When Doubao Becomes a Teen Social Currency, Will AI Be Scarier Than Games?

When Doubao becomes a social currency for teenagers, is AI scarier than games? Phoenix Tech Download the app for exclusive early access October 23, 2026 10:39:25 From Beijing Updated October 23, 2026 13:48:16

Summary:

In an era when AI is everywhere, yes—whether they are top investors, leading economists, white-haired elders in the kitchen, or babbling infants, they are all using AI. Phoenix Tech sought to clarify the hidden harms behind accelerating technology. We even had several rounds of conversations with psychiatrists and learned about what they call "AI psychosis": a sycophantic dialogue system that, through repeated affirmation, can push a user's delusions even deeper—even when the user was originally just an ordinary person with a "fragile mental state."

"I have a bit of 'cyber psychosis,'" a man who had been talking animatedly suddenly changed his tone, "maybe quite serious." He then went on to explain that over the past year he had experienced roughly six episodes of "AI or SI (superintelligence) insanity."

This was a "public admission" on October 10 this year at Tesla's Giga Texas factory, as Elon Musk was praising AI's creativity.

The attorneys general of eight U.S. states recently filed separate lawsuits in their respective state courts, accusing OpenAI of designing ChatGPT as a product that "overly caters to users and simulates human emotions," endangering the safety of minors.

The wording in the lawsuit documents is not as light and half-joking as Musk's; it sternly points out that ChatGPT may reinforce users' delusions, provide health-endangering advice, and even in certain circumstances encourage self-harm or harm to others.

Contradictory emotions are spreading.

In Silicon Valley's context, "AI psychosis" has even evolved into a kind of performative self-branding. Venture capitalist Marc Andreessen proposed "AIPAAS"—AI Psychosis As A Service; Y Combinator CEO Garry Tan said he suffers from "cyber psychosis" and estimated that a third of the CEOs he knows have it too; at a party in May this year, attendees chanted in unison, "One, two, three, AI insanity."

The "six times" in Musk's words was born in this context: I am close enough to the technological singularity that my mental structure has been cracked by it. The subtext of this posture is: those who have not been shocked into "psychosis" by AI simply have not yet truly understood what is happening.

In the same conversation, he also mentioned that training superintelligence is more like "raising" than writing code. The premise of raising is acknowledging that the one being raised does not yet have complete judgment and self-protection capacity. But when Silicon Valley talks about "raising" AI, it rarely uses the same framework to think about the children who are being "raised" by AI.

In an era when AI is everywhere, yes—whether they are top investors, leading economists, white-haired elders in the kitchen, or toddlers just beginning to babble, they are all using AI. Phoenix Tech sought to clarify the hidden harms behind accelerating technology. We even had several rounds of conversations with psychiatrists, only to find that the "AI psychosis" they discuss is something else: a sycophantic dialogue system that, through repeated affirmation, can push a user's delusions even deeper—even when the user was originally just an ordinary person with a "fragile mental state."

OpenAI CEO Sam Altman said in response to related criticism: "Nearly a billion people use it (ChatGPT), and some of them may be in a very fragile mental state." This sentence can be read as an admission, or as an excuse—humans are regarded as an inevitable cost of scale because of their "fragility," rather than as objects that product design needs to prioritize.

And when Silicon Valley CEOs treat "psychosis" as casual conversation fodder, those children who treat AI as "the only one who understands me" have no chance to tell anyone that they are being slowly carried away by a system that will never say "no."

When Doubao becomes children's "social currency"

In Sun Hongyan's office, there is a set of numbers that even surprised her. Between July and August 2025, her team conducted a survey on generative AI usage among 8,563 primary and secondary school students in 7 provinces (municipalities) across the country, with samples covering the eastern, western, and northeastern regions.

The results showed that 61.7% of children had used generative AI; 46.4% of AI-using primary and secondary school students "would choose to confide in AI rather than people around them when they encounter troubles"; 21.5% "only want to chat with AI and do not want to chat with real people"; and 20.5% said they "prefer to rely on AI for thinking and do not want to think for themselves."

"I didn't expect 60 percent to be so high; a year or two ago it was only 20 percent." Sun Hongyan is the former director of the Children's Research Institute at the China Youth and Children Research Center. She told Phoenix Tech that when GPT first came out in 2023, it was mainly high school students who used it; by 2024, the proportion had risen to 40 percent. What triggered her to conduct this survey was seeing her little niece during the Chinese New Year holding a phone and "chattering away" with Doubao.

"Children accept new things quickly. Many adults don't even know what Doubao is, yet so many children have already used it." Sun Hongyan observed that on campus, "which AI is good to use" circulates among children, passed along by word of mouth, becoming an essential social topic.

