TikTok, Perez Hilton, and the thirty four minutes nobody can quite explain

What happened to Perez Hilton on Tuesday night is so horrific and genuinely distressing, and I hope he’s safe and getting the care he needs. I don’t feel comfortable putting a label on exactly what this was, that’s not mine to diagnose from the outside, but from what he was reportedly saying throughout, about feeling worthless and not wanting to be here anymore, it’s hard not to wonder whether this was closer to a suicide attempt than a moment of self-harm. I say that hesitantly, but I think even the fact that it’s a fair question to ask says something about how badly platforms like TikTok need to get their act together around moments like this. What’s more answerable is something narrower: how does a livestream like that stay up on TikTok for the best part of thirty four minutes before it’s pulled? And what does that gap actually tell us about how the platform’s moderation works once a camera starts rolling, rather than how TikTok describes it in its own safety literature.

For full transparency, I saw a clip of this myself this morning, completely by accident, and it left me feeling incredibly unsettled. I’ve spent years working in crisis PR, dealing with genuinely difficult situations, and I’m used to having to witness disturbing footage as part of the job. I only managed about ten seconds before I had to stop. From what I saw, and what’s been widely reported since, he was unclothed, covered head to toe in blood, with cuts across his body, holding a very large butcher’s knife, self-harming and saying deeply disturbing things throughout. I really recommend not trying to find the footage, platforms are beginning to have it fully removed now.

And then there’s the children who could have seen it, which I don’t think can be said enough. TikTok’s minimum age is thirteen, and realistically plenty of younger kids are on there anyway, on a parent’s login or with a fake birthday. That’s an adult mid-crisis, and findable for over half an hour on an app where the average teenager is scrolling for hours a day. Even setting aside the original stream, clips of it were circulating within minutes, and there was nothing stopping any of that landing on a child’s feed completely unprompted. That’s not a hypothetical worry. That’s just how these apps works.

So how is TikTok actually meant to catch this stuff

There’s a computer system that scans videos as they’re happening, looking for red flags like nudity, weapons, or violence (this is what’s called automated content moderation, or AI detection). If it’s confident something’s wrong, it pulls the content straight away. If it’s not sure, it gets sent to an actual person to look at, what TikTok calls a human moderator, who makes the final call. On top of that, TikTok also runs software that scans the words underneath, in the comments, looking for language that suggests people are panicking or alarmed (language moderation, sometimes called keyword or crisis-term detection). And if a video or livestream suddenly gets a huge number of views very fast, that’s supposed to trigger a review too, regardless of whether anyone’s actually reported it.

For a normal video someone uploads, all of that checking happens before it goes properly live to everyone. A livestream doesn’t get that same safety net, because there’s no finished video to check first, it’s happening in real time as it’s being filmed. So the systems are trying to catch things as they unfold rather than before anyone sees them, and the human check only kicks in once enough signal, whether that’s reports, comments, or a spike in views, tells someone to actually look at it.

Why I don’t think TikTok gets to hide behind any of that here

This is where I struggle to give TikTok the benefit of the doubt. A man, unclothed, covered head to toe in blood, using a large knife, cutting himself repeatedly, for over half an hour. There’s nothing ambiguous about that. It’s not a video that could be mistaken for a cooking mishap or a costume or anything else a system might reasonably get confused by. If the automated detection genuinely couldn’t catch that fast, it isn’t working for the exact thing it’s supposedly built for.

The comment side is just as bad. You don’t need me to tell you what a comment section looks like under something like this. People panicking. Begging him to stop. Saying outright that they were reporting it. That is exactly the kind of language the keyword detection is meant to pick up on instantly. If that system exists the way TikTok says it does, and it still took over half an hour, that’s not one system failing. That’s two.

Here’s what makes it genuinely ridiculous to me. I’ve had a client’s livestream pulled within minutes because there was a vape visible at the edge of the frame. Nothing to do with the content of the stream itself, just an object sat in the background. Gone almost instantly, no appeal, no waiting around. So TikTok clearly has the tools to act fast and decisively when it wants to. It’ll catch a vape in seconds but somehow missed what looked like a live suicide attempt for over half an hour. There’s no version of that where TikTok gets to say its hands were tied.

It’s also worth saying plainly that this wasn’t some obscure account with a handful of followers, where thin moderation coverage would at least be explainable. Perez Hilton has built a public profile over two decades, with a large, established following on TikTok specifically. If anything, an account that size should sit higher up the priority list for fast intervention, not lower, given the reach a single stream from it can generate in minutes. That it still took multiple viewers going outside the platform entirely, and ringing Miami-Dade police directly, to get a proper response tells you account size isn’t functioning as a fast-track to scrutiny the way you’d assume it would.

That’s not TikTok’s moderation system working. That’s the public doing the job the moderation system should have done far faster than it did.

What would actually close this gap

Better automated detection for self-harm specifically, rather than leaning on general violence classifiers built for other things. A genuinely low threshold for escalating livestreams to human review the moment comment language spikes with alarm, rather than waiting for a report count to build. And proper thought given to how findable this kind of content is to under eighteens in particular, given how much of TikTok’s audience is exactly that age.

None of the explanation above changes where the fault sits for what happened on Tuesday. Understanding how the system is built isn’t a defence of the outcome it produced. On this occasion, TikTok’s systems failed, badly, in front of an audience that included people, and likely children, who should never have been anywhere near that content, and that failure deserves to be named as exactly that.

If you’ve been affected by anything in this piece, Samaritans are available free, any time, on 116 123. In the US, the 988 Suicide and Crisis Lifeline is available by call or text.