Can AI Predict Who You'll Block On Bluesky? New Research

A new study looked at millions of block events on Bluesky, the decentralized social network, to ask a simple question: can we predict who someone is going to block before they do it? The researchers analyzed over 3 million blocks and 260 million user interactions. Most people doesn't block strangers they've never interacted with—but the study found that only about 5% of blocks were preceded by any recent direct interaction, like a reply or a follow. Most blocks happen for other reasons.

To improve predictions, the researchers looked beyond direct contact. If a user and a potential target both interacted with the same third person—a kind of friend-of-a-friend connection—that accounted for about 22% of blocks before they happened, as long as very popular accounts were left out of the chain. Using these signals, the best predictive tools were right between 61% and 72% of the time, far better than a random guess at around 17%. In other words, the pattern of who you interact with can reveal a lot about whom you'll want to avoid.

However, the study has an honest catch: accuracy depends heavily on how the problem is framed. The "candidates" you ask the system to choose from—who it considers as possible block targets—change the results a great deal. That means building a real product is not just a technical challenge but a design choice. For everyday users, the research suggests future moderation tools could quietly warn, "You might want to avoid this account," before harmful interactions happen.

It also highlights a deeper social media truth: blocking is rarely about a single fight. It's often shaped by the wider web of friends, followers, and shared connections. As platforms become more serious about safety and mental health, tools that predict distress before it shows up could be the next privacy battleground—and the next everyday feature that makes one platform feel safer than another.