Why Bluesky Remains The Most Interesting Experiment In Social Media, By Far

These days, everyone hates big tech, and that’s often for very good reasons. You shouldn’t trust giant centralized companies that have collected a ridiculous amount of data on you. There are few reasonable alternatives, so they can keep you stuck in their silos. They just move more and more rapidly along the enshittification curve, extracting more and more value from you for less value given back.

But, at the same time, one reason why people keep using those big tech services is because the underlying services are useful. The ability to connect with people and information around the world is still quite wonderful. However, the fact that so much of the internet has become controlled by a few giant companies remains a problem.

Some feel that this sort of thing is inevitable, and the answer will be to constantly go through a cycle of breaking up the big companies to allow the smaller ones to grow. But, it would also be nice if we got away from the setup where we just expect we’re going to hand our lives and our data over to one giant company, and went back towards the promise of the early internet, where the power and control were at the ends of the network (the users) and not the giant central nodes.

Enter Bluesky, which remains the most interesting experiment in social media. It has recently both opened up federation, but even more interestingly it has abstracted out the moderation layer (along with open sourcing tooling for people to use). This means that anyone can provide moderation services, and users can pick who they want to moderate their experience.

It may be difficult to wrap your head around how this works and why this matters, but I’m going to try to break it down with this article.

Yes, you can say that I’m biased. A little over four years ago, Jack Dorsey announced that he was going to fund a little project called Bluesky to build a decentralized and open social media protocol, based in part on my Protocols, Not Platforms paper.

So you can argue that I am predisposed to liking whatever it is Bluesky does (though, you might also note that Jack Dorsey grew so disillusioned with Bluesky and preferred the approach of a different protocol, nostr, that he deleted his account on Bluesky and spends most of his social media time on nostr*).

However, at a fundamental level, I still think that Bluesky is the most interesting experiment in social media today, even as Meta has just taken a big step towards integrating with the Fediverse, enabling users on Threads to communicate with users on Mastodon (the most widely used Fediverse service).

And it’s because Bluesky can be seen as something of the antidote to big tech, by enabling the freedom to exit, but without losing the benefits of the service you’re using. The main reason that we get these giant tech companies, and the ensuing enshittification, is that there’s no easy way to leave without losing access to the underlying services.

The Mastodon/ActivityPub/Threads (in theory…) approach is different, and hopefully better than just living in a giant silo. You can leave and still connect, but it’s something of a process. You are basically betting on the services provided by whichever instance you’re on. That could be Threads. Or it could be an instance set up by someone else. Or it could be your own instance, which would then involve having to manage it. This has some advantages, especially if you can find a server you trust and feel comfortable with, but you remain at the mercy of whoever is handling moderation for that instance (or, if you do it yourself, you have to keep up with everything going on).

Bluesky’s approach is different because it’s designed to feel like a more centralized provider, but without the lock-in or power structure. Its method of federation is effectively seamless. Personal data servers can now be federated, so that users can control their own data, or allow someone they trust to host it, but where that data is or who you’re relying on isn’t something anyone else needs to know, because it’s not necessary.

Instead, Bluesky is basically abstracting out each layer of a social media service and allowing anyone to provide alternatives at each layer. I think the clearest explanation of all this comes from a thread put together by Daniel Holmgren, who works at Bluesky. He explained the progression from when Bluesky first launched, in which it was a fully centralized system (with plans to decentralize) to where it is now.

As he explains, at the beginning, the whole service was contained on one server, or the “primordial personal data server (PDS)”:

I won’t post the images of each iteration, though it is worth watching the progression. The team focused on building towards a more truly decentralized model and were careful not to “cut corners” that would make that more difficult over time.

So, after the initial server, they abstracted out the algorithmic recommendations into “feeds.” This meant that, unlike basically every other social media app out there, on Bluesky you can create your own algorithmic feeds or choose from the tens of thousands of feeds that others have developed (or you can just take the ones from Bluesky if you don’t want to bother). Either way, you’re not controlled by “the algorithm” that Bluesky controls.

