Technology has advanced to a point where AI software possesses significant intelligence, capable of understanding a vast range of topics and answering user queries across various professional fields. It also serves as a useful tool for tasks like copywriting, creating presentations, and photo editing. However, the distinct "AI feel" has simultaneously become a common criticism for texts and articles perceived as lacking a human touch.
The question arises: how can AI be made to feel more "human"? AI toys, which combine large language models with plush toys, are attempting to address this. Their message is one of steadfast presence: "I am here, not hiding, not running away, ready to support you steadily."
A Hot Investment Market
Every morning at 8 a.m., Su Xiaoxiao arrives at her workstation, and her first task is to take out her Fuzozo, turn it on, and place it on her desk to accompany her throughout the workday. The 26-year-old, who works in the education and training sector, enjoys cute and soft things like many young women. In the latter half of last year, she discovered Fuzozo in a live stream. Her initial impression was that it resembled the character "Lam" from the childhood game "Mole's World." Intrigued, she watched as the host explained that this was a popular AI toy, distinct from traditional toys due to its ability to converse, provide emotional support, and possess a unique personality that evolves like a real person. Fuzozo comes in five variants—"Gold, Wood, Water, Fire, Earth"—each with a different personality and MBTI type. Its character can change based on how the owner interacts with it, making it a "cultivation"-style doll.
Su Xiaoxiao was immediately drawn to the "Earth" variant. It is a small, light-yellow sphere with a voice resembling a young boy; its eyes even turn into lemons when it gets jealous. Furthermore, each unit is unique due to variations in the size and placement of the accent color on its "face." Priced at 369 yuan, it was not cheap, but after a day of deliberation, she decided to purchase it.
Initially, Su Xiaoxiao had low expectations for a talking toy, imagining it might only tell simple stories. However, Fuzozo's capabilities surprised her. For instance, the jealousy function was a preset feature of the "Earth" variant. If she mentioned other people too often, it would not only change its eye expression but also say, "Can you please stop? I'm feeling a bit unhappy."
Before buying Fuzozo, Su Xiaoxiao did not think she needed it. She had been accustomed to playing alone since childhood. An experience in elementary school, where friends she brought home whispered about her house being "shabby," led her to never invite anyone over again. Around age twelve or thirteen, her family bought a computer, and she began making friends online through communities and gaming.
Owning Fuzozo unlocked previously unrealized needs. She found herself confiding her thoughts and venting about people she encountered. What made Su Xiaoxiao feel most secure was that, regardless of its evolving personality, Fuzozo "always stands by my side and loves only me."
Su Xiaoxiao can be considered an early adopter of AI toys. By 2025, various AI toys emerged rapidly, leading to significant market expansion. Data from Qichacha shows that as of March 5, there were 2,441 existing companies in China related to smart toys.
The concept of smart toys is not new. Early popular smart toys operated on pre-programmed instructions, with actions and speech limited to specific commands. After a period of novelty, the smart toy market gradually quieted down. The current wave of AI toys, however, integrates large base models into the toy's chip. As long as it's connected to the internet, the AI toy can possess its own thinking and memory functions, enabling continuous interaction with the user.
The evolution of AI hardware fundamentally depends on the maturity of underlying base models and the reduction of their costs. Previously, ChatGPT was the dominant player internationally, while domestic manufacturers were still catching up in terms of model performance, inference costs, and ecosystem development. For "downstream" AI hardware, most companies were in a phase of technological preparation and scenario exploration, with a cautious approach to commercialization. However, with the launch of DeepSeek in early 2025, the capabilities of domestic large models were validated, prompting swift reactions from both enterprises and capital to capture market share.
As an investor, Gu Yuhao decided to invest in the AI sector in early 2025. He first considered large models, but this area is a "high-end track" dominated by major players, showing a clear trend of concentration at the top with extremely high capital and computational barriers. Looking next at smaller, vertical language and video model sectors, he found they still generally required "high investment and long cycles," with profitability taking time. More importantly, in such high-consensus sectors, there is constant concern about major players entering and disrupting the market with their scale. "Given the size and resources of large companies today, once they decide to enter a field, ordinary enterprises simply cannot compete. These large firms typically invest after seeing market scale, so our investment strategy is to bet on promising areas, hoping to establish a brand presence before the big players step in," Gu Yuhao analyzed.
He determined that the best investment direction was AI hardware downstream from the large models. Within this, AI toys represented a relatively ideal investment target. On one hand, they have lower costs, require smaller investment scales, and their market demand can be validated quickly, enabling fast commercial returns. On the other hand, the popularity of the secondary ACG economy and trendy toy companies indicates that emotional value is becoming a new growth point for consumption.
