Individually authentic signals often combine into powerful distillations, even if they're noisy.
As long as the signal is authentic–it is from the actual desire of the individual–then it is trustworthy.
If there is a consistent bias to what individual want, then the signals will tend to align, and if you multiply them all together, the noise drops out and the signal remains.
This is true as long as the bias is consistent, no matter the magnitude of the noise.
This is why the Bitter Lesson emerges.
A few places this shows up:
In the web, the vast majority of links on web pages were put there by humans asserting, "this other page is potentially worth visiting", which can then be distilled into PageRank.
In search engines, the user's queries are for their own benefit, so consistent patterns can be used as ranking signals (like the proportion of [images of foo] to [foo] queries as a signal of image intent).
In business transactions freely entered into, both parties agree that there is a useful exchange of value, which allows an emergent signal of price that reveals deep insight about relative value of things in practice.
This authentic, non-performative alignment of individual actions is one of the ingredients in all auto-cohering systems.