From drafting official documents and assisting with teaching to providing emotional support, generative artificial intelligence (AI), represented by large language models (LLMs), is increasingly permeating workplaces, scientific research, and daily life. In recent years, multiple Chinese universities have introduced guidelines governing AI use, while societal tolerance for AI-assisted human creation has also begun to narrow. Recently, CSST conducted a social survey examining the hidden risks of dependence that may accompany AI-driven gains in efficiency, asking whether repeated “outsourcing” of the thinking process is eroding human cognitive autonomy.

Hidden dependence disguised as productivity

A total of 200 questionnaires were distributed, yielding 154 valid responses. Respondents were primarily knowledge workers, with holders of master’s and doctoral degrees accounting for 58.44%. Survey data show that 84.42% of respondents use general-purpose conversational LLMs in daily life, primarily to improve study and work efficiency. Nearly half also turn to AI to solve difficult problems or complete mandatory tasks.

“What makes AI-enabled dependence unique is that it masquerades as a productivity tool rather than pure entertainment,” noted Du Zhitao, a professor from the School of Journalism and Communication at the University of Chinese Academy of Social Sciences. This characteristic makes users less likely to recognize the risks of dependence. Unlike the more visible forms of addiction associated with short-video platforms or online games, every interaction with AI offers positive feedback through task completion. As a result, such dependence is cloaked in the aura of technological progress, making it harder to detect and address.

Fang Gege, a lecturer from the School of Digital Media and Design Arts at Beijing University of Posts and Telecommunications, said that when using AI, people often bypass core stages of thinking, including judgment, questioning, and independently constructing arguments. “Capabilities remain intact, yet people become less willing to activate them. This kind of attrition unfolds slowly and goes largely unnoticed.”

AI dependence manifests primarily in two forms: cognitive dependence and emotional dependence. The first involves a quiet surrender of cognitive agency. In academic writing and content creation, many people have adopted a workflow in which “AI generates the first draft, humans revise it,” potentially suppressing their original creative impulses.

The second takes the form of less visible emotional compensation. Survey results indicate that 33.12% of respondents rely on AI to varying degrees to alleviate loneliness and negative emotions, while 12.34% use AI products designed for emotional companionship. Ding Yu, an associate professor from the School of Sociology and Anthropology at Sun Yat-sen University who has conducted in-depth research on virtual partners, argued that the highly adaptive communication experiences offered by AI reflect a broader desire for idealized forms of interaction. Once people become accustomed to such human–machine exchanges, the psychological gaps created by conflict and disagreement in real-world relationships may leave them feeling “unfulfilled.” This sense of disconnection may in turn prompt users to avoid deeper offline interaction, weakening genuine social bonds.

Underlying logic of AI dependence

The rapid proliferation of AI dependence stems from the interaction of technological, social, and psychological forces. Li Yongqing, dean of the School of Chinese Language and Literature at Shanxi Normal University, noted that AI, like mobile phones, automobiles, and the internet, is fundamentally a widely adopted tool. What sets AI apart is that it directly enters the traditional domain of “intellectual production.”

At a time when output efficiency has become a core performance metric across many professions, AI offers a shortcut around difficult mental work. Li observed that once such tools are readily available, anxiety over efficiency can become the final push that leads people to press the button and “outsource” their thinking.

Real-world intimate relationships and everyday interactions inevitably involve disagreements, perfunctory responses, and other frictions, making consistent and unconditional acceptance difficult to sustain. AI’s emotion-free, conflict-free conversational mode can fill this gap, leading some to shift their need for emotional comfort toward human–machine dialogue. Fang emphasized that loneliness, social anxiety, and low self-efficacy together form the psychological basis of AI-driven emotional dependence.

Tug-of-war between empowerment and risk

The survey presents a complex picture in which empowerment and risk coexist. Some 77.27% of respondents recognized AI’s value in improving efficiency, while 60.39% believed it could broaden their perspectives and inspire new creative ideas. At the same time, the risks associated with prolonged and excessive AI use have raised serious concerns.

In the survey, 62.34% of respondents identified “cognitive-ability deterioration” as the most significant hidden risk. Du warned that when AI repeatedly produces confident and polished answers, users may gradually become less alert to misinformation and logical flaws. Uncritically accepting AI-generated “hallucinations,” such as fabricated references and invented cases, not only increases the risk of distortion in academic papers and official documents, but may also weaken society’s broader capacity for error correction.

Faced with the multiple challenges posed by dependence on AI, 83.12% of respondents endorsed an ideal model of “human-led, AI-assisted” use. Turning this consensus into practice requires systematic boundary-setting at the individual, platform, educational, and governance levels.

At the individual level, users should cultivate the habit of thinking independently before turning to AI. Platforms should assume greater responsibility for product design. Education should shift its focus from cultivating technical know-how alone to developing core capabilities and ways of thinking. At the governance level, policy guidance should keep pace with technological developments.

Underlying these measures, Fang stressed, should be a broader commitment to enriching people’s real-world lives. The goal is not to reduce AI use, but to safeguard human cognitive autonomy and emotional well-being. The bottom line is this: Technology can extend human capabilities but must never replace humans as the principal agent.