You can use any algorithm, and Bluesky might not even know what algorithms you’re using.

From there, Bluesky abstracted out the “appview,” which separated out the view of the content that people received from the data (the PDS). This would (eventually) enable federating the PDS’s. But before that, they extracted out the “relay” which would effectively act as a crawler of all the (eventually federated) PDS, to feed them into the Appview, to determine what each user would actually see (as controlled by which users they chose to follow and which feeds).

From there, actual federation began. While it was invisible to (most) users, Bluesky actually broke up all users into a set of different PDSs, such that their data was decentralized (but still controlled by Bluesky) without people even knowing it. Unlike, say, Mastodon, Bluesky itself acts as an “entryway” so that there’s no confusion when you sign up. You can just sign up via Bluesky (which opened to the public in February), and not even realize it’s a federated system, because it doesn’t directly matter.

But, then, with federation, it allowed users to move their data away from Bluesky entirely and put it wherever they wanted, while again making the experience seamless for users (setting up your own PDS is not yet seamless, but that’s coming eventually, I imagine).

So, according to Holmgren’s chart, that single “primordial” server was now abstracted out into a variety of different services, each of which is open for others to make use of, while the entire system feels like a single whole.

And, finally, that takes us to the stackable moderation layer, which also abstracts out the moderation. This includes an open source moderation tool, called Ozone (I saw a demo of this a few weeks ago, and it’s cool), and the concept of “labelers.” This means anyone, not just Bluesky, can set up their own moderation service, which users can then choose to have moderate for them.

Or, in Holmgren’s visual representation:

Now, this might all seem a bit technical, but it becomes cooler as you understand it. You can already pick and choose the recommendation side of what you see via custom feeds. That’s which content is recommended to you. For example, I have feeds of posts that are “popular with friends,” one that is of “quiet users” (to highlight those who I follow but who don’t post very often), one that focuses just on “internet policy” and one that tries to figure out what content would be most interesting to me. Some of those feeds are developed by Bluesky itself, some are developed by others. And they all work together.

But, now, by abstracting out the moderation layer and labeling, we no longer have to just rely on Bluesky to determine how to label or moderate things as well. In some ways, it’s like the flipside of custom feeds, which allow things to be recommended to you. With custom moderators, you can also choose which things you don’t want to see (or want to see in different ways — for example, maybe in a special feed you only look at once a week). As Bluesky itself explains:

First and foremost, we want Bluesky to be a great and intuitive experience as soon as you install the app. But if you want to customize your experience, you can easily browse and select from other independent moderation services and subscribe to them in the Bluesky app — as easily as you’d follow another account.

For example, someone could make a moderation service that blocks photos of spiders from Bluesky — let’s call it the Spider Shield. If you get a jump scare from seeing spiders in your otherwise peaceful nature feed, you could install this moderation service and immediately any labeled spider pictures would disappear from your experience.

Moderation services can also accept reports, so if you came across an unlabeled picture of a spider, you could report it to the Spider Shield for review.

So, for example, someone could also set up a “fact-checking” moderation service. Or a moderation service to hide all political content. Or all sports content. Whatever people want. They don’t have to change instances to get it. They can pick and choose which labelers to subscribe to, or where to send reports.

While these labelers can be people or organizations, I imagine with the rise of AI, it’s likely that people will be able to build their own custom AI-powered labelers as well, if they’d like. And then more creative things could happen. Maybe I’m sick of tech news during my days off so I could set up a labeler that labels all tech news and then says “don’t show this to me on weekends, only show it to me on weekdays.”

Or, maybe you need a break from all the nonsense, and you set things so that, on Monday, you only see good news.

