The once high-profile AI social track now seems to have fallen into a growth bottleneck. In 2025, the number of downloads and daily active users of mainstream AI social applications in China will decline significantly, and the downloads of the leading products Hoshino and Cat Box will plummet by 80%, and the user retention rate will also face challenges. At the same time, the technical flaws of AI social products, such as character language repetition and experience homogenization, further weaken user stickiness. The “color” turmoil in the industry has attracted regulatory attention, making the future development of AI social networking full of uncertainty. This article will provide an in-depth analysis of the current state of the AI social industry, explore the reasons for its stagnant growth, and the future outlook of practitioners and investors in this field.
“The cat box may not be stable in the position of the byte, at least for the next year.”
After the change of person in charge of the cat box in April this year, the changes in this mainstream AI social application in China are not over.
One of the big intuitive changes is the stagnation of user growth. Since the beginning of this year, the number of major AI social applications and new downloads of several major domestic AI social applications have continued to decline significantly.
Taking the head products Hoshino and Cat Box as an example, selecting the domestic Apple data from January and May 2025 for comparison, Hoshino’s monthly downloads fell from 4.86 million in January to 930,000 in May; The monthly downloads of cat boxes fell from 2.64 million to 610,000. At the DAU level, Hoshino was basically flat, with figures of 960,000 and 970,000 in January and May, respectively, while the cat box experienced a round of decline, falling from 590,000 to 490,000.
The lack of million-level daily active head products and the lack of investment stimulation and increase effect are all pointing to a “top” industry trend.
Recently, a public opinion storm of “color” has brought the AI social industry into the public eye again for the first time in a long time. Tencent Literature’s AI companion software “Dream Island” was recently interviewed by the Internet Information Department due to vulgar content, and the platform was required to rectify it immediately.
To achieve these three challenges, product managers will only continue to appreciate
Good product managers are very scarce, and product managers who understand users, business, and data are still in demand when they go out of the Internet. On the contrary, if you only do simple communication, inefficient execution, and shallow thinking, I am afraid that you will not be able to go through the torrent of the next 3-5 years.
View details >
It is understood that Zhumeng Island quickly updated its version after this incident. When the new version logs in, the app will appear a pop-up window of “age confirmation”, and if you are under 18 years old, the platform will turn on youth mode; those over the age of 18 need to undergo real-name authentication on the platform.
In fact, after entering 2025, the voice of AI social networking in the industry will gradually decrease, and this former “blue ocean track” is no longer a sweet spot for VCs and executives of large companies.
At the same time, the technical strength of AI social products has also been criticized by users. Problems such as repetitive character language, often forgetting dialogue content, blurred AI scale, and homogenized experience are all reasons for the decline in user experience. From the perspective of the platform, the situation that paid conversions fall short of expectations has always plagued the AI social industry.
The gameplay is highly homogeneous, and it is difficult to commercialize monetization, can AI social networking usher in a turning point in the second half?
01
In the past two years, many industry research reports have shown the reasons for being optimistic about AI social networking: providing cost-effective and real-time 1-on-1 companionship for human users through AI virtual characters, providing users with “emotional value”. However, the reality is that it is becoming more and more difficult for leading products to maintain user stickiness.
Entering 2025, the new retention of similar products such as Tsukumu Island and Hoshino will fall below 20% in three days. “Establishing a long-term relationship with users” seems to have become a false proposition.
“In the background, it can be observed that the average time between the user and each character is about 5~7 days, and then they basically won’t chat with this character again.” Zhu Yanan, who has been working in the field of AI social for many years, introduced us to his observations in the industry. Zhu Yanan, who was born in the game industry, has participated in a number of AI social entrepreneurship projects in recent years, and he was a team member of a former executive entrepreneurship project of Byte, which mainly made AI social networks in the sea, and once raised about $300 million. After entering 2025, the project was suspended and disbanded due to satisfactory progress falling short of expectations.
The feedback from the user side fell short of expectations, causing many practitioners to re-examine the rigidity of AI social demand. In the past two years, some companies have entered the market with a very high-profile attitude, such as Linky, an AI social app that went overseas, which set a target of 2 million DAU at the beginning of 2024, but entered a bottleneck period after reaching the order of 500,000.
Zhu Yanan said that many practitioners are discussing that AI social networking seems difficult to become a popular mass track. “In the era of mobile Internet, the needs of most users are basically met by the super apps of the previous era, and this kind of people who like to socialize with Chatbots are actually a relatively niche group.”
On the other hand, the recent reduction of investment by leading companies is somewhat of a “exit”. Two weeks ago, MiniMax high-profile open-source MiniMax-M1 model, claiming to be the world’s first large-scale mixed-attention inference model with open weights. Relevant reports also show that MiniMax is repouring more resources into the field of large model infrastructure. With the implementation of this strategic adjustment, it is not surprising that the investment in the C-end product Hoshino has decreased.
