>I can't see how defining 'substantive' images could ever work
You can't? Seems pretty straight forward, so long as you remember first and foremost that 100% accuracy is completely unnecessary. The "load all images" option can still be there, so all that's needed is a solution that's good enough to suffice the majority of the time and shape norms and standard-case. That leaves lot of low hanging fruit. To start with are basic heuristics like that there is no need in email for 1x1 or 2x2 or 3x3 or whatever pixel images, or images where the pixels are entirely set to transparent. A lot of tracking stuff, particularly mass scale efforts tied to multiple parties beyond the sender, is pretty blatantly content free for even a totally dumb static rule set. Slightly more smarts could be used to discriminate based on simple sender trust heuristics, such as whether the email is cryptographically signed, they're in your contacts book, whether you've ever replied or clicked through, etc. Finally this could be another area where, just as they did here with cookies, Apple applied some basic ML to answer the "likely substantive" question. Again, remember that failing here doesn't necessarily matter because it can still be an improvement over the status quo.
You can't? Seems pretty straight forward, so long as you remember first and foremost that 100% accuracy is completely unnecessary. The "load all images" option can still be there, so all that's needed is a solution that's good enough to suffice the majority of the time and shape norms and standard-case. That leaves lot of low hanging fruit. To start with are basic heuristics like that there is no need in email for 1x1 or 2x2 or 3x3 or whatever pixel images, or images where the pixels are entirely set to transparent. A lot of tracking stuff, particularly mass scale efforts tied to multiple parties beyond the sender, is pretty blatantly content free for even a totally dumb static rule set. Slightly more smarts could be used to discriminate based on simple sender trust heuristics, such as whether the email is cryptographically signed, they're in your contacts book, whether you've ever replied or clicked through, etc. Finally this could be another area where, just as they did here with cookies, Apple applied some basic ML to answer the "likely substantive" question. Again, remember that failing here doesn't necessarily matter because it can still be an improvement over the status quo.