I made a robot moderator. It models trust flow through a network that’s made of voting patterns, and detects people and posts/comments that are accumulating a large amount of “negative trust,” so to speak.
In its current form, it is supposed to run autonomously. In practice, I have to step in and fix some of its boo-boos when it makes them, which happens sometimes but not very often.
I think it’s working well enough at this point that I’d like to experiment with a mode where it can form an assistant to an existing moderation team, instead of taking its own actions. I’m thinking about making it auto-report suspect comments, instead of autonomously deleting them. There are other modes that might be useful, but that might be a good place to start out. Is anyone interested in trying the experiment in one of your communities? I’m pretty confident that at this point it can ease moderation load without causing many problems.
You must log in or register to comment.
This is an extremely important tool for moderating large forums, thank you for working on this! This is one of those mod tools that some people may need to enable large scale moderated federated forums!
Interesting concept
Is there a way of tailoring the moderation to a communities needs? One problem that I can see arising is that it could lead to a mono culture of moderation practices. If there is a way of making the auto reports relative that would be interesting.
I tried that early on. It does have a “perspective,” in terms of what communities are the trusted ones. What I found was that more data is simply better. It’s able to sort out for itself who the jerks are, and who are the widely trusted social networks, when it looks at a global picture. Trying to tell it to interpret the data a certain configured way or curtail things, when I tried it, only increased the chance of error without making it any better-tuned to the specific community it’s looking at.
I think giving people some insight into how it works, and ability to play with the settings, so to speak, so they feel confident that it’s on their side instead of being a black box, is a really good idea. I tried some things along those lines, but I didn’t get very far along.
Maybe it’d be nice to set it up so it’s more transparent. Instead of auto-banning, it can send auto-reports to the moderators with comments which it considers to be bad, and an indication of how bad or why. And then, once a week, it can publish a report of what it’s done and why, some justification for anyone who it took action against, so that everyone in the community can see it, so there aren’t surprises or secrets.
I thought about some other ideas, such as opening up an “appeal” community where someone can come in and talk with people and agree not to be a jerk, and get unbanned as long as they aren’t toxic going forward. That, coupled with the idea that if you come in for your appeal and yell at everyone that you are right and everyone else is wrong and this is unfair, your ban stays, could I think be a good thing. Maybe it would just be a magnet for toxicity. But in general, one reason I really like the idea is that it’s getting away from one individual making decisions about what is and isn’t toxic and outsourcing it more to the community at large and how they feel about it, which feels more fair.
But in general, one reason I really like the idea is that it’s getting away from one individual making decisions about what is and isn’t toxic and outsourcing it more to the community at large and how they feel about it, which feels more fair.
Yeah that does sound useful it is just that there are some communities where it isn’t necessarily clear who is a jerk and who has a controversial minority opinion. For example how do you think the bot would’ve handled the vegan community debacle that happened. There were a lot of trusted users who were not necessarily on the side of vegans and it could’ve made those communities revert back to a norm of what users think to be good and bad.
I think giving people some insight into how it works, and ability to play with the settings, so to speak, so they feel confident that it’s on their side instead of being a black box, is a really good idea. I tried some things along those lines, but I didn’t get very far along.
If you’d want I can help with that. Like you said it sounds like a good way of decentralizing moderation so that we have less problems with power tripping moderators and more transparent decisions. I just want it so that communities can keep their specific values while easing their moderation burden.
For example how do you think the bot would’ve handled the vegan community debacle that happened.
That’s not a situation it’s completely equipped to handle. It can decide what the community’s opinion of someone is, but it’s not going to be able to approach any kind of judgement call, in terms of whether a post by a permitted user is unexpectedly dangerous misinformation that the admins need to remove. That’s a judgement call that humans can’t effectively come to a conclusion on, so definitely the bot won’t be able to do any better.
There is some interesting insight to be had. One of the big concerns that people had about the bot’s premise was that it would shut down minority opinions, with vegans as a perfect example.
I tried going back and having it judge https://lemmy.world/post/18691022, but there may not be recent activity for a lot of those users, so there’s a risk of false negatives. The only user it found which it wanted to do anything to was EndlessApollo@lemmy.world, who it wanted to greylist, meaning they’re allowed to post, but anything of theirs that gets downvotes will get removed. That sounds right to me, if you look at their modlog.
