Given that white noise is random, a recorded copy of white noise is its own perfect watermark.
So if you submit white noise to a copyright database, it could match different white noise recordings.
You (probably) can't copyright white noise because US copyright requires authorship. So in the same way that you can't copyright a phonebook's alphabetical list of numbers, you can't copyright random numbers rendered as sound, unless you did something else unique to it to exert authorship. It's just not something protected by copyright law. So even copying someone else's exact white noise sample is probably just fine.
The problem is that YouTube's system can only apply simple content matching rules and it counts any sufficiently long content match as a violation with no consideration of the work or context of use. Thats not how copyright law works. Copyright is a complicated system with all sorts of issues like fair use, derivative works, public domain, and works that don't qualify for protection. It's not a database query.
Modern codecs don't encode noise - they remove it during decoding and then add back artificial "comfort noise" when decoding, e.g. for film grain or background noise in voice calls.
YouTube itself seems to rely more on webm for audio. That seems to be a container for Opus or Vorbis formats.
Vorbis and Opus themselves are lossy encoders (Ogg Vorbis).
How do they preserve or modify white noise?
Apple Podcasts are AAC. Spotify are Ogg Vorbis. That probably covers 90% of the market right there.
Sufficient to support the latter assertion was that lossy compression is frequently used.
YouTube's webm does in fact use lossy compression as noted. MP3 is, I assert, still significantly used in multiple contexts, including many podcast sites. I haven't specifically surveyed those, and don't use Apple Podcasts myself (never bought into the ecosystem, though as it happens, this specific response is coming from a MacOS system). I do make heavy use of tools such as youtube-dl and mpv, and find that those do in fact frequently find and extract mp3 audio from various sources, including podcasts and IIRC Soundcloud.
Verification is as simple as:
youtube-dl -F <URL>
E.g.:
$ youtube-dl -F 'https://soundcloud.com/danyork/tdyr-362-thoughts-on-wordpress'
[soundcloud] danyork/tdyr-362-thoughts-on-wordpress: Downloading info JSON
[soundcloud] None: Downloading webpage
[soundcloud] None: Downloading webpage
[soundcloud] danyork/tdyr-362-thoughts-on-wordpress: Downloading info JSON
[soundcloud] 637919466: Downloading JSON metadata
[soundcloud] 637919466: Downloading JSON metadata
[soundcloud] 637919466: Downloading JSON metadata
[info] Available formats for 637919466:
format code extension resolution note
hls_opus_64 opus audio only audio@ 64k
hls_mp3_128 mp3 audio only audio@128k
http_mp3_128 mp3 audio only audio@128k (best)
'Opus (audio format) - Wikipediahttps://en.wikipedia.org › wiki › Opus_(audio_format) CELT includes both spectral replication and noise generation, similar to AAC's SBR and PNS'
Does opus have a notion of bitrates similar to mp3 or ogg vorbis? If so, the white noise would vary among bitrates.
Copyright algos are not looking at the whole video, they are searching for pieces of songs.
Because if you use 30sec of someones song you have to pay up and youtube is enforcing that.
Now if you will take another 2h video cut it into pieces and start searching for similar patterns in other 2h video I expect you will find some matching ones.
If the random number generator's period is 2^32 and you use one integer per sample, then at 44 kHz, you would have about a million seconds, or twelve days, before the RNG has gone through a full period.
Most RNGs have periods much higher than this. xorshift128 has a period of 2^128-1 and the Mersenne Twister's period is 2^19937-1.
So you could push it out until practically forever.
But, there will still be repeated sections.