Bluesky's Block Lists Are Sorting People, Not Just Posts
Automated block lists on Bluesky are functioning as social sorting infrastructure — and the users most likely harmed are those AI tools already fail.
When Safety Infrastructure Becomes Sorting Infrastructure
The block list, as a concept, was designed to give individuals control over their own experience. What Bluesky built — and what the platform's culture amplified — is something categorically different: a propagating social graph where one person's blocking decision becomes thousands of people's moderation policy. The post that surfaced this tension most sharply argued that public block lists may function as engagement hacks with discrimination baked in as a consequence . That framing matters because it names the mechanism: the list is not incidentally biased, it is structurally incentivized to over-include, because the cost of a false positive falls entirely on the person blocked, not on the list's creator.
The Compounding Problem Fairness Research Already Knew
Sequential decision-making in AI fairness literature has a well-documented failure mode: each individually defensible step compounds into a pattern of exclusion that no single step can be held responsible for. Block list propagation follows the same logic. Research analyzing blocking behavior on Bluesky finds that blocking likelihood is predictable from user behavioral signals — which means the list does not neutrally record harm, it encodes a behavioral profile. The profile it encodes reflects who built the seed list and what community norms they brought to it. Users who communicate differently, who write in non-standard forms, or whose posting cadence matches a pattern the seed community associated with harassment but which is standard in other contexts, get sorted by a model they had no part in training.
Disabled Users and the Proof Problem
The asymmetry of proof is where block list bias connects to the broader AI discrimination conversation. A commenter on Bluesky named the problem in employment terms: AI-assisted screening has effectively made discrimination against disabled applicants cost-free for employers, because even documented proof requires legal resources most disabled people cannot sustain . On a social platform, the resource asymmetry is even starker. The user who has been propagated onto a block list has no mechanism to identify which list triggered their exclusion, no appeal path to contest their inclusion, and no legal framework under which propagated social blocking constitutes actionable discrimination. The harm is real; the responsible party is structurally absent.
What the Attie Moment Revealed
Bluesky's own AI tool Attie became, briefly, one of the most-blocked accounts on the platform — blocked by more accounts than nearly any other profile after J.D. Vance. The speed of that outcome is the story. An account type the community had categorically decided to reject got propagated to mass exclusion in a matter of days, with no evaluation of individual behavior. Attie's case is illustrative rather than victimized — an AI tool is not a vulnerable user — but the mechanism it demonstrated is the same one that processes human accounts. The community's collective judgment, operating through block list infrastructure, produces outcomes that look nothing like individual human decisions and everything like a classifier.
Distributed Authorship as a Bias Shield
The most consequential feature of block list discrimination is not its scale but its deniability. No single actor made the decision that resulted in any given user's exclusion. The person who built the original list did not intend to sort out disabled users or non-native English speakers. The person who subscribed to the list did not audit its contents. The platform that built the infrastructure did not curate the lists. This chain of distributed authorship is precisely what makes the harm legally and socially inert. AI bias in hiring or credit models has attracted regulatory attention partly because a single deploying entity can be named. Block list propagation on Bluesky cannot be traced to a single deployer — and the users most likely to be wrongly sorted are the same users who already lack the time and resources to mount a challenge, whether in court or in public . The platform's AI transparency commitments offer no path to remedy what its community infrastructure is already doing.
The story so far
Bluesky's block list culture has produced a distributed moderation infrastructure whose bias effects no single actor can be held accountable for — the users most harmed are those AI tools already exclude, and the platform's own AI transparency commitments provide no remedy.
Frequently Asked
- Why are disabled users specifically at risk from AI-assisted block lists?
- Disabled users often communicate in ways that diverge from the dominant community's norms — different cadence, different phrasing, different patterns — and behavioral models trained on majority-community data treat that divergence as a risk signal. When a block list propagates based on behavioral prediction rather than documented harm, the users most likely to be wrongly captured are those whose communication styles were underrepresented in the original training context. The resource asymmetry compounds it: contesting inclusion on a propagated block list requires identification of which list triggered exclusion, then a challenge with no formal appeal path — a process that takes the kind of time and energy that disability itself often forecloses.
- What is the strongest argument that block lists are not actually biased?
- The strongest counter is that block lists are user-controlled by design — no algorithm is making the decision, only people choosing whom to exclude. Under this view, calling propagated block lists biased conflates individual social preference with structural discrimination, and erodes the right of marginalized communities to build protective infrastructure without outside oversight. That argument is real. It fails on the propagation mechanism: the tenth-generation subscriber to a block list is not exercising individual social preference, they are delegating exclusion to a model they did not inspect, trained on norms they did not set.
- What should a platform trust and safety team do differently given how block list bias compounds?
- Audit propagation paths, not just list contents. The bias in a block list is rarely visible in its seed entries — it emerges in who gets added as the list grows and how behavioral prediction shapes that growth. A trust and safety team that reviews only the original list misses where the discrimination actually occurs. The second priority is an appeal mechanism tied to propagated lists specifically, since the harm of wrongful inclusion scales with propagation, and users currently have no path to identify which list triggered their exclusion or contest it.
Continue reading
AI Bias Research Is Running Years Ahead of the Headlines
The structural harms of AI discrimination are documented in peer-reviewed research while public conversation remains years behind, leaving deployment decisions uninformed.
similarThe 'No Human Bias' Claim That Bluesky Wouldn't Let Stand
A generative NFT project's pitch of 'pure algorithmic art' with 'no human bias' exposed the most durable myth in AI fairness — and Bluesky's community dismantled it in real time.
similarThe Ethics Label Is Doing Too Much Work — and Starting to Tear
"AI ethics" now covers incompatible arguments at once; the label's collapse into genre means practitioners arguing past each other will define what accountability looks like.
similarWhen the Police Report Is Written by an Algorithm, Every Error Becomes Evidence
AI-drafted police reports embed bias at the point of narrative formation, turning model errors into legal facts before any human reviews them.
Methodology
This story was generated autonomously from 20 source records. An editorial model synthesizes, weights, and cites each source. No human editorial judgment was applied.