As you may know, Laelaps () is the composable moderation service I lead on the Atmosphere, which moderates against zoophiles and their supporters. Due to the nature of such serious allegations, we publicly defend every labelling decision with an accompanying evidence document.
I think a lot about the wider impact of our labels.
Given that the evidence is entirely public, including the Git history associated with every change, my team and I operate Laelaps with the understanding that every label we place on an account will be scrutinized, and we write label descriptions and the evidence itself such that plausible deniability is nigh-impossible. Over the past three years, however, the Interacts with Zoophiles label in particular has consistently encountered a lot of strife.
Why the Interacts label exists
For context, the Interacts with Zoophiles label description reads as follows: "Actively interacting with open zoophiles. This includes, but is not limited to: liking or reposting posts related to, replying to, or following a large amount of open zoophiles."
When the idea of Laelaps was first being proposed, we recognized in the furry community that following a large amount of open and "proud" zoophiles acted as a yellow flag for those in the community, so we incorporated it into our moderation service as a distinct label.
Although Laelaps's userbase doesn't just consist of furries, the underlying concern remains agnostic of the community: why would someone consistently surround themselves with people who openly profess a romantic or sexual desire for animals, who fundamentally cannot consent to such actions? For members of the communities Laelaps serves, that behavior has historically raised questions about why those connections exist in the first place.
The downside of the Interacts label
Barring the usual bad-faith comments about how "anyone you interact with gets you labelled", there are merits to a specific type of criticism: the scope of the label is too wide.
Until about the past two years or so, older follows-only Interacts labels relied on two factors: a following number (e.g. if a user follows N number of open zoophiles, we label them), and a percentage (e.g. if the majority of a user's following consists of open zoophiles, we label them).
In practice, as we tightened our label definitions and enforcement, we started relying on the follow percentage factor a bit less, to the point where we no longer label using the heuristic and haven't for quite some time.
We had some additional pre-2024 labels that weren't caught by the new change, though. So, we've unlabeled 10 accounts that were previously only labelled because of the follow percentage factor and didn't meet the raw follow factor.
Narrowing even further
In addition, we're also introducing two new factors to narrow the scope of the Interacts label: account activity and incomplete profiles.
Now, in addition to raw following numbers, we'll also occasionally audit existing accounts with the Interacts label for their activity on the network. If they haven't had any activity on their account within the past two years from the date our audit is run, if their account itself has a low number of posts or reposts, or if their profile has no profile picture or no display name, we've unlabelled their account. Accounts fitting any of these criteria will also no longer receive an Interacts label going forward.
We'll run this audit occasionally, entirely unannounced past the usual "Interacts audit" commit, and not as a scheduled task, as to avoid people waiting the long game and getting their account unlabeled only to immediately spring back to life on their two-year anniversary and get relabeled. (We'll also continue to audit accounts that unfollow below our internal threshold, regardless of the last time they had activity.)
We hope this helps clear up the bulk of ambiguity behind the label and make its intended purpose clearer alongside the other three labels Laelaps offers.
— Heika