Automated moderation might be getting it right, then human judgment complicates that
TikTok's automated moderation systems caught 98% of guideline-violating content before users saw it, outperforming Meta (54%) and X (24%) in removal rates. Antisemitic content has surged on TikTok since October 7, 2023; the platform explicitly prohibits Holocaust denial, hateful conspiracies, and supremacist claims.
- TikTok removed 98% of harmful content proactively in Q1 2025 in Australia
- Over 336 million videos posted to TikTok globally in Q1 2025; 270,000+ removed under safety guidelines
- Antisemitic content surged on TikTok after October 7, 2023 Hamas attack on Israel
- TikTok removed 64% of reported offensive videos; Meta removed 54%; X removed 24%
TikTok reported removing 98% of harmful content proactively before public viewing during Q1 2025 in Australia, with over 270,000 videos removed from 336 million posted. The disclosure came before a Royal Commission examining online antisemitism and social cohesion.
TikTok's head of policy, trust and safety walked into an Australian courtroom this week with a number that sounded almost too clean: 98 percent. That's the share of harmful content the platform says it removes automatically before anyone in Australia ever sees it. The figure came during testimony before the Royal Commission on Antisemitism and Social Cohesion, an inquiry examining how hate spreads across digital platforms and what companies are actually doing to stop it.
Zachary Hecht, who traveled from the United States to present TikTok's case, described a system where every video uploaded to the platform gets scanned by automated moderation tools the moment it arrives. The company's community guidelines explicitly ban antisemitic material—Holocaust denial, conspiracy theories targeting Jewish people, claims of ethnic supremacy, and the practice of blaming entire groups for individual actions. When violations are detected, the content gets pulled before it can accumulate views or spread through the algorithm.
The scale is staggering. In the first quarter of 2025 alone, more than 336 million videos were posted to TikTok globally. In Australia, the platform removed over 270,000 of them under its safety and civility rules. The 98 percent figure applies specifically to content flagged by the platform's own systems rather than user reports—a distinction that matters because it speaks to the speed and reach of automated enforcement.
Context matters here. The Royal Commission has already heard testimony about a sharp rise in antisemitic content on TikTok since October 7, 2023, when Hamas attacked Israel, killing 1,200 people and taking hostages, which then triggered an Israeli military campaign in Gaza. The platform became a vector for hate speech during a period of intense global conflict, and regulators wanted to know what was being done about it. TikTok's numbers, if accurate, suggest the company's automated systems are working at scale.
But the testimony also revealed the messiness of content moderation in practice. The commission heard a case study that exposed real tension between machines and humans. An AI system correctly flagged a piece of antisemitic content as a violation and hid it from public view. The creator appealed. A human moderator reviewed the same content and decided it was fine—it didn't violate policy. When the Dor Foundation, an antisemitism advocacy group, escalated the video for further review, another human moderator again approved it. Only after additional escalation did the content finally get removed, with Hecht acknowledging it "absolutely violates our policies." He framed the incident as a learning moment, noting that automated moderation sometimes gets it right while human judgment can complicate the picture, and that reviewers needed better training.
The performance matters partly because of how TikTok stacks up against competitors. The Online Hate Prevention Institute analyzed removal rates across platforms. In a sample of more than 400 videos reported to TikTok as offensive, the platform removed 64 percent. Meta removed 54 percent of roughly 950 Facebook posts flagged as problematic. X removed 24 percent of approximately 1,700 reported posts. By that measure, TikTok's responsiveness to user reports outpaced its rivals, though none of these numbers are perfect.
The Royal Commission's work is far from finished. YouTube, owned by Google, is scheduled to testify next. Australia's public broadcasters are due to appear later in the week, and the ABC has already submitted written evidence to the inquiry. What emerges from these hearings will likely shape how regulators think about platform accountability and whether the current mix of automated and human moderation can actually keep pace with the volume and velocity of online hate.
Notable Quotes
Our focus is on getting content moderation right at scale for billions of pieces of content.— Zachary Hecht, TikTok's global head of policy, trust and safety
In some instances automated moderation might be getting the moderation correct and then human judgement can then complicate that.— Zachary Hecht, TikTok
The Hearth Conversation Another angle on the story
When TikTok says it removes 98 percent proactively, what does that actually mean in practice? Is the platform catching things before they spread, or just before they're reported?
It's the former—the platform's own automated systems are scanning every upload and pulling violations before users ever see them. That's different from waiting for someone to report something. The speed matters enormously at that scale.
But the case study they presented—where the AI flagged something correctly and humans overruled it—that seems to undermine the whole story, doesn't it?
It does expose a real problem. The AI got it right. Humans got it wrong twice. That suggests the bottleneck isn't the technology; it's the training and judgment of the people doing secondary review.
Why would a human moderator approve antisemitic content that an automated system had already flagged as a violation?
The testimony doesn't explain that directly. But it hints at it—Hecht mentioned that human judgment can "complicate" what the machines get right. That could mean inconsistent training, fatigue, cultural context being lost in translation, or just the sheer volume making careful review impossible.
So is TikTok actually better at this than Meta or X, or is it just better at talking about it?
The numbers suggest it's genuinely better at responding to user reports—64 percent versus 54 and 24 percent. But the 98 percent figure is about proactive removal, which is a different metric. You can't directly compare them. What you can say is that TikTok's automated systems are catching most violations before they're seen, and when users do report things, the platform acts faster than its competitors.
What happens now? Does this testimony change anything?
That depends on what the Royal Commission decides to recommend. Right now, TikTok is making its case that it's taking the problem seriously and investing in the infrastructure to handle it. But one case where the system failed—where humans overruled the machines and approved hate speech—is enough to show that scale and automation alone don't solve the problem.