usage, and more.

Then, like a toddler with a shape sorter, I tried to match the titles from YouTube to known songs.

If only it was that easy.

Whose Track Is It Anyway?

Relying on a sea of random YouTube users to be descriptive in their titles is not going to get you far. Typos, missing artist names, meme references - generally incomplete data. For example, many don’t mention the original artist - especially when the song is iconic.

On top of all this, song names can appear in other artist or song names and cause potential false matches.

Take Bad by Michael Jackson. Now imagine trying to match that with titles like:

• Bad Romance A.I. Cover

• Bad A.I. Compilation

• Eric Cartman is a Baddy and Sings Bring Me To Life by Evanescence

And then what about all the songs which have the exact same name?!

This means the song-identifying system I built needed to be hierarchical.

It needed to check titles against popular songs with the longest names first, and in the case of identical titles, assume that if people don’t say the artist name they must be talking about the most popular version of that song.

This matching ran in rounds:

1. Exact match: song and artist

2. Exact match after removing special characters

3. Song name only

4. Song name match after removing special characters

I then searched the unmatched text for names of other artists, singers, cartoon characters, historical figures, and inanimate objects, to find the voices which were being super-imposed onto the track with A.I.

Putting it all together

At this point, I had:

• A rough but solid estimate of total A.I. cover songs on YouTube

• The percentage of them tied to specific songs or artists

• The most common voices being imitated — real, fictional, or mechanical

• View counts for each

• And an estimate revenue impact using Spotify payout figures

According to my research, two artists stood out as losing the most potential money through A.I. imitations of their voices - International sensations BLACKPINK and SPONGEBOB SQUAREPANTS.

And the song being covered by the most A.I. Artists - Not Like Us by Kendrick Lamar.

Want More?

Bottled Imagination published the final article over on their, and their client’s site. You can read it here if you're curious how we wrapped the project up.