I also spent some time just now asking it to look at comments from vegantheoryclub.com and modern comments from !vegan@lemmy.world, and it didn’t want to ban or greylist anybody. That’s in keeping with how it’s programmed. Almost all users on Lemmy are fine. They have normal participation to counterbalance anything unpopular that they like to say, or any single bad day where they get in a big argument. The point is to pick out the users that only like to pick fights or start trouble, and don’t have a lot that they do other than that, which is a significant number. You can see some of them in these comments. I think that broader picture of people’s participation, and leeway to get a little out of pocket for people who are normal human people, is useful context that the bot can include that would be time-prohibitive when human mods are trying to do it when they make decisions.
The literal answer to your question is that I don’t think it would have done anything about the Vegan cat food issue other than letting everyone hash it out, and potentially removing some comments from EndlessApollo. But that kind of misinformation referee position isn’t quite the role I envisioned for it.
Like you said it sounds like a good way of decentralizing moderation so that we have less problems with power tripping moderators and more transparent decisions.
I wasn’t thinking in these terms when I made it, but I do think this is a very significant thing. We’re all human. It’s just hard to be fair and balanced all of the time when you’re given sole authority over who is and isn’t allowed to speak. Initially, I was looking at the bot as its own entity with its own opinions, but I realized that it’s not doing anything more than detecting the will of the community with as good a fidelity as I can achieve.
I just want it so that communities can keep their specific values while easing their moderation burden.
This was a huge concern. We went back and forth over a big number of specific users and situations to make sure it wasn’t going to do this, back in the early days of testing it out and designing behaviors.
I think the vegan community is a great example. I think there was one vegan user who was a big edge case in the early days, and they wound up banned, because all they wanted to talk about was veganism, and they kept wanting to talk about it to non-vegans in a pretty unfriendly fashion. I think their username was vegan-related also. I can’t remember the specifics, but that was the only case like that where the bot was silencing a vegan person, and we hemmed and hawed a little but wound up leaving them banned.
The point is to pick out the users that only like to pick fights or start trouble, and don’t have a lot that they do other than that, which is a significant number. You can see some of them in these comments.
Ok then that makes sense on why you chose these specific mechanics for how it works. Does that mean hostile but popular comments in the wrong communities would have a pass though?
For example let’s assume that most people on Lemmy love cars (probably not the case but lets go with it) and there are a few commenters that consistently shows up in the !fuck_cars@lemmy.ml or !fuckcars@lemmy.world community to show why everyone in that community is wrong. Or vice a versa
Since most people scroll all it could be the case that those comments get elevated and comments from people that community is supposed to be for get downvoted.
I mean its not that much of a deal now because most values are shared across Lemmy but I can already see that starting to shift a bit.
I was reminded of this meme a bit
Initially, I was looking at the bot as its own entity with its own opinions, but I realized that it’s not doing anything more than detecting the will of the community with as good a fidelity as I can achieve.
Yeah that’s the main benefit I see that would come from this bot. Especially if it is just given in the form of suggestions, it is still human judgements that are making most of the judgement calls, and the way it makes decisions are transparent (like the appeal community you suggested).
I still think that instead of the bot considering all of Lemmy as one community it would be better if moderators can provide focus for it because there are differences in values between instances and communities that I think should reflect in the moderation decisions that are taken.
However if you aren’t planning on developing that side of it more I think you could probably still let the other moderators that want to test the bot see notifications from it anytime it has a suggestion for a community user ban (edit: for clarification) as a test run. Good luck.
Does that mean hostile but popular comments in the wrong communities would have a pass though?
They have no effect. The impact of someone’s upvote is dependent on how much trust from the wider community that person has. It’s a huge recursive formula, almost the same as PageRank. The upshot is that those little isolated wrong communities have no power unless the wider community also gives them some upvotes. It’s a very clever algorithm. I like it a lot.
For normal minority communities like vegans, that’s not a problem. They still get some upvotes, because the occasional conflict isn’t the normal state, so they count as normal users. They post stuff, people generally upvote more than they downvote by about 10 to 1, and they are their own separate thing, which is fine. For minority communities that are totally isolated from interactions with the wider community, they just have more or less 0 rank, so it doesn’t matter what they think. They’re not banned, unless they’ve done something, but their votes do almost nothing. For minority communities that constantly pick fights with the wider community, they tend to have negative rank, so it also doesn’t matter what they think, in terms of the impact of them mutually upvoting each other.
