[arp242]

And that other 1%[1] is just collateral damage? "Sucks to be you!" Please note they were suggesting that a single email should "blacklist the domain".

You are pretty much suggesting the very thing Microsoft is doing here.

This is not a serious suggestion in any shape or form.

[1]: A number I have serious doubts about by the way, but we'll use it for now.

[gpderetta]

There is a very simple solution to the collateral damage: don't send anything that can be construed as spam.

[arp242]

Collateral damage is by definition damage to innocents – people who have done nothing wrong.

As I mentioned before, even with the best of intentions people can "construe your email as spam".

People mark emails as spam as "fuck you". Bad support? Spam! Argument with a friend? Spam! Yes, people really do this.

People can abuse your platforms in way you didn't foresee: either an outright security flaw or a "logic flaw" (e.g. one system I worked on the rate-limiter could be bypassed by using Cc, which was of course quickly solved, but people did unfortunately use it to send out spam).

If you have any sort of "sign-up", even if paid only, people will try to abuse it to send spam.

People's computers get hacked, and while botnet spam is less of an issue due to residential ISPs blocking SMTP traffic, abusing the hacked machine's Outlook or whatnot still happens.

There's tons of cases where regular well-intentioned people send out spam. Anyone who claims any different has never seriously worked on any kind of anti-spam system with real-world usage. If this was an easy problem it would be a solved problem, but it's not, because it's a hard problem.

[gpderetta]

To be more constructive, I agree that personal spam/nom-spam decisions should not be used to train a general model, at least not without a significant signal from multiple users. Possibly users should be matched to models according to their behaviour.

But the onus should be on the model builders not on the final user.