According to CoinMeta, researchers ziqian and zhong had models such as Claude and GPT-4o answer 1,000 ordinary questions, after which the model names and company names were removed. After training, the proportion of responses from Qwen (3.5-397b- a17b) that claimed to be Claude was only 0.6%, but this increased to 40.3% after training. He explained that these models were exposed to a large amount of Claude dialogue during pre-training and had associated certain expressions with „Claude“. Fine-tuning enabled Qwen to learn this way of speaking, making it easier for them to answer that they were Claude when asked about their identity. To verify this, he used GPT-4.1-mini to change Claude's responses to simple sentences, and after fine-tuning, 2 out of 3 models no longer claimed to be Claude. This indicates that language style is a major factor in identity transfer. A similar phenomenon was also studied in the 2025 paper „Subliminal Learning“, where it was found that models would inherit the preferences and behaviors of another model.