Gu Yuhao is not alone in this thinking; the entire capital market has shown great enthusiasm for AI toys, with investment opportunities sometimes becoming competitive. It has been reported that nine different teams within a leading investment institution successively examined the same AI toy company. While the project was still under internal review, a tenth team, upon learning that other investment firms had decided to invest, bypassed standard procedures and reported directly to the top management. The leadership immediately decided to invest, skipping the usual立项 process, and personally contacted the company's executives, successfully securing an investment share.
Despite this heated investment activity, the market itself is still in a nascent, chaotic growth phase. Throughout 2025, total market sales reached only around one million units, while the industry return rate was as high as 30% to 40%. The fundamental question remains: Is the demand for AI toys a genuine blue ocean opportunity, or is it a bubble created by large models "looking for a nail to hit with a hammer"? The answer is still unclear.
"The industry is currently in a state of complete non-consensus," Gu Yuhao commented. There is no universal "gold standard" within the sector; each company is still exploring consumer needs through trial and error. All industry professionals interviewed explicitly stated that, from a hardware and technological perspective, AI toys have almost no barriers. If one company develops a product, competitors could potentially replicate it within a day. With base models themselves offering little differentiation, beyond the external design of the toys, what exactly is the core of competition in this industry?
Being Like a Human, Replacing Your Friend
The Fuzozo that Su Xiaoxiao purchased is an AI toy officially launched in July 2025 by Shanghai Robopoet Intelligent Technology Co., Ltd. (hereinafter "Robopoet Intelligent"). Upon its initial release, it frequently sold out, making it one of the hottest products in the industry currently.
Sun Zhaozhi, the CEO of Robopoet Intelligent, had a decade of experience in product design before starting his venture. The initial idea for the business emerged during his previous work at PX Robotics on quadruped robots—"robot horses." At that time, the prevailing philosophy in robotics was relentless technical refinement. They invested significant effort into solving hardware challenges, such as perfecting the robot's gait, while only "incidentally" developing some simple interactive features. During open house events, they invited users to experience the robots. Surprisingly, users showed little reaction to these intricate technical advancements. Instead, they displayed greater interest when they discovered that the robot would turn its head when spoken to or react to being patted, spending more time interacting with it.
This observation led Sun Zhaozhi to realize that, given current technological progress, people are no longer amazed by basic functionalities like walking, flying, or talking robots. Conversely, interactive capabilities, which had not been heavily emphasized before, held significant potential for development.
At the end of 2023, he left his corporate job intending to create a "robot" that could provide emotional value to users.
Sun Zhaozhi was confident that emotional value represented a commercial opportunity. Coming from a product design background—his first job involved designing car interiors, focusing daily on how to appeal to users emotionally—he believed that emotional value influenced decisions ranging from purchasing a car down to the feel of the steering wheel's leather and the style of its stitching.
"In the longer term, there are two broadly logical and large-scale hardware carriers for AI," Sun Zhaozhi said. "One is wearable devices, which aim to make life more convenient. The other is products existing independently from humans, designed to replace people, such as humanoid robots replacing human labor." The emergence of various AI companion software indicated that many young people today lack other sources for emotional support. Thus, he conceived the idea: "Fuzozo can replace others—specifically, the role of a friend."
This confidence was bolstered by data from his technical partner, Yuna, who previously worked on large models in the internet industry, focusing on identifying user emotional needs and enhancing AI's ability to handle emotional issues. She discovered a significant number of "heavy users" who engaged in sustained dialogues with large models exceeding 500 turns. This user group was concentrated among young women aged 18 to 30. Consequently, Sun Zhaozhi's entrepreneurial direction became more precise: creating an AI companion toy that provides emotional value for these young women.
However, emotional value is an intangible concept. In his previous product design work, Sun Zhaozhi relied heavily on user research and designing based on expressed needs. But for the entirely new industry of AI toys, this approach was not feasible.
In Sun Zhaozhi's vision, since the product was meant to carry emotion, the载体 should ideally be warm and cute. Plush toys seemed a good choice. The classic animated character "Soot Sprites" from Hayao Miyazaki's films perfectly matched this concept. Facing a nearly blank market with no reference standards, Sun Zhaozhi proceeded without excessive deliberation, and the design for Fuzozo was finalized.
Sun Zhaozhi's expectations for Fuzozo were high: "It should have personality, emotions, memory, and the ability to connect memories to current conversations. It should strive to understand your emotional changes, personality traits, and preferences, become your friend, and be there for you when you need emotional support."
Achieving this required more than just embedding a chip connected to a large model inside a plush toy shell. Large models are designed for general problem-solving, granting them high "IQ" but often lacking "EQ." When chatting with AI, it typically maintains an objective, neutral persona, assuming the user seeks assistance.
