Your Songs Know What You Did in the Dark
Recently I made a data export request to Spotify for a copy of my listening data. I also asked for the technical logs, that dwarfed the streaming history thirty to one. What came back was 12,500,872 data points about me, spread across 218 JSON files totalling 462.2 MB. Together it is an itemised record of a considerable portion of my life from July 2009 to February 2026.
The excessive data retention didn’t surprise me, but I was struck by just how much someone could infer about me from it. I was expecting to write something light about my more eclectic tastes, but instead this has ended up being a bit more serious.
Some numbers
Let’s start with some data. Across 16.6 years, I streamed 307,148 tracks totalling 8,951 hours, which is just over a year of continuous listening. I listened to 44,153 unique tracks, 19,059 artists and 31,054 albums. There were 65 distinct platform strings in my data, which were enough for me to reconstruct my full device ownership history over the period. You can even see the timing of individual OS updates. Spotify has retained a far more complete record of this than I have.
Every playback record also includes an IP address. Spotify has logged me listening from 4,018 unique IP addresses spanning 15 countries. I geolocated 3,350 of them, and used this to create what is essentially a map of my movements.
So far, as expected, but I started to see patterns in the listening data that were quite revealing.
My sleep
By converting the UTC timestamps to the local time of where I was, and searching for gaps of 4–14 hours between consecutive streams, I identified 1,618 plausible sleep periods across the dataset. Here’s a snapshot.
| Year | Avg Bedtime | Bedtime Variability | Nights Measured |
|---|---|---|---|
| 2019 | 00:18 | ±107 min | 134 |
| 2020 | 00:05 | ±133 min | 91 |
| 2021 | 23:03 | ±118 min | 115 |
| 2022 | 23:10 | ±85 min | 222 |
| 2023 | 22:56 | ±51 min | 270 |
| 2024 | 23:53 | ±79 min | 188 |
| 2025 | 00:12 | ±116 min | 105 |
I found evidence of 35 all-nighters, that is to say, continuous streaming between midnight and 8am with no gap of two-hours or more. Almost all of these were in the last few years. The longest was nearly eight hours of German rap in April 2024. I know that that was because things were difficult at work, but even without the why, Spotify had still captured the what in much higher resolution than my own recollection had.
The breach
The data showed a period when 3am streams spiked from zero to ninety-one in the space of a fortnight, starting in early August 2023. Out of curiosity, I pulled more data from that period, and quickly realised that this exactly matched the timing of a serious data breach that I had had to deal with. Those ninety-one streams are evidence that I wasn’t sleeping. I knew this, of course, but not the extent of it. It wasn’t workload, rather having a full understanding of what the consequences could be for those affected and their families.
| Period | 3–5am streams |
|---|---|
| June 2023 | 0 |
| July 2023 | 9 |
| August (pre-breach) | 4 |
| August (breach period) | 91 |
| September 2023 | 14 |
| October 2023 | 0 |
I can see that I woke up at my usual time the day it started, that I went for a run whilst listening on my GPS watch, then logged into my work laptop about 8am. I was clearly in a good mood, as the tracks are mainly light and playful. I was listening to “Walk Like an Egyptian” by the Bangles when I took the file down, but this had escalated to Ride of the Valkyries by 9pm.
From the data, I can see that I was awake from 3:14am on the morning after the breach listening to tracks that were not typical for me, indeed all of the 3am sessions contained at least some religious music or tracks normally associated with grief. I have no memory that this is what I was listening to, or that I had reached for my phone in the middle of the night, but Spotify never forgets. They could see that something had changed.
This made me curious as to what else the data might allow them to infer about my state of mind or wellbeing.
My mood
I ran further analysis, and found that a distinct pattern of listening had appeared during a difficult period of content moderation in early 2024. I hadn’t understood the impact of that until more recently, but it is clearly visible here. The data shows long listening hours in the week and triple what was normal for laptop streams on the weekends. Over that time, my night-time listening more than doubled compared to the same period a year earlier. There was much less skipping or shuffling than usual, representing long periods of intense focus. The content got steadily faster in terms of bpm, then darker and angrier as time went on. The algorithm saw this and responded by serving me up more of the same - I still have some of the playlists that it generated.
I’ve wondered for a while whether algorithm-driven listening might feed the moods it detects as much as it tries to reflect them. If the algorithm decides that I am feeling low at 2am and serves me darker music, does hearing it keep me in that mood for longer than I would have stayed there on my own? The data by itself can’t answer this, but the question still troubles me.
There were days where I can see that I frequently skipped many successive tracks within seconds. This browsing pattern repeats the most during what I know to have been quite stressful times. Spotify has captured a snapshot of a slightly agitated state of mind that I don’t think I could have described from memory alone. It is a very strange feeling to see it mapped out in that way.
Then there are the daylists. Spotify generates these several times a day based on what it thinks I might want to hear. The export showed the 634 descriptive labels from these, tagged by day of the week and time of day. Most are benign like “retro,” or “classic rock.” Others a bit less so. Spotify has seen fit to label my listening as “furious”, “gloomy, cry it out”, “venting, fearful”, “lovesick, emotional” and even “lycanthropic”!
Those are not genre labels in any conventional sense, rather products of a system that seems like it is trying to pinpoint your emotional state to keep you engaged. On any given day, viewed in isolation, they seem like a bit of fun, but taken together, I’m much less sure. My late-night daylists are three times more likely to carry dark or sinister labels than the daytime ones. At midnight on a Tuesday, Spotify labelled me “fearful, strange.” At midnight on a Wednesday: “venting, weird, fearful.” Their algorithm knows that a late night version of me might be different from the daytime one and tries to label both accordingly.
My thoughts
I could repeat this sort of analysis in greater detail for other life events. Spotify has data on the voice in my ear, and a glimpse of geolocated thoughts and feelings, from weddings, births, funerals, holidays, house moves, new hopes, endings, beginnings, and focused intent. For much of that, they would not need any input from me to try and read me.
Up until now this was data that I had considered to be mundane. It’s just what songs I played. And it is. And yet, at sufficient scale and over a sufficient period of time, it is also in part a health record, a mood diary, and a travelogue that forms a detailed portrait of my life. I was not prepared for the extent that 307,148 streams across 16.6 years could allow a competent analyst to reconstruct periods of insomnia, deep emotions (evidenced through repeat listening spirals and genre shifts) and evolving political views.
The data that has been collected here is excessive. The storage of such granular data over such a long period of time is not necessary for any justifiable reason. It is the sort of thing that GDPR was meant to rein in.
The next step for me, now I have a copy of the data myself, will be a Right to Erasure request.
There were 31 streams of My Songs Know What You Did In The Dark (Light Em Up) recorded in the export, which inspired the title of this post.
© 2026 foimonkey. All rights reserved.