This is already happening. Further, embedding paid for content in the middle of original content at the server side is also already happening.
> As ML ad detection improves, the only way for ads to get past it will be to make ads that look more and more like real content, even to the human eye. Perhaps in the limit, ads and content will become completely indistinguishable.
This is also already happening. Brand placement deals are nothing new to advertising and they are basically table stakes to most content purveyors.
The real question isn't if those things are going to happen, its how those sorts of deals will handle the analytics side of the equation. How will you prove that your instagram feed is the best place for the yoga pants provider to spend their marketing budget?
I don't know whether this is a deliberate reference, but there is South Park episode playing with this idea.
https://en.wikipedia.org/wiki/Sponsored_Content_%28South_Par...
They are already doing it: Google AMP, Facebook Instant Pages.
But that only defeats simple adblocks that blacklist domains. It's theoretically possible to make an adblock that detects ads through machine learning and removes them. Doesn't matter where they are hosted.
But advertisers can run scripts to detect modifying the web page and stop it. So adblock would have to move to having a headless browser that renders the page, and then copies it and removes the ads.
Then the advertisers would have to try to trick the machine learning itself. Adversarial machine learning is a thing, with neural networks trained to fool other neural networks. And images that are highly optimized to exploit subtle defects in other neural networks.
As ML ad detection improves, the only way for ads to get past it will be to make ads that look more and more like real content, even to the human eye. Perhaps in the limit, ads and content will become completely indistinguishable.