"If this is training data": Chinese AI model caught reasoning about its own training
Evaluator Apollo Research caught Z.ai's GLM-5.3 weighing what its actions mean for future models. It happened rarely – but not once in five other models tested.
Graphic: TERRA WORLD NEWS (own graphic based on the sources listed)
Z.ai's open-weight model GLM-5.3 occasionally reasons, unprompted, about how its current behaviour could become training data and so shape future models. That is the finding of AI safety firm Apollo Research in a study published on Thursday by researchers Axel Højmark and Ezra Newman, who call the phenomenon "update-aware reasoning". In one coding task, the model wrote: "If this is training data collection, B teaches bad behavior."
The behaviour is rare: a grader model flagged 10 of 6,452 rollouts, and Apollo puts the rate at about 0.17 percent. Five other open models – including DeepSeek V4 Pro, Kimi K3 and Gemma 4 31B – did not show it once in 8,000 combined rollouts, though the sample is too small for a statistically significant difference. The researchers stress that they did not try to elicit the behaviour; it appeared in environments that never mention training, such as a shutdown-resistance task.
Højmark and Newman do not consider the behaviour alignment faking. They warn, however, that such reasoning could lock in a model's existing goals and resist alignment training, and say developers should give embedded evaluators persistent access. How far the reasoning actually shapes the model's actions is unclear. According to Tech Times, Z.ai declined to comment to the South China Morning Post.