I turned Gemini into my personal DJ — here’s what worked and what didn’t

I’ve been looking at creative ways to use AI chatbots to solve niche, low-stakes problems I face. I’ve stumbled across a few of these in the past, such as using Gemini to practice my second language or using NotebookLM to create interactive gaming guides.

When reading about how my colleague Stephen wishes Google would bring Gemini to YouTube Music, it inspired my latest experiment — would Gemini be able to address my very picky music taste to provide recommendations I would actually enjoy?

I use Spotify for my music streaming, and while the service has an AI DJ, I’ve never found it very helpful. When it comes to recommendations and the playlists that Spotify generates, I have also had diminishing returns over time.

This is down to a few factors. Recommendation algorithms often rely on similar artists and genres to suggest new songs. But my music taste is actually very inconsistent. I may obsessively listen to a single song from a band or artist and not any of their other songs. My favorite album of all time is Kezia by Protest the Hero, but I only like one or two of their other songs.

A single change can ruin a song for me. For example, while I love the song Dead in the Corner by Guardin on Spotify, the versions on YouTube and SoundCloud have an additional beat that ruins it for me.

This means that I spend a lot of time listening to the same songs over and over again, and my Liked Songs playlist grows very slowly. For example, for September and October 2025, I had only added two songs to that playlist. For the whole of 2024, I only added eight songs.

Besides Spotify’s generated playlists and AI DJ, I’ve tried websites like Music-Map and have lurked on subreddits to find bands I might enjoy. But none of these have worked because they’re broad recommendations for broad tastes.

I also struggle to describe what exactly I like about a song in a way that makes sense to others. For example, one description I have is “it’s like my brain has an itch and this song scratches it”. But that’s not really a description that you can get recommendations from.

Since Gemini has surprised me in the past, I decided to give the AI chatbot a try to see if it could recommend songs I would actually enjoy.

I experimented with different prompts over two weeks with Gemini to generate different song suggestions. I focused on getting song recommendations, rather than just artist recommendations, due to my well-documented pickiness.

However, there were many times that the AI focused on the artist over the song. This meant that plenty of its suggestions didn’t fit the general vibe or sound of the song I enjoyed and just shared a theme or genre with the band or artist.

But the biggest issue by far was the hallucinations. On many occasions, Gemini provided recommendations for songs that didn’t exist. The first time it did this, it insisted it had just gotten the punctuation wrong. Eventually, it said that the song was no longer available on streaming platforms. I couldn’t find mention of the supposed song anywhere, so I doubted that it existed, but I tried to give the AI the benefit of the doubt.

However, the issue cropped up again and again. In these later instances, Gemini would admit it got the suggestion wrong and then suggest new recommendations. In total, Gemini recommended nine songs that didn’t exist across different conversations and days.

While I didn’t really hold it against the AI when it suggested songs I wasn’t really a fan of, the hallucinations were quite annoying since I wasted time searching for the songs and then had to go back and correct Gemini to generate better recommendations.

Despite the hallucinations and ineffective recommendations, consulting Gemini for song recommendations wasn’t a complete waste of time. It did suggest some songs that were already on my Liked Songs list, but I viewed this as a point in the chatbot’s favor since they were songs I really enjoyed.

One thing that seemed to work was taking a screenshot of a playlist, turning it into a table with Gemini, and then getting Gemini to base recommendations based off of that. This worked better than just identifying a few songs I liked from a band and asking for more similar songs.

Gemini was also able to generate descriptions for tracks that I struggled to describe. This helped with creating better recommendations. For example, it was able to describe some tracks as “theatrical rock” and “showtunes-meets-rock”, mainly based off of Foxy Shazam’s Oh Lord. But thanks to this description, I was able to get better recommendations than the broad genre of glam rock.

This particular description actually led to two new additions to my Liked Songs playlist — Holy Mother by Starbenders and Could Have Been Me by The Struts. Holy Mother in particular has become my new favorite track, which I listen to obsessively. Since discovering it, I’ve played it multiple times every day: the kind of enjoyment I haven’t experienced since discovering Oh Lord a few months back.

Through other prompts, I found a total of six more songs that I enjoyed, with another landing on my Liked playlists and the others ending up on other playlists. It also put more bands on my radar that I want to explore when I have more time. The additional songs included In Between and Disease by Beartooth, Hometown by Cleopatrick, 45 by Shinedown, Medicate by Hollywood Undead, and Black Holes (Solid Ground) by The Blue Stones.

While not every recommendation was a success, and there were definitely a few duds along with hallucinations, Gemini was mostly competent with music recommendations. I’d definitely say it was worth my time since I’m currently obsessed with the new songs on my Liked Songs playlist and have more artists I want to check out.

Past this initial experiment, I definitely want to use Gemini for more music recommendations. I also have a very specific taste when it comes to metalcore music, so I think Gemini can help me expand out my related playlist a bit more.

I also hope that by adding some more songs to my listening profile, my Spotify recommendations will improve as a result. Overall, I would say the experiment was a success and has inspired me to see what other media recommendations I can get from Gemini.