2025-11-11

Chicken Caesars: they're messing with your Bluesky feed

It seems there's some underhanded algorithmic fuckery happening in the ATmosphere. My best guess, based on the available information, is that the Bluesky team is running some sort of LLM-based social experiment affecting reply visibility. There's a fair bit of confusion as to what's actually happening, and a lack of clear communication (perhaps combined with attempts at outright obfuscation) hasn't helped.

I've tried to be as careful and methodical as I can while, as a non-expert, digging into the odd behaviour of the platform that I've observed first-hand. I wrote a short bsky thread [1] about it, but figured it would be good to do this longer blog post and have all the info in one place. If you have good reason to believe I've gotten something wrong here, feel free to message me on the fediverse or on bluesky and I'll correct the record in a transparent way.

Hidden replies

A couple weeks ago, I saw some Bluesky users raising concerns that the Prime Minister of Canada's account was hiding replies from critics. Meaning, a reader would have to scroll down to the very bottom of the comments, and then click a "show more replies" button in order to see them.

That's an eyebrow-raising accusation, and if true, would be a bit of a scandal. It wasn't immediately clear if the PM's team were in fact responsible, but one thing was undeniable: a lot of replies were (and still are!) being hidden. And although the policies [2] say that the PM's social media team may remove comments and/or block users for things like hate speech, advertising, or "coarse" language, many of the hidden comments are pretty innocuous, and in some cases they're even supportive.

So is the PM's account really hiding these replies?

The PM is not hiding the replies

Well, here's the thing. If I make a post, there are only two available options for hiding a reply: "Hide reply for me" or "Hide reply for everyone".

But, based on other comments, many of the hidden replies were only being hidden for *some* users, not everyone. (I was able to confirm this first-hand later.)

So I'm pretty sure it's *not* the PM who's hiding the replies.

It's not moderation labels, either

Regardless, I wanted to know what the deal was. Conveniently, I'm also an annoying weirdo whose hobby is to clap back at the bland, focus-group tested, consultant-approved, empty pablum technocratic rhetoric that gets half-heartedly delivered by The Right Honourable Mr. Mark Carney - so I started double-checking the visibility of my own replies to the PM using an alt account.

Sure enough, a few days later, one of my own replies to a post by the PM [3] got hidden behind a "show more replies" button.

Maybe it was something to do with the Bluesky Moderation Service? Maybe it flagged me as 'rude' because (gasp) I say 'fuck' or 'shit' sometimes?

Well, if it did, here's how that's supposed to work:

- Bluesky first filters out illegal content.

- Then, posts and/or accounts may have 'labels' (such as 'rude', or 'adult content', or 'extremist' content) applied to them by the Bluesky Moderation Service. This is the moderation service that Bluesky users are defaulted into using.

- For each of these labels, there's a user-defined setting that defines how the label affects the visibility of a post or reply. A user can choose to either ignore that particular label ('off'), place a content warning on a post or reply with that label ('warn'), or hide it entirely ('hide').

So I tried turning off all the moderation settings for my alt account, to see if anything changed. No - my post was still hidden.

Someone pointed me to Kuba Suder's label scanner tool, where you can enter a post's url or an account name, and see if there are any labels attached to it. [4] So I tried my username: no labels. Then I tried the URL [5] of my reply: there's a 'politics' label, but that's not one of the Bluesky Moderation Service labels, so it wouldn't be affecting visibility for my alt.

Inconsistent and opaque. Open protocols FTW!

I asked around. One helpful suggestion was that this could be due to a known bug [6] where long threads, or conversations with many branching replies, can hit some kind of database query rate-limit (or something like that). This didn't explain it either though, because my reply was one of the first dozen or so, and as far as I could tell, had been hidden immediately.

Maybe there's some other software edge-case bug, and an API response was being cached on the server side or client side? I can't do anything about the server side cache, but I did try clearing the cache in my browser and app. No change. (And, for the record, still no change almost two weeks later.)

What about the fact that there was a link preview in my reply? Maybe that's extra server load or something, and who knows, maybe that could trigger a rate-limit if the system is under strain? But that didn't seem right, since there are also regular old text posts being hidden.

I couldn't explain this, however:

For some reason, my mutuals can see the post just fine, but those without a follow/following relationship with me can't. So clearly there's some connection between an existing follow/following relationship and a reply's visibility.

More strangeness: someone else pointed out that the post is hidden for them on bsky.app, but *visible* on deer.social. I signed into deer.social with my alt, and yep - the post is visible.

Uhhhhh - so what in the hell is going on?

Protocols not Platforms? No, Caesars!

Well, a few days later, a massive clue appeared in the form of a blog post from the Bluesky team. [7] A lot of people focused on the 'dislikes' bit, but this wasn't the only part that caught my attention.

The blog post also notes they've been

"testing a mix of ranking updates, design changes, and new feedback tools — all aimed at improving the quality of conversation and giving people more control over their experience", and "developing a system that maps the 'social neighborhoods' that naturally form on Bluesky — the people you already interact with or would likely enjoy knowing. By prioritizing replies from people closer to your neighborhood, we can make conversations feel more relevant, familiar, and less prone to misunderstandings."

The post ends by saying,

"Over the next few months, we’ll keep refining these systems and measuring their impact on how people experience Bluesky. Some experiments will stick, others will evolve, and we’ll share what we learn along the way."

Now sure, it's still possible this isn't the reason why a huge number of users are noticing what amounts to the automated suppression of political speech at a critical moment in Canadian politics. [8] But it all fits!

Mass Confusion

And I'm not saying the Bluesky team is *intending* to suppress political speech with these changes - although to be clear, that's absolutely one of the effects.

