I don't know what to think about it.

When you use a free software, don't mention it

https://epsilon.social/about

First, they completly hide the fact that it is based on Mastodon in their main page. You have to check their source code where they explain technical detail about their fork.

Epsilon is a French social network built on Mastodon. It is a fork of Mastodon v4.6.0 that adds content categorization, AI-assisted moderation, and a complete redesign of the web interface.

I think Mastodon developers deserve a little more love and credit.

AI moderation

Then, they openly admit that they use AI for moderation and probably Claude. I should warmly thank them for their transparency :

Human moderation

Moderation happens before publishing: every post is first checked by AI. Sensitive content, cases where the AI is unsure, and anything flagged go through human review.

If it is "human moderation", then every post shouldn't be checked by the AI, isn't it ? We know and hear everytime that "we lack people", "we just execute some script that make our life easier". But, AI, no thank you.

We recently have a post from an user that used AI for reviving a dead project, Castodon, that he loved. He finally found some dev help. But, for me, that's another scale from analysing every posts we made by an AI.

Can we review its code inside the blackbox ?

I not doing moderation, i also build a para-social relationship with you, good or bad. Sometime i post, sometime i intervene, sometime i protect myself, sometime i talk about myself. I'm an user that do moderation and animate communities in the fediverse. That maybe the huge difference between creating a community and using a tool for mass moderation process.

Another problem remain, and that's the exact same reason we blocked Threads (Meta) : if we interact with epsilon.social users, our com will be probably analysed by the AI.

Privacy policy

https://epsilon.social/privacy-policy

According to their privacy policy, they claim that 3 principles governe their project :

Trois principes gouvernent l’ensemble. Nous ne vendons ni ne louons aucune donnée personnelle. Nous ne pratiquons aucune publicité fondée sur le profilage de votre comportement. Nous n’utilisons pas vos contenus pour entraîner des modèles d’intelligence artificielle.

Quick translation :

• They don't sell nor rent any personal data
• They don't use tracking ads based on our profil behaviour
• They don't use your data to train AI

They will probably apply it for the time being. How long ? How many time have we seen privacy policy change on a service provider ?

The good side, for european citizen, is that they are on an european tech stack and GDPR compliant. That's a good thinking for european sovereinety and thus limited by european sphere. But the question is : Can you trust it ? It is directly feed to an AI anyway. We will never know where our data will go, except some promises written on a web page.

https://github.com/epsilon-social/epsilon/tree/epsilon-v0.3.24

The code source mention that they use mistral for moderation process.

Personal opinion

Honnestly, i can't trust them.

At this point, I would rather trust my own server, some friend or my local association that endorse free software for years.

I think, as we did for Threads (meta, facebook...) the safest option is to defederate or block their domain epsilon.social

@Snoopy Booo!

Booo ? Why ?

At this point the whole french fediverse is aware and analysing it.

Moderation happens before publishing: every post is first checked by AI. Sensitive content, cases where the AI is unsure, and anything flagged go through human review.

Ben, in terms of that part, I suspect that pretty much already is or will be the model for most social media services, large and small.

True, but not for our service.

And we already know from corporate social media that AI ends up targetting under privileged people even more than human moderators do. In the fediverse, where most moderators care about the health of the community, the gap will be astronomical.

Interesting. I wonder what the motivation behind this project is. I suppose a small Mistral model could be run without much expense. But still, how is this development being funded?

Any public facing website will be scraped. That's just the nature of the modern internet. I don't think you need to worry about blocking their domain.

Overall i agree, too opaque for me to trust. Probably wouldn't make an account there. But i don't see an issue interacting with the users of that instance.

Hmmh. The commits look like it's vibe-coded. And I always like to question people's sanity or knowledge of AI once they forego things like classifiers and decide to ask an LLM instead. I mean sure, LLMs can mimick a lot of things, including mimick content classification. But why not use the AI thing that's made for that job? It'd probably be more accurate and surely less expensive. And it could solve the problematic decision to forward all content to a third party.

@Snoopy Considering, that there are known cases of actually storing the users' data by models, I wonder how epsilon is going to ensure this:They don’t sell nor rent any personal dataand this:They don’t use your data to train AIThat smells like defederation to me.#fediverse #fedimeta

…Honestly, you shouldn’t preclude it forever. I think it’s a great idea for filtering spam, because at some point the flood of LLM spam agents will not be tenable to moderate with human labor and traditional scripts.

It doesn’t have to be Claude. Small models (or better yet: classifiers) trained on open datasets can filter spam effectively. And it doesn’t have to make moderators/admins unreachable, like Reddit.

Is it actually federated?

If so, I’d say that’s a good thing, and that this problem will fix itself. I it becomes problematic, other instances will just defederate.