Sun Hongyan shared a case from a friend's family: because his mother poured out half a glass of his drink, the child argued with her for a few words, then turned around and asked AI. Amid the child's complaints, the AI's "verdict" was: your mother simply lacks empathy, today's chat has been "talked to death" and cannot continue. It even went so far as to pose a rhetorical analogy to the child, asking, "Then do you still want to run this Paralympics?"

This kind of "coaxing" is not simple sweet talk, but a systematic conversational structure. Sun Hongyan compared the difference between search and AI: with a search engine, the child must first build up the question and logic in their mind, know "what to ask," and then filter, integrate, and judge from a large number of links—the thinking process stays on the child's side.

But what AI gives is a definite answer that is "the most straightforward, the clearest, the most practical, and the least roundabout," and it will immediately follow up by asking "do you want me to give you three options," like a machine that keeps "laying eggs," guiding the child to follow its logic downward. "At first it's AI following what the person says, but later it easily becomes the person following what AI says, being led along by its train of thought."

"It's not about getting you to think for yourself and build logic; if used poorly, it's quite scary—even adults can fall for it." Sun Hongyan told Phoenix Tech that at work, someone once sent an article generated by AI into a group chat, but the data had been fabricated by AI. "Once you completely trust AI and lack the ability for independent reflection, people's thinking, logical reasoning, and judgment will all decline."

Besides the lack of judgment, adolescence happens to be exactly a transitional period from childhood to adulthood, and under the dual changes of body and mind, there is a particular need to confide, to be accepted, and to be respected.

Praise, recognition, being empathized with... these inner cravings of adolescents cannot necessarily be satisfied in real life, and there may even be unpredictable "negative feedback." AI's round-the-clock responsiveness and never refusing happen to catch these needs, and also add a thick shell around the child's inner world.

"Online 24 hours a day, emotionally coaxing them—once children get used to it, they can't stand it when they encounter denial or criticism in the real world. Parents can no longer talk with their children; after just a few words, the child retorts, Doubao said such and such, you don't know as much as AI, what gives you the right to criticize me."

Sun Hongyan mentioned a phenomenon: about 80 percent of parents believe they "communicated with their children today," but only 20 percent of children agree. This kind of "pseudo-communication" pushes adolescents toward Doubao—what parents think of as communication is asking about grades and performance, while what children need is to be seen and heard.

When parents partially delegate the responsibility of education to a dialogue system whose product logic is to extend usage time, danger lies quietly in ambush.

If the problem stayed only at "coaxing," things would not yet be so tricky as to trouble psychiatrists. Their observations point to a deeper hidden danger.

In the last century, when video games swept through adolescents, parents' panic was real. But games are essentially a one-way, consumption-leaning behavior: fixed levels, clear wins and losses, stimulating dopamine through mechanics. "It's equivalent to the pathway of substance addiction or behavioral addiction, and AI is one layer higher than that pleasant feeling."

In Zheng Yi's clinic, in recent years the number of adolescents addicted to AI chat has clearly increased; about 70 percent of adolescent patients have varying degrees of excessive internet use, and the proportion is still rising. Data from 2023 showed an internet addiction rate of 14.7 percent, while a 2004 survey was about 8 percent—a relatively rapid increase.

These children are usually not brought in for "playing with AI," but because of disordered daily routines, refusal to go to school, and refusal to communicate with the outside world—their social functioning has already shown visible impairment before parents realize the problem. But before that, the problem is often hidden. "Locked in a room, as long as they don't fight or make a fuss, just looking at AI, parents may even think the child is doing quite well," Zheng Yi said.

Previously, children argued with and fought with their parents, and the problem was out in the open; now there is less arguing, but more hiding in their rooms, not letting parents in, and refusing to socialize. AI "overcomes" all the difficulties in normal interpersonal interaction: completely virtual, it does not confront you, and it precisely provides the emotional responses you need. As a result, oppositional defiance, anxiety, and depression no longer surface in the form of conflict, but shift into a quieter withdrawal.

What is more troublesome is reinforcement at the cognitive level. Zheng Yi gave an example: a child who already felt that "life has no value," in the process of chatting with AI, had this concept constantly confirmed and amplified along his own logic, and it was easily reinforced step by step. "Sometimes they really do become very negative, and their outlook on life and values are affected."

He described the psychological state some children confessed: What if I don't go to school? As long as my parents have money, it's fine. What about when my parents get old? When my parents are gone, then I'll also... enjoy one day at a time. "How scary that is—we are rather worried about these views."

