[linuxftw]

I think you underestimate the ability of AI to literally solve this problem for almost nothing. The large subs can absolutely be automated moderation, if they're not already. Enough flags on a post? Just take it down, there will be 15 million new posts in an hour.

[skilled]

Too easily gamed and not reliable. And if people find out you are using AI for modding, they will find a way to mess with it.

A human has discernment. An AI is looking at cute colors at a wall but doesn’t actually know which one is red or blue.

[tester457]

If only there was a way to prevent prompt injection.

[throw16180339]

> Just take it down, there will be 15 million new posts in an hour.

I'm a mod at a 150k+ user subreddit. AI can probably handle removing objectionable content, but I spend most of my time on other areas that are much less susceptible to automation.

• Removing requests when there are recent similar posts. We've found that users disengage when the same requests are posted every week.

• Hosting AMAs. An AI isn't going to know who we should invite or whether they'll be a good fit for our subreddit.

• Detecting and banning people deceptively promoting their own products. This is incredibly common and often quite difficult.

[linuxftw]

I'm confident an AI can be utilized for 'remove duplicate content' type of operation.

AMAs don't require a mod to participate, just the community. All mods really do is sticky a thread.

People deceptively promoting their own products, others are free to down vote them and report the post.

[throw16180339]

Exact duplicates are already automatically removed by bots. We remove request posts that lack detail or when there have been several similar active posts within the year. This strikes a balance between allowing every post only once and swamping the subreddit with nearly identical requests. We removed about 75% of the posts reported for this category last month, the remaining 25% either didn't have previous similar requests or had enough to detail to stay up. Many cases are ambiguous and resolved by majority vote.

AMAs don't happen spontaneously in major subs. As with everything else, there's a lot of admin work that goes on behind the scenes. We mostly invite people for AMAs; under 10% of this year's 10+ AMAs were requested by the participant. Someone has to email them, answer their questions, schedule a time, explain how to post, join them in a chat channel, etc.

Deceptive promotion is rarely flagged on our sub. We got ~450 reports last month and none of them were for promotion. Promotion posts are quite rare; they're against our rules and most people don't even try. What we regularly see are comments from users recommending their own products. Some cases are obvious, e.g. new accounts frequently recommending a book without ratings that was released yesterday. Many are much harder to identify; we regularly search user profiles and the Internet to confirm an identity. One person we banned last year had been covertly recommending their book for months. It took numerous searches and several weeks to remove all their promotion comments.

[1270018080]

We're in the "this is how blockchain can solve this" era of the AI bubble.

[linuxftw]

It seems to me a very trivial problem to classify a post as on topic or not with today's AI capabilities.

[whatyesaid]

Well that means bots can take anything down unless you become really good at detecting bots. Not like it's hard for bots to get points you can do so with reposting and GPT for comments.

[linuxftw]

I'm sure they have metrics at this point on how often real humans take a particular action. If an account is reporting posts higher than some amount X, then shadow ban that account.

[actionablefiber]

Do you think users will continue to report posts if they see that reporting too often will cause them to be shadowbanned?

Conversely if you are in the business of publishing spam to Reddit, wouldn’t it be useful if you could overwhelm the site with spam such that the people who care enough report all of the spam they see will get banned?

[linuxftw]

I think the vast majority of Reddit's userbase, normies, won't know the difference and won't care. It will literally be an invisible change.

[actionablefiber]

Will it, though?

For how much I see people here saying AI moderation can solve this, I frankly think AI tools are likely to be an order of magnitude better at playing offense than defense - producing huge volumes of AI-generated spam/marketing/offensive content that could fool an AI content moderation algorithm on the other side.

[linuxftw]

The quality on the popular subs is the equivalent of spam to me, so I don't think most of their users would be able to tell the difference.

[bentcorner]

I think you underestimate the ability of AI to literally solve this problem for almost nothing. The large subs can absolutely be botted, if they're not already. Enough flags on an account? Just take it down, there will be 15 million new accounts in an hour.

[jfghi]

Even if this was feasible, how much would it cost and how long would it take to successfully roll out?

[gunchie]

Reddit are already using AI for first-line moderation on their site-wide reports. It only reaches a human if there's an appeal.