And they can all work together (hence, stackable). You can choose what you want to see, not see, what you want labeled, and even more granular levels. Maybe (using Bluesky’s examples) you need to know about spiders, but don’t want to see them. You can still see the posts labeled as “spider” by subscribing to the spider labeling system, but you could automatically set your moderation controls to “hide” those images. Voila, you still get the posts about spiders, without ever having to see the images.

When all of this was being developed, my fear was whether the labeling/moderation systems would find outside developers. But, already seeing how many thousands (as of last month over 40,000) custom feeds had been developed, it really did feel like people were willing to jump in and build. But labelers are a slightly different breed, and a little more complex to wrap your head around (hence some of this post). It’s not a surprise that people may be slower to adopt setting up their own labelers, though I hope that changes.

Already, in just a few weeks, we’re seeing some very early fascinating uses of the labeler / abstracted moderation layer. Someone created a labeler for screenshots from other social media platforms. Tom Eastman once complained that the internet had become “a group of five websites, each consisting of screenshots of text from the other four” and people who don’t want to see that (or want it otherwise treated differently) can subscribe to this labeler, and set their own rules accordingly.

Or there’s a labeler designed to label “spoilers” content. You could subscribe to the labeler and set it to hide such spoilers from appearing in your feed if you don’t want to have pop culture content spoiled for you.

I’ve also seen labelers that will label content that includes images that don’t properly use “alt text,” which could be useful for those who use screen readers not to bother with.

And, even more practically, I’ve already seen a few labelers set up for other languages, like Chinese. Or a community moderation network that wants to go further in labeling trolling, antisemitism, transphobia, COVID denialism, engagement hacking, intolerance for sex work and more.

I’m reminded, somewhat, of one of the features on Mastodon/ActivityPub that is often (unfortunately) mocked by outsiders: the ability to add “content warnings” to some posts, in order to enable them to be posted minimized, in a manner that can be expanded by users. I find this to be a cool feature for certain things (including “spoilers”).

On Mastodon, however, users themselves must choose that option and adopt the content warning. This leads to fighting over norms on Mastodon that sometimes becomes contentious. Some users don’t add content warnings, while other users wish they would. On Mastodon, the only real solution is to either scold people into using content warnings or somehow punish them (such as by blocking or defederating) for not doing so. This makes the environment unfortunately hostile and uncomfortable in some cases.

But, with a setup like Bluesky’s abstracted Ozone moderation system, people could set up a content warning labeler. The labeler could use people, AI, or a combination of both to add content warnings to certain types of posts. Users who wish for those content warnings to be there can subscribe to that labeler, without requiring the people posting to do anything directly themselves.

In such a scenario, everyone is better off. Those who just want to post without thinking about content warnings can do so. Those who want content warnings on certain types of content can get them. And there doesn’t need to be any public shaming or blocking or anger about the norms.

Of course, none of this is perfect. Nor does it guarantee that all of this will work. There are certainly ways that such systems can fail. Or that they can be abused.

But, abstracting out all these layers opens up a real opportunity to have a better overall social media experience. It creates a system that feels as simple and comprehensive as a centralized system, but which is actually decentralized, and enables anyone else to jump in and provide additional services and features at each layer.

It’s one that has the benefits of a centralized system, without the serious risks of enshittification (if Bluesky goes to shit, you can just move your PDS elsewhere and adopt other feeds/moderation tools). This creates incentives for Bluesky itself not to turn to shit. But, at the same time, it opens things up for lots of others to come in and make the overall system better, but without the cruft and complexity of a system where you have to worry about which server to join, or where you stored your private key or whatever.

It just works.

But it opens up all sorts of clever possibilities that go way beyond just “building another Twitter clone,” which is what everyone else seems to be doing.

And that’s why Bluesky remains the most interesting experiment in social media.

* The last time I spoke to Jack, well over a year ago, he told me that nostr was closer to the vision of my paper than Bluesky. Having spent a fair bit of time on both, I don’t think that’s true, though I do think nostr is pretty cool as well, and could create something cool, if it wasn’t completely overrun by people who only want to talk about Bitcoin.