In 2020, we didn’t just get the pandemic but a new type of music competition: the AI Song Contest.

Dutch broadcasting giants VPRO, NPO Innovation, and NPO 3FM created a battleground where human composers team up with AI, blending their creative spirits with machines' cold, calculated logic.

Think of it as Eurovision with a techy twist.

This playground for tech geeks and music nerds is forecasting a new era for the music industry. One that has AI side by side with human composers. It’s a cultural shift to accept AI more into our lives.

But we should push it even further: completely AI-based music—a true AI Song Contest that could compete with Eurovision.

AI can do more than just crunch numbers, but is it enough to surpass human music?

AI vs. Human Composers

Last year, researchers from the University of York wanted to determine if deep learning methods (i.e., Music Transformer) are better than other (i.e., MAIA Markov), less trendy methods for automatically generating music.

To give some context, deep learning methods use complex algorithms (i.e., a set of rules and calculations) to analyze tons of music data. It learns patterns from existing music and then tries to create new music that sounds similar.

In other words, it’s like AI getting a crash course in music and then writing its songs.

So, researchers gathered 50 people who knew a lot about music to listen to 30-second clips of music. These clips were a mix — some were created by various music generation systems (like the deep learning algorithms), and others were human-composed. Participants rated them on six musical aspects like melody, harmony, and how much they sounded like they belonged to a particular style.

Think of a music talent show for robots.

Hype vs. Reality

The results show that newer isn’t always better.

This means that all that deep learning hype doesn’t necessarily translate to better music generation. It’s like finding out that the flashy new smartphone isn’t much better than your old one. Sure, it’s cool, but it doesn’t necessarily do the job any better.

But there’s more.

Compared to actual human-composed music, deep learning and traditional methods were still way behind. The ratings for computer-generated music were significantly lower than for human-made tunes.

This AI couldn’t quite capture the style and finesse of music composed by real humans.

Note: Consider that this study is from last year (2023), so it’s a pretty recent one.

Humans are not obsolete…yet.

Even with all the tech advancements, there’s still a significant gap between computer-generated music and the real deal. There’s still something about human creativity that machines can’t replicate.

The AI music dilemma

The noticeable gap in quality between human and AI music shows that deep learning still needs to catch up in music nuance and complexity.

That’s a breather for human composers.

However, deep learning systems are in even bigger trouble.

They are copying chunks from their training data and creating music similar to their dataset. This points to a lack of originality and plagiarism or copyright infringement. If these songs are released, they could potentially be sued!

Consider the lawsuit against Anthropic by Universal Music Publishing Group. If AI companies use copyrighted material for training large language models (LLMs) like Claude (Anthropic’s AI assistant), what type of songs will it generate?

Most likely those similar to the dataset.

Universal Music Publishing Group sued Anthtropic for using music without proper licensing agreements, but what if they got these training datasets legally?

How to craft a new tune with AI

It’s the creative domain that’s in danger.

The quality and nature of the training data profoundly influence the AI model’s outputs, so there’s more work to do on what AI can generate from these datasets.

Here are some things to consider:

• Filtering repetitive music: before training, remove overly repetitive or dominant patterns. This will prevent the AI from learning to replicate specific pieces too closely.

• Diversify the data: include a wide range of styles, genres, and sources. This reduces the likelihood of replicating specific works.

• Constrain creativity: steering the AI toward certain themes, structures, or styles could create more original outputs.

• Have a hybrid approach: why not have human input as well? AI could be a tool for initial ideas or drafts, which human artists then refine or make the other way around.

There are many other things developers could do to reduce the risk of ending up with a similar song to those in the dataset, but it’s a good first step.

Final thoughts

The high-tech apprentice (i.e., deep learning methods) can’t outshine either older tech or the human touch in music creation.

At least not in a way that stands out to trained human ears…for now.

Music is more than just how it sounds but how it makes us feel. AI music has a lot of ground to cover.

However, with LLM, we’ve seen huge progress. Just look at this article I wrote about Google’s MusicLM. The pace of progress is astonishing, and I would guess it won’t take much more time to make AI music sound very similar to human music.

It’s just a matter of when not if.