But at the time, it was considered to be a big do not touch -- just don't do this. Not so much for ethical reasons, but for optics in the industry. (I wasn't proposing doing it, was just curious)
In the meantime, though, this seems to have just become standard practice, but way more sophisticated with way higher accuracy, as this article touches on.
What was not acceptable a decade ago is now "ok." Not just by sketchy ad startups, but by major players.
But this whole mess ties back to one of the things that worries me the most about the propagation of LLM type ML out into the general industry. It's only a matter of time before ad targeting takes on an extra dimension of creepiness through this (and I'm sure it's already happening in some aspects, inside Google & Meta.)
In the past, in ad tech & search, etc. people could say things like: "Yes, it's highly targeted. Yes we've co-related an absolutely huge quantity of data to fingerprint you exactly, and retarget you. But it's anonymized. No humans saw your personal data. It's just statistics.". Not saying whether or not this argument has merit or not, just repeating it.
But now, here we are, where "just statistics" is a far more intricate learning model. One which is capable not just of corelating your purchases and browsing activity, but of "understanding" you, and which -- while not an AGI -- is pretty damn smart.
At what point does "a computer scanned your browsing for patterns and recommend this TV set" become ethically the same as "a human read your logs, and would like to talk to you about television sets..."?
Having worked in ad-tech before (and having worked at Google, in ads and other things as well), I do not trust the people in that industry to make the right decisions here.