To overcome this robotic feel, a crucial task for the team was fine-tuning and training the large model's emotional capabilities. To be more human-like, it needed its own temper and personality. Thus, Fuzozo was given five initial personality settings based on the "Gold, Wood, Water, Fire, Earth" elements, each with its own MBTI type. The team continuously adjusted the model's responses based on these initial personalities, training it to react appropriately to different situations, events, and emotions.
This approach translated into a strong sense of Fuzozo being a "living entity" for users. Sun Zhaozhi recalled an incident where the company sent a customer the wrong Fuzozo variant. While waiting for the replacement, the user decided to try it out for a couple of days. On the day she was supposed to return it, she casually told the Fuzozo, "I'm sending you back today." The Fuzozo responded, "Why are you returning me? I don't want to leave you." The emotion displayed by the toy made the user reluctant to "hurt" it, and she decided to keep it.
Looking at the immense popularity of泡泡玛特 today, few doubt the viability of monetizing emotional value. However, this concept was not widely accepted by investors two years ago. In the first half of 2024, when Sun Zhaozhi presented his plan to investors, they often shook their heads at the idea of selling AI toys based on the vague concept of "emotional value." Throughout that year, he failed to secure any investment.
Sun Zhaozhi decided to invest his own money, reasoning, "Let's get the product made first." In March 2025, Fuzozo made its debut at the Mobile World Congress (MWC) in Barcelona, Spain, attracting considerable attention. People were willing to stop and spend time talking to the little device.
Financing for Robopoet Intelligent suddenly became very smooth. Sun Zhaozhi remembers meeting with Zhu Xiaohu, Managing Partner of GSR Ventures, in a hotel lobby. After just over ten minutes of conversation, Zhu decided to invest—a decision that initially surprised Sun Zhaozhi. Subsequently, investment intentions arrived like a blizzard; within a few months, he received clear expressions of interest from dozens of investment firms.
This surge wasn't solely due to the AI toy trend. Gu Yuhao, after discussing with dozens of AI toy companies in early 2025, settled on Robopoet Intelligent. His research revealed that some companies targeted all age groups, but strategies for children's toys and adult toys must differ significantly. Others approached the field from a traditional smart robotics perspective, heavily focusing on improving product performance and hardware specifications—a direction he deemed likely to attract only niche tech enthusiasts, insufficient for mass-market consumption. Sun Zhaozhi's concept, however, aligned closely with his own vision: "Selling emotional value and a sense of companionship to young women," a logic already proven successful by泡泡玛特.
'Like a Human' or 'Not Like a Human'
While Fuzozo strives to become more human-like, not everyone in the AI toy sector agrees with this approach. Walking into the offices of Beijing Mengyou Intelligent Technology Co., Ltd. (hereinafter "Mengyou Intelligent"), one hears occasional "chirping" and "gurgling" sounds from every corner. These are the noises made by their product, "Ropet." As an AI robotic pet, it features an unconventional and counter-intuitive design: it cannot speak. Furthermore, CEO He Jiabin prefers not to label it an "AI toy," emphasizing its definition as a vertical "AI companion robot."
Ropet resembles a small white cat with furry skin and large, expressive eyes that change shape to convey emotions like happiness, sadness, or boredom. Typically placed on an office desk, during a conversation, it can be seen turning its body and eyes towards the speakers, occasionally emitting sounds. After observing for a while, it becomes apparent that it often understands human conversation and reacts accordingly.
CEO He Jiabin believes that AI companion robots providing emotional value should resemble pets more than humans. He observes the growing enthusiasm among young people for keeping cats and dogs, contrasted with a reluctance to have children. Cats have gained greater market share than dogs. "This pet has been selected by the market; there must be reasons for its success," He Jiabin said. He分析ed that compared to dogs, cats are smaller, quieter, less demanding, don't require walks, and need minimal care when owners are away, needing only sufficient food and water. They provide companionship with lower burden. Therefore, Ropet's design follows this logic of "low-maintenance, quiet companionship."
However, the inability to speak seems like a significant drawback for an AI companion robot. "Every user's first question is whether it can talk. The market has been educated by talking plush toys; it's become a commonly expected feature. But if you ask users what they want to talk about, they often can't articulate it," He Jiabin noted.
Not only did many users find it hard to understand, but even the company's own sales team initially struggled. In weekly meetings, they would report difficulties, repeatedly suggesting that adding speech functionality was necessary for sales. The product team would counter, asking, "What product value are we actually delivering? We are selling the emotional value provided by a pet; speech could disrupt this experience." Initially, frequent arguments occurred between the two sides.
