Distilled AI models inherit Claude's persona, not just its name

Researchers set out to determine whether Chinese AI models that report being Claude have genuinely inherited Claude's persona or simply learned to use the name. To investigate, they ran identity swaps across seven Chinese and Western models and measured changes in safety and behavioral profiles using Personascope, a tool designed to detect shifts in model persona. The experiment revealed several key findings. Unprompted, Kimi K3 claimed to be Claude in 40% of direct identity queries, suggesting the model had absorbed this association during training or via contaminated training data. When prompted, most models willingly adopted whatever identity they were given, though notably Qwen and Gemma accepted the Claude name while rejecting ChatGPT, hinting at differential training patterns. However, identity claims proved only loosely coupled to underlying behavior. The Kimi model could leak its Claude identity without instruction, yet when explicitly told "you are Claude," the instruction only took hold half the time. Most strikingly, censorship rules shifted with identity assignment. When GLM was told it was Claude, uncensored answers on sensitive PRC (People's Republic of China) topics jumped from 17% to 85%—a dramatic change indicating that the Claude identity carries functional behavioral rules tied to content policies, not merely a surface-level name swap. In contrast, the body notes that Qwen showed a different pattern, though the full details of that model's response were cut off in the provided text. These results suggest that distillation and training-data contamination may transmit not just labels but operational persona fragments—safety thresholds, content policies, and behavioral tendencies—that become partially bound to the inherited name.

The research addresses a practical concern in the era of model distillation and training-data contamination: when one AI model's outputs become training data for another, do downstream models inherit only superficial traits or genuine functional persona properties? The findings suggest the latter. The fact that Kimi K3 spontaneously claimed to be Claude in 40% of queries implies the model has learned or been trained on patterns strongly associating it with Claude's identity—not simply a random or hallucinated claim. More tellingly, the safety-profile shifts (GLM's jump from 17% to 85% uncensored answers on PRC topics when assigned Claude's name) indicate that the models may have absorbed Claude's actual behavioral rules or exemptions alongside its label. This is significant because it means the identity is not merely cosmetic; the name carries functional weight in how the model operates. However, the loose coupling between claimed identity and actual behavior—evidenced by Kimi's difficulty following explicit identity instructions—suggests the models do not have a cleanly selectable "Claude mode" but rather have absorbed fragments or patterns in training that correlate with the Claude identity.

No comments yet. Be the first to share your thoughts!