Varun Srinivasan
@v
Warpcast Spam labels are improving Our new system uses LLMs to make better decisions. It will make fewer mistakes in both directions and will prevent legitimate users from getting the spam label. We expect fewer users to get the spam label as a result of this change. It may take a few weeks to roll out fully.
27 replies
42 recasts
260 reactions
Varun Srinivasan
@v
What’s changing? - labels are now binary instead of a three point scale. They will be either “likely to be spammy” or “unlikely to be spammy” - our new models use llms to look at profiles and check for inauthentic behavior - they also factor in verifications and many new signals to get more accurate
4 replies
7 recasts
83 reactions
Varun Srinivasan
@v
What makes a user spammy? There are many signals of inauthentic activity. Doing one or two of these will not get you the spam label, but doing many will. - profile that is clearly automatically generated with no effort put in - posting content from the internet and pretending it is yours - using ai to automatically post messages for you - replying and mentioning many users who never interact with you or follow you - posting and resharing repetitive content, just like a bot might - operating in rings, where accounts coordinate to like or boost certain accounts - casting, following, liking or resharing things in an automated fashion to artificially get others to follow you - impersonating well known accounts on the network - getting muted or blocked by many other users on the network
3 replies
2 recasts
41 reactions
Olimpius
@olimpiusinferno
Thanks for the transparency — the changes sound solid. A few follow-up questions: Will the label be dynamic and update over time if behavior changes? Will users be able to appeal if they feel they’ve been mislabeled? Who reviews or audits these decisions to ensure fairness and prevent bias? Building trust in the system will be key, especially as it scales.
0 reply
0 recast
0 reaction