What I *am* saying is that it looks a whole lot like they've been secretly and intentionally performing experiments on their userbase, deliberately bypassing the 'stackable, composable moderation' [9] that was supposed to be the selling point of this network - a network "without Caesars" [10] - and these experiments are having the side effect of suppressing political speech. Can you really say people have "algorithmic choice" [11] if they can't see how the algorithms work? Hey, Mike Masnick, Mr. Protocols Not Platforms: anything to say about all this? You're on the company's board! [12]

* * *

Meanwhile, no one seems to know exactly what's going on, and it feels like there's a lot of bad or incomplete information being thrown about.

The official bsky.app account, for instance, reposted a 30k-follower account confidently assuring everyone that the 'dislikes' functionality is merely to allow you to "see less of topics you're not interested in on certain feeds". [13] Yet Bluesky's blog post itself makes it clear your 'dislikes' will also affect what others see: "Dislikes are private and the signal isn’t global — it mainly affects your own experience and, to an extent, others in your social neighborhood." [14]

Then a well-intentioned (and to be fair, *partially* correct) thread went viral, explaining that the people and posts missing from your feed are entirely a result of over-zealous labelling. [15] It's true that turning off some of the moderation settings may well make some things more visible - but it's far from the whole story here, and it may give people the false impression that they're still more or less in control of what they see on this platform.

* * *

These algorithmic tweaks are causing significant problems. There has been confusion amongst users who - like the folks who worried that Canada's PM was suppressing their speech - assumed their replies were being hidden by the person making the original post, and were annoyed at the OP - and/or the OP was worried that people would think they had hidden the replies. Either way, it's causing unnecessary social friction and paranoia. [16]

A search of the network for the phrase "hidden replies" reveals users' serious concerns about their personal safety; some folks are in personal or professional situations where they need to reliably be able to know if someone is making threats against them. [17,18] This is not at all a trivial thing, and there should be alarms blaring at full fucking volume in the Trust & Safety department because of this.

Others have pointed out that by hiding innocuous replies behind a button, the platform is training users to click that button every time, and potentially expose themselves to content they *don't* want to see, even if it's just people they've chosen to mute. In other words it's making other features - ones which users understand the mechanics of - useless. This seems like bad design, but what do I know!

Chicken Caesars

It's amazing to me that the Bluesky team apparently didn't consider that manipulating the visibility of people's replies with an opaque and unpredictable AI system would have serious unintended consequences. And after digging through posts & replies by some of the Bluesky team, you have to wonder if they've been deliberately trying to avoid owning up to what they're doing:

• Back on Oct. 24, Paul Frazee quote-posted Bluesky Head of Product Alex Benzer's post, about a bug that caused some accounts and posts to be missing in search. Frazee's quote-post helpfully added that the problem was with something named, appropriately, 'kafka'. [19,20] Someone in Benzer's replies asked about an uptick in posts being automatically hidden. [21] "We'll take a look", said Benzer - but there's no follow-up since. And although 'kafka' may not be directly involved in whatever the reply-hiding system is, I did notice that it seems to primarily interact with a part of Bluesky's moderation system called Osprey [22] which according to Bluesky's Alex Garnett, leverages LLMs and private internal heuristics to perform moderation actions. [23] Interesting!

• In a reply to a user's question about hidden replies on Oct. 28, Bryan Newbold mentioned that "the 'hide [sic] more replies' section ... is not driven by the regular moderation labeling system", but didn't follow up or elaborate. [24]

• Seemingly contradicting Newbold, on October 31st Frazee replied to a now-unavailable post, saying that whatever the OP was referring to "wasn't hidden by any ranking system we're running. The person you're replying to might be on a mutelist for you?" [25] To which other users responded that, um, loads of people are noticing replies being hidden for no apparent reason, so what the heck, dude.

• And in a leaflet post on Nov. 3, Frazee refers to changes outlined in the Oct. 31 blogpost in the future tense, as "coming updates", saying "our inclination is to shape the core experience of replies to anchor more closely toward your social cluster". [26] Ok but Paul, it's pretty obvious you've already been rolling a bunch of this stuff out to unsuspecting and non-consenting users - are you worried people will be upset if you admit to secretly running social experiments on them?

Good News and Bad News

The good news is, there are projects like BlackSky and NorthSky that aim to run alternative implementations of the full AT Protocol, so people will have the option to use infrastructure and moderation tools run by folks who *don't* treat their userbase with utter contempt. A few days ago, Rudy Fraser confirmed that Blacksky has its own AppView, and although it doesn't appear to be fully deployed yet, it sounds like it's close to production ready. [27] Also, for whatever reason, using the deer.social front-end seems to negate the problems with hidden replies, at least for now.

The bad news is, most people probably aren't going to switch providers or front-ends any time soon. So even if *you* move, these experiments will still be affecting most of the people you interact with - and will therefore still affect you.

Sort of like how defaulting people to the "Discover" feed instead of the "Following" feed will end up (via comments and reposts) skewing *everyone's* feed (even people who exclusively use the Following feed) toward whatever Bluesky wants people to Discover, and away from what they don't.

By the way, does everybody know that you can *delete* the Discover feed from the app? Just putting that out there.

* * *

Even on a decentralized social network, decisions made by one major entity have a huge effect on the others. And we know from the last two decades that the choices made by social media operators can have huge and devastating consequences outside the confines of the platform. The Bluesky team need to start listening to users' concerns, accept criticism, and be open and honest about what they're working on. They need to stop treating their platform like a social science experiment where they silently manipulate people's social environment while acting like they alone know what's best.

If recent history is any guide: this bunch are not capable of doing that. Too much ego, too little self-awareness, too scared to self-reflect. Chicken Caesars. Maybe I'll be proven wrong, but I'm not holding my breath.