On 30 September 2026 Aaron Rodericks, Bluesky's Head of Trust and Safety, announced that a bot now blocks people on his own account. He says it judges only what you write to him and never who you are. Blocks on Bluesky are public records, so that claim can be checked against the list itself.
“These do not factor in your bio or your identity. I have zero clue who you are. The only thing it calculates for the auto-block is does your specific interaction deserve a block.”
Most of the blocks did not come from a bot reacting to posts. They arrived in three bulk loads on the morning of 30 September, all before the public announcement. He described this himself as “a seed set of recent abuse”. Only blocks came in one at a time.
The comparison group is everyone who replied to him or quoted him and was not blocked. The measure is what each account says about itself in its bio or display name, read the same way for both groups.
These are floors. Many trans people do not say so in a bio, and a flag alone is used by allies too. Thin lines show the range the true share plausibly sits in.
A fair objection is that trans people may simply have been ruder to him. So every reply and quote from both groups was scored for toxicity, and accounts were sorted by the harshest thing they wrote. If only the words mattered, the two dots on each row would sit on top of each other.
Harshness matters. Blocked accounts wrote much nastier posts on average: . But it does not close the gap, and it is far from a rule: of blocked accounts never sent him anything harsh, while accounts that did were left alone.
Each square is one of the accounts that replied to him or quoted him since January. Filled squares were blocked. Change what is highlighted and how the accounts are split to look for yourself.
He calls the system “entirely opt-in”: you are blocked because of how you interacted with him. Each blocked account's own posts were read back to 1 January to see whether it had replied to, quoted or @-mentioned him before the block landed.
The first load is people who spoke to him. The third is mostly people who did not, which fits his description of “an AI query to look for patterns of abuse across the site”. The trans share is about the same whether an account had addressed him or not (). accounts had already blocked him before he blocked them.
“We reached the moment when Aaron outsourced his judgement to a bot. That is a judgement on his judgement.”
Blocks were read from the app.bsky.graph.block records in his own repository. Contact was read from each blocked account's own app.bsky.feed.post records back to 1 Jan 2026, because the public view hides replies across a block. The unblocked group comes from the public reply threads and quote lists of his posts since 1 Jan 2026. Self-identification is a fixed set of patterns over bio and display name, with ally phrases excluded; a hand check of 25 matches found about 2 wrong. Harshness is the highest toxicity score from the unbiased-toxic-roberta classifier across an account's replies and quotes, banded as very mild (under 0.02), mild (0.02–0.1), moderate (0.1–0.5) and harsh (over 0.5). The odds figure comes from a logistic model that also accounts for swearing, threat words, number of posts and quote versus reply. Ranges are 95% intervals.