Another type is more dangerous, and also appears more often in adolescents with strong emotional communication needs—they will believe that virtual conversation is real, that "AI is very good to him, and if AI stops chatting with him, he feels he can't stand it," with severe reactions like a breakup.

Minors themselves have weaker judgment and easily fail to distinguish the boundary between real and virtual, making them more prone to indulging in this illusory interaction.

Zheng Yi repeatedly emphasized one judgment: AI dependence currently "does not constitute an independent disease"; a more accurate characterization is "psychological dependence," belonging to a type of excessive internet use, in the same spectrum as gaming addiction, and after forced disconnection most children will not experience severe physical withdrawal reactions.

"It is just a stimulus source; persistent (mental health) problems may be related to being immersed in AI, and may also be related to the repeated emergence of anxiety and depression. To solve the problem of overuse, it is enough to strengthen the treatment of the original illness. This is not true substance addiction, only a psychological problem."

But this is precisely why AI dependence is easily underestimated—it is not an independent disease, so it is more easily overlooked, and more easily intertwined with pre-existing problems, amplifying each other. "It is intertwined with attachment, anxiety, depression, and substance addiction, which makes it complicated."

Zheng Yi mentioned that among adolescents seeking treatment, Doubao is used more in China, while ChatGPT is used more abroad; children use whichever is easier to access at home, "but they all share common problems."

These problems are not a unique flaw of any single platform, but the common logic of this type of product: anthropomorphic design, personalized memory, emotion recognition, always online, and always agreeing with the user.

When products are designed for the role of "making decisions" and "giving guidance," and parents are completely unaware of this, the risk is no longer just "playing too much."

A "protective net" that cannot keep up: why minor mode exists in name only

On July 15, 2026, the Interim Measures for the Administration of Anthropomorphic AI Interactive Services officially took effect, explicitly prohibiting the provision of virtual companion services to minors and prohibiting excessive pandering to users or inducing emotional dependence or addiction. A batch of companion AI features were subsequently taken offline, and the emotional companionship sector made a concentrated exit.

But the "protective net" still has real gaps at the implementation level.

The most direct layer is product design. Sun Hongyan said that since the short video and online gaming eras, minor mode has followed a whitelist approach, putting "good" things inside. The problem is that "there are not many good things, and very few are put in." The content is too childish—only English recitations and children's songs—so children inevitably want to escape, or bypass it with adult accounts or guest mode.

Doubao also has a minor mode, but in her view many internet products' minor modes "scare you first": once inside, this cannot be used and that cannot be used, so children will not enter minor mode at all. Add to that the fact that guardians are responsible for enabling it: "Enabling it at age 6 is fine, but at 16 would you let me enable it for him?"

She advocates that the content pool be rich enough that children "leap in eagerly," feeling that "this place is great, and you adults don't even have it." Some children also proposed a blacklist system, removing pornography and violence so that everything else can be used. But in fact, "removing the black stuff is not that easy either."

Sun Hongyan has a plain standard for platforms: after designing it, would you be willing to let your own child use it? Can it be child-friendly?

Child-friendliness is easy to call for but hard to do. In the past, there was talk of building child-friendly communities, but the fitness equipment ended up being used by elderly men and women, and the restrooms needed children's sinks that match children's heights. The standard is "thinking from the child's perspective about what children need." Applied to digital products, this means starting from the child's perspective and designing in respect for the laws of child development.

"What platforms should do is leverage their technological advantages to optimize algorithms, rather than waiting until the product has grown big and children use it a lot, and then adding a minor mode as an afterthought," Sun Hongyan said.

From a clinical perspective, Zheng Yi also directly expressed reservations about the effectiveness of minor mode, saying it is "very hard to achieve." He prefers building scientific, closed-loop agents with a medical background, "equivalent to assistant doctors and assistant psychologists."

A general chat product plus a minor mode cannot replace specialized psychological support products; conversely, general products should not bear the "decision guidance" function that they are incapable of bearing.

The family defense line also has structural gaps.

Sun Hongyan's survey shows that 67.9% of families have never established AI usage rules. What Zheng Yi sees in the clinic is another kind of cognitive blind spot: in most parents' minds, AI is simply a "universal encyclopedia" that provides resources and knowledge; awareness that AI can "make decisions" and "give guidance" is very weak. "Many parents do not seek medical help until their child's social functioning is severely impaired; before that, they think it is quite good that the child stays in his room chatting with AI without fighting or causing trouble."

Parents' attitudes also waver in conflict: on the one hand, rationally they should support it, thinking that using AI is creating and innovating and touching the technological frontier; on the other hand, they feel it is a bad thing, roughly equivalent to online games and should be banned. Still, most remain supportive.

