If there’s timed annotations (like say for closed captions or chapters/sections), then there will be some sort of mechanism to line them up with the modified stream. Then compare that with a stream without ads (which might require manually removing all ads or using a premium account where ads aren’t inserted) and you’ll be able to estimate regions of the stream where ads have been inserted. If the timed annotations are dense, you could see where the ad begins and ends just from that.

Also if the ads themselves include timed annotations, there would be a difference in that meta data that would give it away immediately.

Or if ads are supposed to be unskippable, the metadata will need to let the client know about that. Though they could also do that on the server side and just refuse to stream anything else while it’s serving an ad.

Given that, the solution might be to have a seperate program grab the steam and remove the ads for later playback. Or crowdsource that and set up torrents, though that would be exposing it to copyright implications.