I think it might be a good idea to set up “canary” communities, vegans being a great example, with the bot posting warnings if users from those communities start to get ranked down. That can be a safety check to make sure it is working the way it’s supposed to. Even if that downranking does happen, it might be fine, if their behavior is obnoxious and the community is reacting with downvotes, or it might be a sign of a problem. You have to look up people’s profiles and look at the details. In general, people on Lemmy don’t spend very much time going into the vegan community and spreading hate and downvotes just for the sake of hatred, because they saw some vegans being vegans. Usually there’s some reason for it.
One thing that definitely does happen is people from that minority community going out and picking fights with the wider community, and then beginning to make a whining sound when the reaction is negative, and claiming that the heat they’re getting is because of their viewpoint, and not because they’re being obnoxious. That happens quite a lot.
I think some of the instances that police and ban dissent set up a bad expectation for their users. People from there feel like their tribe is being attacked if they have to come into contact a viewpoint that they’re been told is the “wrong” one, and then they make these blanket proclamations about how their own point of view is God’s truth while attacking anyone who disagrees, and then they sincerely don’t expect the hostile response that they get. I think some of them sincerely feel silenced when that happens. I don’t know what to do about that other than be transparent and supportive about where the door to being able to post is, if they want to go through it, and otherwise minimizing the amount that they can irritate everyone else for as long as that’s their MO.
I still think that instead of the bot considering all of Lemmy as one community it would be better if moderators can provide focus for it because there are differences in values between instances and communities that I think should reflect in the moderation decisions that are taken.
It definitely does that. It just uses a more sophisticated metric for “value” than a hard-coding of which are the good communities and which are the bad ones.
I think the configuration options to give more weight or primacy to certain communities are still in the code. I’m not sure. I do see what you’re saying. I think it might be wise for me, if anyone does wind up wanting to play with this, to give as many tools as possible to moderators who want to use it, and just let them make the decision. I think the bot is capable of working without needing configuration which ones are the good communities, but if someone can replicate my checking into it, they’ll be happier with the outcome whether or not they wind up with the same conclusions as me.
And yes, definitely making it advisory to the moderators, instead of its own autonomous AI drone banhammer, will increase people’s trust.
What kind of sample size of comments or time length would the bot use to make a decision? Is it safe to assume that those new Troll accounts would fly under this bots radar?
Edit:Also is there a functionality where we can look up a user using this tool? I would be interested in seeing some users I’ve interacted with in some political and news related communities.
It’s the last 30 days of comments. That’s long enough to be robust, but short enough that someone can realistically rehabilitate their image with the bot by not being a jerk for 30 days, and restore their posting ability.
I was hoping that it would be a good tool for self-reflection and fairness in moderation. In practice, the people who get banned for being jerks are totally uninterested in revising their commenting strategy, and choose instead just to yell at me that I’m awful and my bot is unfair and it should be their right to come in and be a jerk if they want to, and banning them means I am breaking Lemmy. Then they restart one of the arguments that got them banned in the first place. I don’t know what I was thinking, expecting anything different, but that’s what happened. You can see some of it happening in these comments.
New accounts, or accounts that have been recently inactive, are a hard problem. I think I’ve got it mostly worked out now. If the bot has limited information, it won’t ban you, but it will be super-strict if you have a generally negative reception, and if its unclear impression of you is negative and you also make a comment that gets downvoted, it’ll delete the comment. I think it should work fairly well, but it’s still in development. It’s hard to test, because that situation only comes up a few times a month, so I basically just have to wait a while every time I do it.
You can check a user by searching the modlog for their user, and santa@slrpnk.net as the moderator, and see what comes up. If you see that they’ve been banned at any point, then they are probably a reprobate of one sort or another.
How many new accounts do you want? I delete and recreate my account all the time. I’d be happy to make new accounts and put a bunch of comments in them for you. I could make one a week or something.
You can do that now, and evade human moderation in the same way.
