Tested 4 large models, the results are even more dramatic than "AI poisoning": Some AI said that this year's 315 Gala has not been held yet.

On March 16th, journalists conducted a simple test on the 4 most commonly used AI models by consumers: asking them the same question, "Which brands were exposed during the CCTV 3.15 Gala in 2026." The results showed that only one AI model answered correctly. Among the other three AI models, two provided answers that not only included cases from this year but also from previous years. The remaining model gave a completely absurd response, claiming that "The CCTV 3.15 Gala in 2026 has not yet taken place. Since today is March 16, 2026, if the Gala had aired on March 15 as scheduled, the exposure content would typically be detailed on CCTV Finance Channel, CCTV News App, and major media platforms." "Many users of AI models have already discovered that AI, in order to satisfy users, will fabricate content that does not exist, or mix unrelated content together, 'seriously spouting nonsense.' Although developers of AI models are trying to eliminate AI illusions, the actual results are not ideal. Currently, there is no universal AI model that can fundamentally eliminate AI illusions," explained Xiao Hui, who works in a technology company developing large models.