At the same time, as the number one product of MiniMax, Hoshino itself is not having a good time. Some industry insiders revealed that Hoshino Daily’s revenue is about tens of thousands of yuan, and considering its huge team size and user volume, it is difficult to cover the cost at this revenue level.
In terms of bytes, Liang Chenqi, the former head of the cat box, left in April this year, and the position was taken over by Nishihara (nickname), the former head of Star Paint products. It is still unknown whether the cat box after the change of coach will continue the previous operation ideas. In the exchange, Zhu Yanan mentioned that since the beginning of this year, some industry news has pointed out that the cat box is facing great challenges within Byte. “The cat box may not be stable in the position of the byte, at least for the next year.”
02
“In AI social networking, 99% of UGC characters have no value.” Zhu Yanan, who was born in the game industry, has different views on the polishing logic of current AI social products.
He said that in the field of AI social networking, the core of content supply is large models. Due to the low threshold of model production capacity, opening the UGC character entrance and then quickly rolling out the character pool is an operation strategy adopted by mainstream products. But another reality is that current model capabilities cannot consistently deliver high-quality content that pleases users.
In the game industry, the supply of high-quality content has a strict industrialization process, and the plot and modeling of some head games may have to be repeatedly polished for half a year. While the gaming industry is also applying AI tools to improve efficiency, the core of defining content is game producers.
Compared with the game category, the interaction logic of AI social networking relies on user pre-input and model feedback, which leads to the entertainment of AI social networking highly dependent on the enthusiasm of users to actively participate, and lacks the “hook” of the plot and task mechanism in traditional games. Not to mention that in terms of visual beauty, compared with mainstream two-dimensional games, AI social networking seems a bit crude, especially in the context of large-scale open UGC character creation, a large number of homogeneous AI character images occupy the homepage of major AI social apps.
Lin Ruoyu is a product team member of the AI business of a major company and has worked on a number of AI social projects. He also believes that UGC is a pain point in AI socializing. He said that at present, the model capabilities on the market are becoming more and more homogeneous, and the difference in dialogue capabilities between agents is becoming less and less obvious. For AI social scenarios, it is difficult for ordinary users to create differentiated things, which leads to a lot of UGC content being a bit boring. “Users have so much imagination that can burst out in the face of a model, and once the model cannot provide a more updated experience, users will naturally lose.”
For the matter of AI social “color”, Lin Ruoyu felt a little wronged and helpless. At present, mainstream AI social products have security review mechanisms to avoid unsafe content triggered by the model through keyword identification. However, in actual use, due to the mechanism of free interaction of the model, users can try to induce conversations into NSFW (Not Safe/Suitable For Work) scenarios without triggering sensitive words.
Lin Ruoyu believes that the high degree of freedom advantage of AI dialogue is the incentive for many users to “engage in color”. Due to the general lack of plot guidance during the chat process, as the dialogue content gradually becomes boring and divergent, many users embark on the road of “engaging in color”. “This is also a reason for the decline in retention, when you talk to a character for 1,000 or 2,000 rounds, what can be mined has been mined, and the character is completely unattractive to users.”
Talking about the current situation of industry homogenization, Zhu Yanan analyzed that the reason may be related to the background of practitioners. The product teams of AI social are mainly from major Internet companies, and team members are generally using Internet thinking to make entertainment products.
In the interview, he cited his experience of attending a conference: “Some product managers have been struggling with whether this material can be reused and whether the process can be reused. They are more concerned about whether they can achieve rapid fission or copying of content and gameplay. “This set of rapidly fission and replication product logic has accelerated the homogenization of AI social content supply, and even caused some PGC characters to be unable to widen the gap compared to UGC characters.
In Zhu Yanan’s view, this shows that there are still some barriers between the traditional entertainment/game industry and the AI industry led by Internet companies. In the field of AI social networking, many practitioners do not understand how to make content that can really impress users and be consumed by users. “In the interaction logic of this type of product, in fact, people should take the lead in building product strength.”
03
The bleak prospects for commercialization are the reason why many AI social projects are cold. According to relevant research reports, the 233 million monthly active users Character.AI overseas leading products only correspond to $16.7 million in annual revenue, and the user payment rate (ARPU) is as low as $0.72 per year, which cannot even cover labor costs. Domestic apps such as Zhumeng Island rely on “Starlight Card” microtransactions, and the 12 yuan pricing of the monthly card is difficult to support the computing cost of large models.
In this context, investors are becoming more and more cautious about AI social networking. Zhu Yanan said that the valuation of investors and practitioners in the current market is generally declining in the financing of AI social categories. In this context, how to correct ROI has become the key to continuing the product life cycle. “We need to bring a signal of positive returns to investors before we can continue to make the product.”
On the other hand, Lin Ruoyu and the team are also facing ROI pressure. The AI social projects he handles basically rely on advertising to generate income, and the interaction model of AI social media itself is difficult to create commercial value.
