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

I’ve been exploring creative ways to use AI chatbots to solve niche, low-stakes problems I encounter. Over time, I’ve discovered a few interesting applications, such as using Gemini to practise my second language or employing NotebookLM to create interactive gaming guides.

When reading about how my colleague Stephen wished Google would bring Gemini to YouTube Music, it sparked my latest experiment — could Gemini provide music recommendations that cater to my very specific tastes?

Would you use an AI chatbot for music recommendations?

I’ve been stuck in a bit of a musical rut, so I decided to turn to Gemini. I use Spotify for my music streaming, and while it offers an AI DJ, I’ve never found it particularly useful. Over time, the playlists and recommendations generated by Spotify have become less effective for me.

This issue can be attributed to a few factors. Recommendation algorithms typically rely on similar artists or genres to suggest new songs, but my musical taste is rather inconsistent. I might listen to one song obsessively from an artist, yet I have no interest in their other tracks. For instance, my favourite album of all time is Kezia by Protest the Hero, but I only enjoy one or two of their other songs.

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

I’m extremely picky about my music, and this inconsistency across artists, albums, and genres makes it difficult to even pinpoint my favourite band. As a result, I often find myself listening to the same songs repeatedly, and my “Liked Songs” playlist grows at a snail’s pace. For instance, during September and October 2025, I added only two songs to the playlist. In 2024, just eight songs were added in total.

Beyond Spotify’s AI-generated playlists and DJ, I’ve explored websites like Music-Map and browsed subreddits for recommendations, but these have proven unhelpful as they offer very broad suggestions.

I also struggle to articulate exactly what I enjoy about a song in a way that makes sense to others. A common description I use is, “It’s like my brain has an itch, and this song scratches it,” but this isn’t exactly a helpful way to describe music preferences to others or to an AI.

Since Gemini has surprised me before, I thought I’d give it a try to see if it could recommend songs that I’d actually like.

Some things really didn’t work

I experimented with different prompts over two weeks with Gemini, aiming for song recommendations rather than artist suggestions, given my well-known pickiness. However, there were many instances where Gemini focused more on the artist than the song itself. As a result, many of its recommendations didn’t match the general vibe or sound I enjoy, often only sharing a theme or genre with the artist.

But the biggest issue by far was what’s known as “hallucinations.” On several occasions, Gemini suggested songs that simply didn’t exist. The first time it happened, it claimed the song’s punctuation was incorrect, but eventually it stated that the track was no longer available on streaming platforms. I couldn’t find any trace of the song anywhere, and I began to doubt its existence. Still, I gave the AI the benefit of the doubt.

Beyond the occasional generic recommendations, Gemini had a tendency to hallucinate song titles. This issue cropped up repeatedly. After each false recommendation, Gemini would admit its mistake and suggest new tracks. In total, it recommended nine songs that didn’t exist across various conversations.

While I wasn’t too bothered by suggestions that didn’t suit my tastes, these hallucinations were frustrating. I wasted time searching for non-existent songs, only to have to go back and correct Gemini in order to get better recommendations.

But a few things worked really well

Despite the hallucinations and the occasional misfire, consulting Gemini for song suggestions wasn’t a complete waste of time. It managed to suggest songs that were already on my “Liked Songs” list — a definite win in my book, as these were tracks I truly enjoy.

One of the more successful methods involved taking a screenshot of a playlist, turning it into a table, and then using Gemini to generate recommendations based on that. This approach worked far better than simply identifying a few songs I liked from a specific artist and asking for more similar tracks.

Additionally, Gemini was able to provide descriptions for songs I found hard to explain. For instance, it described certain tracks as “theatrical rock” or “showtunes-meets-rock,” drawing from Foxy Shazam’s Oh Lord. These descriptions helped refine the recommendations far more than simply classifying the genre as glam rock.

Despite the challenges, this approach led to some fantastic discoveries. Thanks to the more accurate descriptions, I added Holy Mother by Starbenders and Could Have Been Me by The Struts to my playlist. Holy Mother in particular has become my new favourite track, and I’ve been playing it obsessively every day, much like I did with Oh Lord a few months ago.

Through other prompts, I found six more songs I enjoyed, with one making it onto my Liked Songs playlist and the others landing on various other playlists. It also introduced me to bands I’m keen to explore further when I have more time. These 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 hit, and I encountered a few duds along the way, Gemini generally proved competent when it came to music recommendations. In the end, I think the time spent experimenting was worth it. I’m now obsessed with several new tracks on my playlist, and I’ve got more artists to check out.