You know what would be a good #AutoMod type feature for #Mastodon, or any other instance software for that matter? Detecting if a new user is posting more than a certain number of posts in a brief period of time.

This thing was posting like five messages per minute. I will say that it's good that 127 of the people who got those messages apparently hit the report button, that really helps us as #moderation.

Thank you everyone who helped bring attention to this!
But looks like it was shut down by Twitch.

They misinterpreted the whole thing, perhaps intentionally. AI is being used to flag/ban without considering context and I hope that this is reconsidered.

"Twitch Global AI AutoMod does not understand context."
#Twitch #AI #BadAI #AutoMod #AIModeration #AIAutoMod #ModerationMeta #Moderation

Here it is: "Spirit AI" using "proactive" technology.

I fear this has backfired to instead mean presumed guilt until proven innocent. Makes me think of Minority Report. The road to hell is paved with good intentions and all that I guess.

From a 2022 Twitch blog post:
https://safety.twitch.tv/s/article/An-Update-on-Our-Work-to-Stamp-out-Targeted-Hate?language=en_US
#Twitch #AI #BadAI #SpiritAI #AutoMod #AIModeration #ModerationMeta #Moderation

@cragsand
Discussion regarding Twitch moderation AI spread to Reddit where I clarified some questions that arose:

Since this global AI AutoMod remains an undocumented "feature" of Twitch chat from a while back a lot of the conclusions I've listed in the thread are based on deduction from watching active chatters get suspended and tell their stories on Discord and social media.

Most can luckily get their account reinstated after appealing but it relies on having an actual human look at the timestamp of the VOD and take their time to figure out what actually happened as well as get the complete context of what was going on on stream when it occurred. I've seen many apologies from Twitch moderation sent in emails after appealing, but if you get unbanned, an apology or stay banned seems mostly random.

Being banned like this will also make it much less likely that you want to participate and joke around in chat in the future, leading to a much worse chatting experience.

I see some discussions are arguing that all AI flagged moderation events are actually reviewed by humans (but poorly) and this is a possibility. Because of the lack of transparency from Twitch regarding how this works it's very difficult to know for sure how these reviews are done. A manual report in combination with an AI flag is almost certainly a ban. One thing is sure though, and that is that too much power is given to AI to judge in these cases.

Seeing as permanent suspensions from accounts who have had active paying subscriptions for YEARS on Twitch can be dished out in seconds, either those reviewing are doing a lousy job, or its mostly done by AI. Even worse, if those reviewing are underpaid workers who get paid by "number of cases solved per hour" there is little incentive for them to take their time to gather more context when reviewing.

It's likely that if Twitch gets called out for doing this, they have little incentive to admit it as it may even be in violation of consumer regulations in some countries. Getting a response that they "Will oversee their internal protocol for reviewing" may be enough of a win which results in them actually turning this off. Since there is no transparency we can't really know for sure.

A similar thing happened on YouTube at the start of 2023, where they went through all old videos speech-to-text transcripts and issues strikes retroactively. It got a lot of old channels to disappear, especially those with hours of VOD content where something could get picked up and flagged by AI. For the communities I'm engaged in, it meant relying less on YouTube for saving Twitch VODs. It was brought up by MoistCritical about a year ago since it also affected monetization of old videos.

#Twitch #Moderation #BadAI #AI #Enshittification #AutoMod #AIAutoMod #ModMeta #ModerationMeta

@supernovae following up: here's the moderation stuff i was working on. shelving it to get back to Feditext. might be useful to others.

• https://github.com/VyrCossont/m1000 implements a webhook listener automod bot in Rust which automatically creates reports when posts or users match on configurable words, regexes, or link domains. should be able to handle another #thegx situation, or a small link-spam campaign. proof of concept, not quite prod-ready, but i'm already running it on my home instance.

• https://github.com/VyrCossont/mastodon/commit/df8be12f1190769aa530590163f9cbf56294dd52 removes the `if local?` check from Mastodon's `status.*` and `account.*` webhooks, allowing `m1000` to moderate remote posts and users.

• https://github.com/VyrCossont/mastodon/commit/528afe989e1aa1e1a1069e2f420c498e78fc9f08 adds a `report.updated` webhook to Mastodon, useful for integration with external ticketing systems.

• https://github.com/VyrCossont/mastodon/commit/0a869f0d31e6ecbf57d25451ab60b52bf35730d8 creates an API for Mastodon's audit log, which currently doesn't have one. this would be useful for admin apps.

#MastoAdmin #MastoDev #MastoMod #moderation #automod #webhook #webhooks