Tools:
– Custom *horrible* Python code (numpy, pandas, nltk, matplotlib, plotly)
– photopea for extra image flair and rotating the hue for colorblindness
I don’t want to release any code because it’s bad.
I recently heard to a fake ai-generated Lady Gaga leak that included several of the hard slop words: endless, electric, neon, etc. I suspected it was Suno generated, but now I’m positive.
Suno really seems to like common but “vivid” words that might be suggested in a lesson on songwriting but probably wouldn’t ring true to the average real songwriter.
So AI has to use more “bitch” and “pussy” to sound human 🙂
Not the biggest endorsement of peak human.
“ayy shawty bust,” the rallying cry of the human resistanceÂ
It’s not clear to me what we see on the horizontal and the vertical axes, and it’s also not clear to me what the font size signifies. Could you please explain?
The vertical axis appears to be totally random, so there’s no point in e.g. comparing the top ten percent to the bottom ten percent, right?
The horizontal axis is apparently the interesting one, the one that indicates “likely word usage”. But how did you calculate that? It certainly can’t be the case that the words on the extreme left occur exclusively in “Suno” lyrics. I for sure know a few human-written lyrics that contain “joy” or “laughter”, so they must have a “likely word usage” larger than 0.0 for human-written lyrics as well. Is this something like a difference in probabilities, i.e. something like *P*(“suno”) – *P*(“genius”)? Or did you use some sort of [keyness](https://en.wikipedia.org/wiki/Keyword_(linguistics)) measure? But most keyness measures that I know aren’t restricted to a fixed data range, which your points on the x axis certainly are.
With regard to the font size, this may be related to absolute frequencies, as it’s the usual suspects like personal pronouns and articles that use a bigger font size (you know, those words that are usually filtered out in the first place). Is that really all that there is to it? If so, why even bother?
I love that “ai” is all the way over on the right by itself.
Great post, saved it! What is this possible Rap bias about? Could you tell us a bit about your dataset & method?
Stay on the left for 5-cent ice cream cones, Stay on the right to have your car keyed.
GenANI assist tools are usually censored and aligned against swearing, no wonder GenANI assist has an hard time swearing.
You can often tell text is AI generated because it’s so ridiculously over-descriptive/verbose
this would be really helpful if it didn’t include rap/hip hop
This means that using AI isn’t stealing because it actually uses different words. I thought it would use the same words since it was trained on real songs.
>might be skewed towards rap
i guess the title of the post should convey that
I hear the words on the left side every time I listen to some heavy metal
also, were the suno generations comprised of only rap songs? My guess is you’re comparing apples to oranges right there.
![[OC] Slop cloud: Likely words to appear in AI-generated audio vs real songs](https://www.europesays.com/wp-content/uploads/2025/02/cqlwg3ylg8le1-1920x1024.png)
14 comments
Data source:
– Genius: [Kaggle](https://www.kaggle.com/datasets/dell4010/song-lyrics-from-genius)
– Suno: [Kaggle](https://www.kaggle.com/datasets/rafyaa/suno-ai-music-prompts)
Tools:
– Custom *horrible* Python code (numpy, pandas, nltk, matplotlib, plotly)
– photopea for extra image flair and rotating the hue for colorblindness
I don’t want to release any code because it’s bad.
I recently heard to a fake ai-generated Lady Gaga leak that included several of the hard slop words: endless, electric, neon, etc. I suspected it was Suno generated, but now I’m positive.
Suno really seems to like common but “vivid” words that might be suggested in a lesson on songwriting but probably wouldn’t ring true to the average real songwriter.
So AI has to use more “bitch” and “pussy” to sound human 🙂
Not the biggest endorsement of peak human.
“ayy shawty bust,” the rallying cry of the human resistanceÂ
It’s not clear to me what we see on the horizontal and the vertical axes, and it’s also not clear to me what the font size signifies. Could you please explain?
The vertical axis appears to be totally random, so there’s no point in e.g. comparing the top ten percent to the bottom ten percent, right?
The horizontal axis is apparently the interesting one, the one that indicates “likely word usage”. But how did you calculate that? It certainly can’t be the case that the words on the extreme left occur exclusively in “Suno” lyrics. I for sure know a few human-written lyrics that contain “joy” or “laughter”, so they must have a “likely word usage” larger than 0.0 for human-written lyrics as well. Is this something like a difference in probabilities, i.e. something like *P*(“suno”) – *P*(“genius”)? Or did you use some sort of [keyness](https://en.wikipedia.org/wiki/Keyword_(linguistics)) measure? But most keyness measures that I know aren’t restricted to a fixed data range, which your points on the x axis certainly are.
With regard to the font size, this may be related to absolute frequencies, as it’s the usual suspects like personal pronouns and articles that use a bigger font size (you know, those words that are usually filtered out in the first place). Is that really all that there is to it? If so, why even bother?
I love that “ai” is all the way over on the right by itself.
Great post, saved it! What is this possible Rap bias about? Could you tell us a bit about your dataset & method?
Stay on the left for 5-cent ice cream cones, Stay on the right to have your car keyed.
GenANI assist tools are usually censored and aligned against swearing, no wonder GenANI assist has an hard time swearing.
You can often tell text is AI generated because it’s so ridiculously over-descriptive/verbose
this would be really helpful if it didn’t include rap/hip hop
This means that using AI isn’t stealing because it actually uses different words. I thought it would use the same words since it was trained on real songs.
>might be skewed towards rap
i guess the title of the post should convey that
I hear the words on the left side every time I listen to some heavy metal
also, were the suno generations comprised of only rap songs? My guess is you’re comparing apples to oranges right there.
Comments are closed.