[parliament32]

> To try to do something about it requires curation, somebody needs to decide what is junk, which is completely antithetical to open protocols.

The contra-example, of course, is email. SpamAssassin figured this out 24 years (!) ago. There is zero reason you couldn't apply similar heuristics to detect AI-slop or whatever particular kind of content you don't want to accept.

> Radical recentralization to companies that had effective spam management

Only for the lazy.

[gjsman-1000]

A. SpamAssassin has never been tested at Gmail scale, and would likely fail in such a scenario.

B. SpamAssassin is benefiting from centralized players, like Gmail, harming spam's economics. You're a free rider from the onslaught that would occur if spamming actually worked. Spam is at 45.6% of email globally with aggressive spam filters, but could easily double, triple, quadruple in volume if filters started failing even moderately. Weaker filters, and we'll start seeing the email DDoS for the first time.

C. Heuristics on AI Content? What are you going to do, run an "AI Detector" model on a GPU for every incoming email? 376 billion of them every day to Gmail alone? This only makes the email DDoS even more likely.

D. Lazy = 99%+ of global computer users - and that's changing as soon as everyone becomes their own paramedic. If you can't convince most people to learn how to save other people's lives, and probably didn't bother yourself, despite it being disproportionately more important, you're never teaching them technical literacy.

[parliament32]

I think you misunderstand what I'm getting at. SpamAssassin is older than Gmail. It's an old example, much newer and better spam-filtering-at-scale solutions exist (although SA is still maintained). Trying to claim that only the big boys can filter spam is an uninformed opinion.

No, you don't need an AI model to detect AI content (lmao). Heuristics already exist, and you see people mention them online all the time -- excessive use of lists, em dashes, common phrases, etc. Yes, a basic text heuristic scorer from the 1980s can pick these up without much difficulty. The magic of auto-learning heuristics (which have also existed since the 1980s, and performed fine at scale with less processing power than your smartwatch) is you can train them on whatever content you don't want to receive: marketing, political content, etc. You can absolutely apply this to whatever content suits your fancy, and it doesn't really take any more effort than moving messages you want filtered out to a Junk folder or similar.