Abstract:
On September 1st, Douyin experienced a large-scale abnormality in its recommendation service. A large number of users reported that the recommendation flow on the homepage seriously deviated from their personal interests, and was concentrated on pushing AI short dramas, health care, parenting, movie commentary and other content that had nothing to do with their preferences, as if they had "stuck into a middle-aged and elderly channel." What confuses users even more is that even if they press and hold the video repeatedly and click "Not interested", similar low-quality content still appears repeatedly. Some users even encounter basic functional abnormalities such as empty popular lists and loading lags.
Relevant terms such as “Douyin recommendation is ugly”, “Douyin mistakenly enters the channel for middle-aged and elderly people” quickly became hot searches on Weibo, triggering heated discussions. In response to this, Douyin customer service responded to Pear Video by saying that it has received user feedback and the platform attaches great importance to it and is currently investigating the issue urgently. Please wait patiently for the specific reasons. Users can try to restart the APP to recover.
Douyin customer service response Image source: Li Video
According to feedback from many users on social media, this recommendation anomaly first appeared on the evening of August 31 and has not fully recovered until this morning. A common feeling among affected users is that the daily personalized recommendation flow based on likes, viewing records and interest tags suddenly lost its memory and was replaced by a large amount of homogeneous and low-quality pan-entertainment content.
A young user complained that he usually pays attention to technology, finance and digital content, but all he saw in one night were health care, emotional quotes from middle-aged and elderly people, and AI-generated short dramas. "It feels like my dad has taken away my phone." Some users reported that the recommendation stream seemed to be distributed randomly, and the content they received had almost no correlation with their interests. There were even cases where the same video was pushed continuously.
The AI short drama content that users have complained about this time is one of the fastest growing content categories on the current short video platform. Data shows that domestic AI short drama users have exceeded 600 million in the first half of this year, and the number of new AI dramas and comics native to Douyin has reached 221,900, with a cumulative playback volume of 515.738 billion times. At the same time, a large number of shoddy AI short dramas rely on platform traffic distribution to gain exposure. Once the recommendation algorithm deviates, this type of content will easily flood into users' vision.
On June 15 this year, Douyin adjusted its AI video traffic distribution rules, established an exclusive traffic pool for compliant AI content, granted additional natural traffic to creators who had done a good job in explicitly labeling AI traces, and restricted the flow of low-quality AI videos that applied templates in batches. However, the speed of governance obviously failed to keep up with the expansion of supply.
For a national-level application like Douyin, even a short-term local abnormality in the recommendation algorithm may affect the experience of tens of millions of users. The "China Mobile Internet Strength and Value List for the First Half of 2026" released by QuestMobile shows that the number of monthly active users of Douyin's main APP has reached 1.029 billion, making it the second domestic super application to reach one billion monthly active users after WeChat. It has about 830 million daily active users, an average daily usage time of nearly 2 hours, and an average daily video playback of more than 20 billion times.
In the face of a huge number of users, any local adjustments to Douyin’s recommendation algorithm may trigger a chain reaction. According to the algorithm principles disclosed by Douyin’s Safety and Trust Center, the core logic of its recommendation system is to comprehensively predict user behavior feedback—including positive and negative signals such as completion, likes, comments, collections, sharing, and “not interested”—to push videos with the highest feedback value to matching users. The entire process involves multiple links such as behavioral probability prediction, value model evaluation, and minute-level real-time feedback.
At the end of 2025, Douyin completed the largest upgrade of its recommendation system in recent years, from the previous single-tower content model to the "creator plus content" dual-tower model. The system not only judges the quality of a single piece of content, but also evaluates the comprehensive weight of the creator, including multiple dimensions such as originality rate, fan stay time, depth of interaction, content quality records, and fan return rate. In the daily operation of such a complex system, any model update, parameter adjustment or server cluster failure may cause the recommendation flow of some users to deviate.
Comments