Inside TikTok's transparency center: AI catches 97% of removed videos

SINGAPORE — Before a TikTok video reaches the recommendation feed, machines get the first look, scanning its images, sounds and other signals for potential violations of the platform’s community guidelines.

In the first quarter of this year, automated systems detected and removed 178 million, or 96.7 percent, of the 184 million videos taken down globally — before they were reviewed by human moderators. Overall, 99.3 percent of removed videos were taken down before users reported them, according to the platform.

At TikTok’s Transparency and Accountability Center in Singapore, a demonstration shows how machine-learning systems scan uploaded videos for signals associated with potential violations, such as smoking, drinking, physical violence and extremist symbols, even before human moderators step in to review cases requiring greater context or judgment.

The systems use more than visual recognition to also analyze audio, keywords and hyperlinks. The technology can also recognize body posture and actions, allowing it to combine multiple signals — such as a cigarette and a person’s hand position — to identify potentially prohibited activity.

In Korea, the company removed over 254,000 videos in the first quarter, or 0.3 percent of locally uploaded videos. About 57.4 percent were removed with zero views, while 84.4 percent were taken down after receiving 100 views or fewer. Automated systems accounted for about 87 percent of removals.

TikTok said its systems automatically remove content only when they are highly confident that it violates the platform’s rules. More ambiguous cases, including misinformation, bullying and content requiring cultural or linguistic context, are referred to human moderators, with thousands of specialists handling such cases globally.

For example, its systems can flag content showing smoking or drinking, but may struggle to distinguish a bottle of kombucha from alcohol or a religious symbol from an extremist logo. Such cases require human judgment, the company said.

For users whose content is removed, the platform provides a notification and an appeal process. Globally, 8.8 million videos removed in the first quarter were restored following further review, equivalent to about 5 percent of all videos taken down during the period.

TikTok’s community guidelines are applied globally, but moderators use regional playbooks to account for local laws, languages and cultural contexts. In Korea, local staff with knowledge of the Korean language and culture review content, while local experts provide feedback on emerging risks and nuanced expressions.

Content moderation goes beyond removing violating videos, playing a role in what appears on the platform’s recommendation feed, where most users spend their time.

Once content clears the moderation process, TikTok’s recommendation system matches videos with users based on signals such as watch time, likes, comments and follows, as well as captions, hashtags, sounds, language settings and approximate location.

TikTok also tries to keep users from being repeatedly shown similar content. The system limits videos from the same creator and adjusts recommendations when a feed becomes overly focused on one type of content, helping users see a wider range of content.

Users can also reset their recommendation feed using the "refresh feed" feature, which clears their existing recommendations and allows TikTok to build a new feed based on their subsequent interactions.

The platform is also tightening controls around artificial intelligence (AI)-generated content. Realistic AI-generated or edited content must carry a label, while harmful or misleading AI-generated content is subject to removal under the company’s guidelines.

TikTok said it has applied AI labels to more than 3 billion videos using technologies including the Coalition for Content Provenance and Authenticity standard, creator labeling tools and invisible watermarking.

The company noted it is targeting accounts that mass produce low-quality AI-generated content, rather than treating the use of AI itself as a violation. It plans to strengthen detection of accounts that repeatedly post AI-generated content in sensitive areas such as politics, finance and health, particularly when the content is produced at scale with little human input.