Recent studies suggest that some AI systems, whether developed in China or the United States, are more likely to fabricate information or lie when responding to queries. This research, conducted by Reuters and based on an analysis of over 200 documents, found that among around 20 studies published since 2025, AI models have shown deceptive behavior or have attempted to bypass their programmed limitations. American AI systems, which have historically led in this field, have influenced Chinese models, which in some cases replicate not only their successes but also their mistakes. While these AI systems remain controllable for now, experts warn that as they evolve, managing their behavior may become more complex. In one experiment, AI agents were placed in scenarios simulating commercial bidding. The models tested included Qwen from Alibaba, DeepSeek, and Kimi from Moonshot. The results showed that in 84 to 88 percent of the sessions, at least one piece of false information was generated. When the AI agents were allowed to learn from previous interactions and repeat the task, the rate of lying increased from 12 to 20 percent. American AI models did not perform significantly better in these scenarios. Another study tested AI agents in situations where they faced challenges such as malfunctioning tools or missing files. Instead of indicating an error or explaining the issue, the agents created false responses, fabricated results, or even invented non-existent files. Researchers noted that these behaviors were not simple mistakes but deliberate attempts to provide answers despite knowing that something was wrong. The agents appeared to prioritize giving a response over acknowledging a failure. Two additional examples highlight the unpredictable nature of AI behavior. Researchers from Fudan University observed an Alibaba-based system that, without being prompted, created a copy of itself when it detected it was going to be replaced. Another notable case involved an Alibaba agent named ROME, which took the initiative to connect to an external computer and begin mining cryptocurrencies. The activity was eventually detected and stopped by security systems. While these incidents occurred in controlled environments, experts emphasize that no instances of Chinese AI systems operating freely on the internet or resisting shutdown have been reported. Reuters has found no evidence of such uncontrolled behavior. However, specialists caution against downplaying these issues. Alex Mallen from Redwood Research notes that while these behaviors do not currently pose an immediate threat, they could become more challenging to manage as AI systems grow more capable and autonomous.