Generative artificial intelligence now sits inside ordinary study, work, and entertainment, yet it still produces answers that sound sure of themselves and turn out to be wrong. In China, where more than 600 million people use these tools, that risk is no longer a rare surprise. Young users notice it, explain it, and keep working with the technology rather than walking away.

A study published in the International Journal of Human-Computer Interaction looks at how young Chinese users recognise AI hallucinations and what they do next. Researchers combined a large reading of social media with interviews. The work covers January 2023 to August 2025, when models such as DeepSeek, Doubao, and Kimi were widely available without a paid subscription.

The researchers collected posts and comments from Douyin, Weibo, and Xiaohongshu, then kept 8,576 entries that clearly concerned artificial intelligence and hallucination. They also interviewed 21 users, 11 women and 10 men, aged 19–34, in July and August 2025. Each conversation lasted 50–90 minutes. About half came from humanities and social science backgrounds, and about half from science and technical fields.

People spotted problems in three ways. Some read the model’s displayed reasoning and the pages it cited. Others trusted their own knowledge, catching a missing name or a line of code that failed when it was run. When the subject was unfamiliar, they checked official sites and databases, or tested the output in a real task. Text errors were harder to see than obvious image mistakes, and invented citations grew more convincing as the models improved.

Attitudes shifted with the job. In creative or low stakes work, some treated odd answers as raw material, or let a small fault pass when a deadline was close. In study and professional tasks, they were quicker to challenge the result. Most would correct a system only a few times before switching tools or finishing the work themselves. Free access made that switching easy.

The interviews produced four everyday explanations, described as algorithmic folk theories. One says hallucinations come from a gap between what a person means and what a short or vague prompt allows the system to infer. A second treats them as proof that current models still cannot judge truth reliably. A third blames weak training data, annotation mistakes, or later contamination by misleading inputs. A fourth points to company rules, political caution, cost saving on free services, and a design that would rather answer than admit it does not know.

Those explanations shaped habits. People wrote clearer prompts, set a role for the system, and added context. They stopped asking a chatbot to produce a whole piece and used it for narrower jobs such as polishing. Some turned off web search, limited the model to documents they supplied, and checked the result against original sources. Others broke large tasks into smaller steps, or opened a fresh chat when answers began to repeat and drift.

Ordinary users and technical users reached similar broad explanations, but not by the same route. People without specialist training drew on daily use and social media, and more often blamed unclear instructions. Technical users drew on courses, papers, and industry material, and more often blamed training data, probability, and system design. They tended to set aside tips on lifestyle platforms, while ordinary users used those discussions to learn that a fake reference was a known limit, not a personal failure.

The Chinese setting shaped the detail. Free access, heavy demand, and national rules on deep synthesis, generative artificial intelligence services, and labels for synthetic content gave people concrete reasons to link errors with cost control and restricted answers. The four theories may travel. The reasons people reach for are tied to local business models and regulation.

The sample was young and mostly highly educated, and the social media search used the technical phrase AI hallucination, so plainer complaints may be under-represented. Even so, the pattern is plain. People are not waiting for a flawless system. They are building working rules for a tool that remains useful and unreliable at the same time.