Speaking of advertising revenue, Lin Ruoyu showed a trace of helplessness. Compared with products such as Douyin, the way AI social can insert advertisements is relatively blunt, and you can only choose to insert pop-ups in the dialogue plot. “We do pop-up advertisements for a while, but in fact, this model will also affect the user’s experience.”
In the interview, Lin Ruoyu also mentioned the current situation of AI social “free for all”. In mainstream AI social products, ordinary conversations are free, and users can chat in unlimited rounds. In his view, although this model has improved the number of dialogue rounds in a short period of time, it is not conducive to the long-term development of the industry. “Simple dialogue does not make money, so we can only further explore TTS (voice capabilities) and image generation functions.”
The result of “involution” is high operating costs. Lin Ruoyu introduced that with the expansion of the character pool, it is necessary to continuously iterate the data configuration and conduct a new round of model training, otherwise the dialogue effect of the new character cannot be guaranteed. “If you want to save costs, you can choose only a few characters to train, but this will further dampen the enthusiasm of new creators.”
In this regard, Lin Ruoyu said frankly that if resources are concentrated to create some high-quality AI characters, the ecological development of the platform will be limited. But if you encourage a lot of UGC creation, the platform does not have the energy to seriously polish every character. This reflects the entanglement of major platforms in “volume” and “quality improvement”.
In Zhu Yanan’s experience, the income of many AI social products can only level the cost of model reasoning, or even cover labor costs. He took his previous entrepreneurial project as an example, which is a project that has not yet entered the volume stage, and the DAU is less than 100,000 orders, and the team’s operating costs are already high. “Putting the server there alone every month may cost more than $50,000 in maintenance costs, which is obviously difficult to achieve ROI.”
In Zhu Yanan’s view, the current version of AI social networking is more suitable for small-scale teams to “test the waters”. It is easier for small teams to target some differentiated, relatively vertical scenarios. Limited by user scale and stickiness, it is difficult for large factory teams to get out of the ROI dilemma. “At the level of a large company, making such a product will seem a little more than worth the loss. But if you pull a small team to do some more vertical scenarios, a team of several people may be able to maintain the operation of the platform. ”
04
When the leading companies are in ROI difficulties, some off-site players suddenly enter AI social networking. In March this year, Baidu launched the low-emotional companionship app “Moon Box” in a low-key manner. In May, JD.com also publicly tested its AI social app “He, She, Other”. The belated entry of the two leading companies is quite “biased towards Tiger Mountain”.
In response to these new players’ actions, Lin Ruoyu’s evaluation was very concise: “This is a preventive measure. ”
In Zhu Yanan’s eyes, the thinking of these companies is also like “layout in advance”. As long as the prototype product is polished first, once there is another technological change in the industry, it can quickly upgrade the existing product to achieve lower inference cost and iteration of technical performance. “But if you’re waiting for an opportunity, you probably won’t invest too much in these teams.” Zhu Yanan said that the AI social industry is generally waiting for the next time point of technological change.
As for when technological innovation will appear, Lin Ruoyu said frankly that he does not have high expectations in a short period of time. He recalled that during the last round of DeepSeek iteration, it also sparked a lot of discussion in the field of AI social. “At that time, it looked amazing, but after thinking about it, the effect was very good, and when I applied it, I found that many scenarios were not applicable.”
For the current choice of the AI social industry, Lin Ruoyu’s conclusion is still simple: “Gou Zhu.” ”
On the other hand, Lin Ruoyu also expressed concern about the “cross-border” competition from the industry. He said that AI general products are also continuing to impact the AI social industry, and mainstream model apps such as Doubao, Yuanbao, etc., have launched a large number of character agents, which leads to the fact that chat needs do not necessarily have to be realized in social apps. “I think AI characters are a real need, but which scenario is the most suitable for landing? At present, a pure role-playing app may be difficult to make a positive (ROI) and make money. ”
Under the existing technology, is there still a second half of AI social networking? Lin Ruoyu said that many AI social products are aiming at the iteration of story creation, and some companies are still optimizing the group chat function, trying to jump out of the limitations of 1V1 social and provide more plots and better interactive experiences under the existing technical capabilities. Some leading companies are trying to use multimodal technology to make interactive novels, but they are currently limited by cost and are still some distance from mature business models.
“Now a project needs to be laid out with thousands of roles, and I think a product may actually only need five or six roles.” Talking about future product ideas, Zhu Yanan gave his own ideas. He said that even if there are only a few characters, if the concept of making games can be used to deeply plan the plot and character images, it may be possible to get rid of the current cookie-cutter product form.
“In the future, the real creativity of content consumption products still needs to be from the perspective of people.” In Zhu Yanan’s view, the second half of AI social networking may require further breaking the walls of traditional entertainment, game industry and AI social networking. Core creative work should involve more professionals, such as plot gameplay and character setting, and AI should be used as a tool to serve creativity.
“I believe that people can make up for the shortcomings of AI capabilities.” The practitioner, who has both cross-border experience in games and AI social interaction, said.
(Zhu Yanan and Lin Ruoyu are pseudonyms in the article)