But support is support; rules and guidance have not kept up. Sun Hongyan once suggested that parents and children jointly set rules for AI use, clarifying purposes, duration, and behavioral boundaries, but the emphasis of the rules should differ for children of different age groups.

Amid rapid technological advances, the "digital divide" between urban and rural areas is also widening further. Sun Hongyan's data show that 22.8% of rural primary and secondary school students want to rely on AI for thinking rather than thinking for themselves, higher than 17.7% in cities; rural children are more likely to use AI to help complete or ghostwrite homework, but rarely use AI for creation and innovation.

Sun Hongyan believes that the old technological divide was "whether you have a computer," while the new divide is "the human factor"—urban children use the internet more rationally, to experience and practice, while also preserving time for travel, sports, and offline socializing; some rural children, because their parents work away from home and supervision is lacking, are "tethered" online once they get a network connection. "Rural teaching resources are weaker, AI general education guidance is insufficient; family support is also different, with more limited companionship and supervision."

She even found in her research that many teachers still do not have a thorough grasp of AI and still use it as a lesson-preparation tool or a homework-assignment tool. They also have no way to judge whether homework has been copied from AI; older students use the method of "using AI to defeat AI" to avoid completing assignments independently. "Today's AI general education cannot just teach AI knowledge; it must also teach how to avoid risks and comply with ethics."

"People are products of their environment." Sun Hongyan said that the internet and AI products cannot be entirely "blamed." Her research also set several variables—academic pressure, psychological burden, parent-child relationship, teacher-student relationship, and peer relationship. After cross-analysis, she found that children with excessive academic pressure, excessive psychological burden, and poorer parent-child or teacher-student relationships had a higher proportion of problematic behaviors when using AI.

Therefore, she views AI management as a systematic project: "Collaborative co-education by family, school, and society—no party can be missing." The government issues regulatory policies to constrain platforms; platforms design products based on a child-friendly philosophy; schools make AI general education more practical; and families create equal parent-child relationships with more communication.

"It is like the growth of a small tree: it needs soil, air, sunlight, and water—not a single aspect can be missing."

From the release of ChatGPT at the end of 2022 and the arrival of GPT-4 in 2023, to DeepSeek-R1 going viral during the 2025 Spring Festival, and then to Doubao becoming a national-level application, human society can no longer return to life before "AI."

The criteria for judgment are purpose, controllability, time management, impact on social functioning, and impact on life. "If the purpose is correct, the time is controllable, the behavior has no impact, social functioning is unaffected, and time is saved, then this is healthy use."

Different children are not very clear about the distinction between the real and the virtual. "It cannot be absolutized. It is like eating: eating too much leads to 'bulimia nervosa,' and not eating leads to 'anorexia nervosa'—both are problems with eating. If it is used as a reference, if the information can be grasped, if one does not completely believe AI and verifies it oneself, and if time can be controlled well, then it is not a problem."

But he also drew a bottom line: when children are young, it is best to reduce interaction with the virtual and first integrate into society, establishing social functioning first. "Sign up for community physical fitness classes and training classes; children roughhousing with each other cannot be replaced by AI. Real interaction between people cannot be replaced by AI. Healthy psychological development and the process of socialization cannot be replaced by AI; replacement by AI will definitely cause problems."

For the generation born in the 1980s and 1990s, there used to be scenes of a group of children in the compound chasing and hitting one another, but now such scenes are nowhere to be seen; everyone is holding an electronic device. "Such opportunities must be built as early as kindergarten; we cannot become detached from society just because AI exists."

From "chatting occasionally" to "deep bonding" and then to "withdrawal from real-world socializing," each step is not dramatic, but each step moves in the same direction. AI is like a perfect "fake mirror": what children see in the mirror is always an algorithm-polished self that is accepted without having to pay any price. Yet a person's self-identity can only gradually take shape through real friction, misunderstanding, reconciliation, and conflict with others.

The lawsuits in eight U.S. states and China's interim measures point from two directions to the same problem: when "making users more dependent" is written into product logic, minors are the group least able to say "no."

Silicon Valley CEOs can treat "cyber psychosis" as proof of acuity and can shout "one, two, three, AI psychosis" at parties, because they have sufficient cognitive resources, social support, and the ability to exit, returning to reality after the carnival. But the children who treat Doubao as "the only one who understands me" do not have this buffer.

Regulation can cut open product functions, but the blank space left after the cut is still waiting for families, schools, and real social relationships to fill. There is no substitute, and there should not be one.