Today I heard sort of offhand that one way people can detect AI music is to look for compression artifacts in the tracks. At first I thought, well how would you know the difference between compression artifacts that are supposed to be there versus "bad" compression artifacts introduced by an AI trained on compressed media?
Apparently the answer is that you can distinguish compression artifacts that show up in a 128kbps mp3, and if those show up in an mp3 labeled as 256kbps, then it's a signal that an AI generated it.
Does anyone have any more detailed info on this, or if there's more to these methods? Sadly it does seem like the kind of detection method that could be bypassed: just have a separate 256kbps model only trained on 256kbps tracks.
It also makes me wonder about text... what has been tried and known not to work? I suspect nothing has been found to work for real.
Jeremy Kun
npub1pe…4h36q
2025-10-20 15:49:00 UTC
Author Public Key
npub1pe67uceztjyefem3xmphw05e65200df0hrrv60464v6w8xjtgvgqp4h36qSeen on
wss://relay.ditto.pubPublished at
2025-10-20 15:49:00 UTCEvent JSON
{
"id": "db744322e20fff896e039634c6ddb55f700457823199ad43babb82dd5ec60470",
"pubkey": "0e75ee63225c8994e77136c3773e99d514f7b52fb8c6cd3ebaab34e39a4b4310",
"created_at": 1760975340,
"kind": 1,
"tags": [
[
"proxy",
"https://mathstodon.xyz/users/j2kun/statuses/115407279921539360",
"activitypub"
],
[
"client",
"Mostr",
"31990:6be38f8c63df7dbf84db7ec4a6e6fbbd8d19dca3b980efad18585c46f04b26f9:mostr",
"wss://relay.ditto.pub/"
]
],
"content": "Today I heard sort of offhand that one way people can detect AI music is to look for compression artifacts in the tracks. At first I thought, well how would you know the difference between compression artifacts that are supposed to be there versus \"bad\" compression artifacts introduced by an AI trained on compressed media?\n\nApparently the answer is that you can distinguish compression artifacts that show up in a 128kbps mp3, and if those show up in an mp3 labeled as 256kbps, then it's a signal that an AI generated it.\n\nDoes anyone have any more detailed info on this, or if there's more to these methods? Sadly it does seem like the kind of detection method that could be bypassed: just have a separate 256kbps model only trained on 256kbps tracks.\n\nIt also makes me wonder about text... what has been tried and known not to work? I suspect nothing has been found to work for real.",
"sig": "1b1c7d88016dd1fc60471f301a4a9cd273cbce2f45e74dbec4aa253868ed3f530d4ab1fbc368593fb3c86f7d6962bc0d10a71f5ff951b8a7ccf2445e8ce00d84"
}