So, some intrepid archive-diver re-found the publicly available training data that the AI Dungeon devs fed to the AI to make their own model which was being used with AID.

That does mean that I was misguided with my earlier explanation of the AI model thing; I'm still not 100% clear whether they re-trained their own model or simply added to OpenAI's training data, but one thing that is clear is that they specifically trained a model off of CYOA text adventures so that it would process things more correctly for their format (i.e. 2nd-person, actions preceded by a >, et cetera).

But besides revealing where commonly recurring characters that the AI would bring up nearly every adventure (names like Dr. Kessler, Count Grey, and "svelk women" became a common in-joke among the community), it also revealed the sheer lack of curation that that the AI Dungeon dev team did wrt the content they fed their AI. The first few sections (the only bit I've read so far) has child abduction, slavery, and abuse with several key words hinting that it's going to get worse, and that's just from the cursory reading of a page or two. Others who have been reading it have reported a lot of torture, murder, and rape of adults and children alike, and running this data (which I'm going to hammer home the point that the AI was trained on these stories and thus tries to match them) through word parsers reveals somewhere between 9 - 17% of it would not pass their current filter.

I also want to get up on my soapbox again about "lobotomizing the AI" because of something that I hadn't even considered until seeing some users still messing with the AI troubleshooting: words like "amusement" and "analytical" also trip the filter, for the same reason the British town of Scunthorpe does. While you can surely bypass the filter with oblique references just like you can any filter (see: rice bunny in China), you cannot escape a bad censor (see: naming something Grape in Monster Hunter) that also doesn't understand context the same way the AI generating text doesn't understand context.

This is apparently something that just came out today: that the AI Dungeon devs are crowdsourcing, possibly directly from 4chan human reviewers.

Not just that, that these human reviewers have access to stories/scenarios that are supposed to be private and that have not yet triggered the filter (confirmed via users and screenshots).

Allegedly, some of these human reviewers are minors (unknown, could be true if they are sourcing from 4chan).

Some real people have been identified due to what was a self-insert story -- name/highschool matched a dean at a quick google -- though obviously there's some room for error and might not be the same person. Supposedly there was talk around 4chan to "find and kill the pedos" from this.

At the same time, an official update has been pushed out to automatically suspend (read: ban) "the 5% of users who most frequently violate our content policies" (and this has not yet been added to their ToS), and they're denying that they're crowdsourcing moderation, but there are paper trails proving that people are being given access so at the very best they had another major data breach that has not been reported, because it includes days-old info.

I think there was more but I'm so wound up by this I'm finding hard to keep my thoughts straight.