Stream

We wrote up how we built moderation for live video and audio. getstream.io/blog/live-call-…

Your blocklist is the worst moderator on your team. It has no context, it never learns, and it treats every instance of a word identically. “I will kill you” and “you killed the presentation” are the same event to a blocklist. It’s also the first thing most teams reach for, because it’s the one part of moderation that feels completely under your control. I’m not saying turn it off. I’m saying go read what yours actually caught last week. Most people never have.

Moderating live video means accepting you can’t undo anything. Chat has a queue. You can hold a message, score it, remove it before anyone reads it. Live video has no queue. It has already been seen. So the question changes. Not “can I stop this frame” but “can I stop the next one.” That’s why enforcement acts on the participant. Warn, then blur or mute, then remove them from the room. The content you can’t take back. The next ten minutes you can.

Five weeks to Black Friday. Here’s what I’d change in your moderation config before then. 🔵 Set flood and rate limits against your peak minute, not your daily average. 🔵 Check any rule keyed to account age, because a lot of your December users will be brand new. 🔵 Make sure something counts flags across messages per account, not just per message. 🔵 Default gray-area categories to limited reach rather than removal while your team is thin. 🔵 Staff the review queue to the hours after the event, not during it. The whole list is an afternoon of work, and every item is harder to do in December.

Sharing some advice as a trust & safety leader: Don't staff up for a big event like it’s a different animal. Case in point: the first two World Series games had 45% more flagged content. But? The harm mix didn’t move at all. Racism was at 10% both nights + on a day with no game. Harm typically looks the same on big nights, you just get more of it. Build the config that fits your Tuesday 🙃

Indiana beat Michigan in overtime on Saturday. I’m a Hoosier, so I spent most of it with my hands over my face. 🙈 Look at the chat. Nobody was being abused.... they were reacting to the game. “This team is ass” and “you are ass” look the same to a keyword filter. But to a person, they’re completely different. If your moderation can’t tell them apart, you have a context problem.

Stop stitching Here's how most platforms end up moderating... One vendor for text. Another for images. Another for video. A separate tool for the review queue. Four contracts. Four dashboards. Zero shared context. At Stream, text, images, video, audio, usernames, and livestreams all run through one engine and land in one queue. Your mods deserve one place to work 🙃

We published our moderation benchmark. Latency and throughput measured on live production traffic over seven days, running customers’ own policies rather than a test set. 35ms median on text. 300ms to score a video frame. Over 1,500 requests a second at peak. getstream.io/blog/moderation…

Bounce. Don't ban. Unpopular opinion: most people who post something nasty aren't nasty people. They're fans who just watched a brutal call at the buzzer 🫠

I’d bet you think marketplace moderation is mostly about catching scams and abuse. The numbers say otherwise. Here’s what one marketplace’s flagged messages broke down to last month: Shared a messaging app 45% Everything else 41% Shared a phone number 8% Shared an email 5% Impersonation 2% 57% of everything flagged was one behavior, which is two people trying to take the deal off the platform. That surprised me until I thought about what a marketplace is actually protecting. Not the conversation. The fee. Buyer protection. The dispute record. The review. All of it disappears the moment someone says “just WhatsApp me.” Profanity and insults were 0.1% of what got flagged. If you bought moderation to catch bad words, you’re covering the part that doesn’t cost you anything 🙂

Here’s your surprising moderation stat for the day... When an account is banned, 53% of the time it happens within an hour of that account’s very first flag. For everyone else, the 90th percentile is 87 hours, which is almost four days. There’s almost nothing in between, so you either catch someone immediately or you don’t catch them until Thursday. My guess is that most escalation configs are optimized for fast responses but don't handle slower cases well.

Don't feed the trolls. Hide them. The best troll strategy I’ve seen of is one they never notice. Breaking it down in the thread 🧵

I’d bet you think nudity is the #1 flag in video moderation. But, the numbers tell a different story. Here's what one customer's flagged keyframes actually broke down to last month: Threats 38% Hate speech 38% Sexual content 18% Self-harm 4% That surprised me until I thought about what's actually on screen in a live stream. Text overlays. Signs. Gestures. Things people hold up to the camera. If you bought video moderation to catch nudity, you only solved 18% of your problem 😊