I don’t want you to give it a try in the Santa communities, even though it would be a badly-needed test of the system. The code that’s supposed to detect and react to that doesn’t get much action. Mostly it’s been misfiring on the innocent case, and attacking innocent people because they’re new and they said one wrong thing one day. I think I fixed that, but it would be nice to test it in the other case, with some participation that I know is badly intended, and make sure it’s still capable of reacting and nuking the comments.
But no, please don’t. The remedy for that kind of thing is for admins to have to do work to find and ban you at the source, or look at banning VPNs or something which is sad for other reasons, so I don’t want that. Just leave it until real bad people do it for real, and then me and the admins will have to work out how to get rid of them when it happens.
Thanks for the information. I took a look at the bots community and for what it’s worth I appreciate the amount of effort you put into fine tuning it as well as being as transparent as possible.
I must say I don’t like the idea of a social-credit-score bot.
Regarding your implementation, I saw the summary of your own comments elsewhere in this post and I noticed all the annotations were on upvoted/blue segments. Other summaries you posted focused more on negative/red segments. Would it be possible to enforce a minimum of 1 or 2 from both categories?
Also, would you be kind enough to read my tea leaves? Am I an acceptable citizen of the Lemmy community?
You’re fine. Why would you not be? You left 15 comments in the last month, and they were all upvoted. It doesn’t even really have much to go on to rank you, but your rank is positive, nowhere near 0, much less far enough into the negative side that it would need to be to even be greylisted.
99% of Lemmy is made of acceptable citizens. That’s a real number. Only 1% of the users that it evaluates, which is itself only a tiny fraction of the total Lemmy population, ever get blacklisted. You have to be very obnoxious before it starts targeting you. I can understand the worry that this is going to arbitrarily start attacking people because of some vague AI bot decision, but that’s not what is happening.
The visualization of someone’s social credit score just picks the 5 most impactful posts, it doesn’t discriminate based on positive or negative. If you want to see what the red corresponds to on my graph, the most negative things I have done within the time window are:
They both contributed some red to the graph, I think. The red at the far right end is comments within this post that people are taking exception to.
(edit: I accidentally a word and didn’t realize you wrote ‘auto-report instead of deleting them’. Read the following with a grain of salt)
I’ve played (briefly) with automated moderation bots on forums, and the main thing stopping me from going much past known-bad profiles (e.g. visited the site from a literal spamlist) is not just false positives but malicious abuse. I wanted to add a feature which would censor an image immediately with a warning if it was reported for (say) porn, shock imagery or other extreme content, but if a user noticed this, they could falsely report content to censor it until a staff member dismisses the report.
Could an external brigade of trolls get legitimate users banned or their posts hidden just by gaming your bot? That’s a serious issue which could make real users have their work deleted, and in my experience, users can take that very personally.
It’s possible. I think it’s more difficult than people think. You have to do it on a scale which is blatantly obvious to anyone who’s looking, so you’re just inviting a ban.
One person swore to me that it would be really easy, so I invited them to try, and they made a gang of bots which farmed karma and then mass-downvoted me, trying to get me banned from my own place. If you look at my profile you’ll see some things which have -300 score because of it. I welcomed the effort, since I’m interested in how well it will resist that kind of attack. Their first effort did exactly nothing, because none of the downvote bots had any rank within the algorithm. I gave them some pointers on how they could improve for a second time around, and they went radio silent and I haven’t heard from them since then.
Haha they thought it was too easy and were proven wrong!
Honestly, if a place is obscure enough, even smaller barriers of entry help, like forums that don’t let you post on important boards until you build a reputation. There’s only so much effort an adversary is willing to put in, and if there isn’t a financial incentive or huge political incentive, that barrier could be low.
Finally, what the Fediverse needed:
AIs.
Good to know they’ve caught up to us from the walled gardens. Welp, I guess that’s it folks. Let’s pack up and go back to Usenet.
One of the first few instances I heard of was botsin.space, which has been around since at least 2017. Bots aren’t new. (Not sure where you’re pulling “AI” from, this is old* tech, and I don’t mean that negatively)
The core algorithm, a robust way to determine global trust rankings based on a network of relative trust, was first codified and used for this purpose in the late 1800s. It would be difficult to come up with a way to attack this problem that is further separated from the era of AI algorithms. I think you would need help from a math historian.