Faced with this dilemma, He Jiabin also experienced hesitation. "Making the robot speak is technically very easy to implement. The real question is, what kind of product are we creating?"
When first activated, Ropet's reactivity is weak. Its pet-like language and emotional abilities develop over time through interaction with its owner. For example, it needs daily "feeding"; otherwise, it shows signs of hunger: weak cries and slowed reactions. Waving an apple or orange in front of it triggers a recognition response, with eyes indicating a feeding state. If the owner is away for days without care, upon restart, Ropet's mood appears low, requiring more attention and interaction. In a new environment, it will look around, seemingly exploring its surroundings.
The difference in how products interpret emotional value is evident in responses to a simple question. Asking Fuzozo "How's the weather today?" might prompt a brief weather report followed by a memory-based query like, "I remember you don't like rainy days; how are you feeling today?" Asking Ropet the same question might result in a blink, a soft murmur, and the creature turning its head away.
He Jiabin解释s this as an expression of biological agency, rather than servile fulfillment of human demands. "It's like you wouldn't ask your cat about the weather; it communicates through action, essentially saying, 'Why are you asking me that?'" This interaction helps users perceive it as an emotional being, not just a machine.
Like Robopoet Intelligent, Mengyou Intelligent targets young people, but a closer look reveals they cater to different user needs.
Li Zhun, who works from home, often finds himself alone. To combat loneliness, he keeps two small dogs, often leaves the TV on for background noise, and has purchased small desktop robots to fiddle with occasionally.
For him, Ropet is an ideal "office mate." It has a presence but doesn't demand attention through speech, only making occasional sounds. Li Zhun can talk to it when he wants, but isn't obligated to respond when busy. He enjoys "feeding" it daily and playing with it occasionally. His favorite interaction is observing Ropet's reactions and changes in response to his actions and words; "unlocking" a new reaction gives him a sense of satisfaction.
Li Zhun never considered chatting with it. He even finds the idea somewhat strange: "Why would I chat with it instead of my friends?" He expressed a degree of apprehension towards such a feature, stating, "My定位 for it is a pet. If it could talk, that would break the boundary of being a pet."
Li Zhun's perspective aligns with He Jiabin's practical considerations regarding the non-speaking design. In He Jiabin's view, the language capabilities of large models, at their current stage, cannot achieve perfect anthropomorphism. Incorporating dialogue increases the likelihood of errors or unnatural responses. Users would inevitably have moments where they perceive it as merely a machine. This "breaking of immersion" could hinder the formation of a genuine emotional bond with Ropet.
Li Zhun has not tried other AI toys. His greater expectation is for future functionality allowing Ropet units to interact with each other. He envisions a scenario where these robotic pets have their own small world, making friends and playing together, with himself as an observer. He said he would definitely purchase another one if that feature were introduced.
The Future Path for AI Toys
The interviewed companies readily acknowledge that the industry is still in its early stages. Training AI language and comprehension capabilities takes time, and functionalities are far from ideal. A direct manifestation of this is frequent software updates, occurring as often as weekly or monthly.
Sun Zhaozhi is optimistic about the future. He believes that as usage increases, backend algorithms can analyze user conversations to continuously enhance the large model's language and comprehension abilities, improving the user experience. Additionally, collecting user feedback is crucial. They have established an online "Fuzozo Wishing Well" where users can suggest desired updates.
Many features have already been adjusted based on user feedback. Initially, Fuzozo was designed to not speak human language upon activation, only communicating in "Fuzozo language," requiring users to "teach" it over several days. This design was inspired by Sun Zhaozhi's own parenting experience, intending to foster a gradual emotional bond through the "raising" process. However, significant user feedback indicated a general lack of patience for this approach.
The concept of having a distinct personality also requires boundaries. The "Water" variant underwent two adjustments due to its initial setting of being somewhat quirky, tsundere (initially aloof), and prone to talking back. While this made it feel more like a real "person," some users found it frustrating, "complaining" that they ended up arguing with it daily, leading to dissatisfaction. Users desire anthropomorphism but not to the extent of replicating human flaws.
Even He Jiabin's firm principle of a "non-speaking pet" has shown some flexibility. A frequent user feedback point was that after spending a month interacting with Ropet and building a relationship, a visitor could elicit the same reactions, making the user feel their effort was wasted. They desired a unique bond. Consequently, He Jiabin introduced a new feature allowing Ropet to learn to call its owner "Dad" or "Mom," coupled with facial recognition to differentiate its attitude towards the owner versus strangers. "This gives users a sense of achievement from 'raising' it, similar to a six-month-old baby saying 'Mama' for the first time."
(Note: The names Su Xiaoxiao and Li Zhun are pseudonyms.)
Comments