This is super helpful feedback! And obviously not good. Although we are using a third party classifier for the weather report, I’ll chat with my collaborators about ways for us to adjust this. Thanks Tea!

Thanks Ian. The goal here isn’t to automate anything. And unchecked hate speech has often hurt marginalized ppl more than the insulated majority. The hope here is to reduce harm.

And the tool doesn’t alter your post (unless you use the rewrite tool as an assist); it doesn’t stop you from posting.

The goal is to provide a visible indicator of how content is often scored. It’s not an enforcement tool. It’s making visible classification that often happens on the backend. There’s a lot of critique of downranking that happens after posting; this is a pre-post indicator that you can take or leave

We/I don’t own this model but all classification models have errors and fall short. It’s one of the reasons who chose a broad range of weather scores that are directional and not a binary, to allow the user to take it or leave it.

Hey Karl. The score is not just word based — it’s a tone classifier more broadly.

That’s a really valid critique. I appreciate you bringing it up.

Hey Tea. The tone classifier it not analyzing the political position or accuracy of the words but their overall tone, so if a word that is negative is included the weather score will change. Thanks for playing around with it! Appreciate the feedback.

Having worked on moderation @ large platforms I’m painfully aware that moderation at scale gets things wrong. And AI by nature is biased. The idea here was to make a common scoring framework visible as a signal that authors can use (or not!), a response to the critique of the invisible “hand” of mod

Hey All. Thanks for your input. @clrskyai.bsky.social is a project I am testing as part of a research project so your input is welcomed. The weather scores model is based on the # of ppl out of 10 who might find something offensive. The highest score = 9/10. A lot of moderation is hidden. This isn’t

Hey Erica! Great question and one we should always ask of both AI and moderation systems. In this case, the classification is defined by Perspective API, which is scored on what X/10 ppl surveyed believed is toxic in nature. I don’t know who was included in that survey pool, so representation is key

It is current 'written over', although we are working on a 'undo' button so you can revert to the original if you prefer it. No storage of anything you type.

@christiannaj.bsky.social I think you might find this project interesting - making a common invisible content moderation tool visible to folks on Bluesky through an overlay called CLR:SKY. www.clrsky.ai