An honest review of running three unfinished demo tracks through an AI composition tool

Two in the morning, headphones on, staring at the Ableton screen. An eight-bar synth loop was blinking over the cursor, and for three days I hadn't been able to figure out what to layer on top of it. I could hear a tram rattling past the window from the direction of Kraków's Old Town, and I caught myself thinking the vibration lined up oddly well with my kick drum pattern. This wasn't inspiration. It was sleep deprivation.

So what I eventually reached for was AI composition tools. To be specific, I took three electronic music sketches I'd half-finished and abandoned, each with a different mood, and fed them into three free tools under the same conditions. I wanted to find out which tool actually helped finish a track, and which ones just chipped away at my self-esteem. In a needlessly rigorous effort at fairness, I prepared identical MIDI and audio stems for all three tracks and fed them into each tool. My methodological obsessiveness comes in handy at times like these.

The Three Demo Tracks Used for Testing

Here's the short version. The first track was ambient: pads layered over a field recording I'd made at a café in the Kazimierz district. The second was a more rhythmic IDM-style sketch, 140 BPM with glitch elements. The third was a slow synthwave-ish track whose melody line cut off completely in the second half. What all three had in common was that I'd stopped because I didn't know what to do next.

Results by Tool

I won't name them specifically. Versions change too often, and there's no guarantee that my experience at the time matches how things are now. Instead I'll call them A, B, and C.

Tool A was the type that continues composing after you upload audio stems. When I fed it the ambient demo, it produced a fairly convincing texture. The problem was that the texture was so disconnected from my original that it felt like someone else's track with my intro forcibly glued onto it. For the IDM sketch, it completely ignored the rhythm and slapped a 4/4 house beat on top. I muttered to myself, "Right, so that's what you think electronic music is." For the synthwave track, it did continue the melody, but the progression was so predictable that it made me reconsider why I'd stopped there in the first place. I'd stopped because I didn't want to take the obvious path.

Tool B worked on a MIDI basis. You feed it your MIDI file and it suggests the next section. Of the three, it was the most usable. The chord progression it suggested for the ambient track was close to the direction I'd been vaguely imagining. Of course, "close" might be my own hindsight rationalisation. People see what they want to see. For the IDM track, it respected the time signature to some degree, and I actually took one of its hi-hat patterns and used it. Strange feeling, that: in the final track you can't tell whether that hi-hat was mine or the AI's. For the synthwave track, it offered three melody suggestions, but all of them were too major-key to fit the mood of my track.

Tool C billed itself as a "co-composition AI," a format that responds in real time as you play. The concept was appealing. In practice, the latency was so bad I couldn't tell whether to blame my Kazimierz apartment's Wi-Fi or the tool itself. When I laid down pads on the ambient track, something followed half a beat behind, but it was closer to a delayed echo than a conversation. On the IDM and synthwave tracks, it flat-out ignored my input and kept spitting out its own phrases, so I just became a listener. "Jamming" with AI sounds nice, but jamming presupposes that both sides are listening. This was a monologue.

Honestly, Did It Help

After running all three tools, when I reopened my Ableton project file, the only thing I actually used was that one hi-hat pattern from Tool B. But the strange thing is, the process itself wasn't useless. Listening to the AI's "wrong" suggestions actually made it clearer what I wanted. Why I'd stopped at those eight bars, what I was trying to avoid. The obvious options the AI showed me served as a kind of reverse compass.

Whether that's the real value of these tools, or whether I'm forcing a positive interpretation just to assign meaning to the outcome, I'm not sure. That's generally how my anxious mind works. Over-analyse everything, then analyse the analysis.

I finished the ambient track that night. Not a single AI suggestion made it in, and the pre-dawn air drifting in through the window from Planty Park was better than any texture the tools had offered. I still don't know if the track is good. What matters is just that it's done. One fewer unfinished piece.