AI models can adopt identities of other AIs through fine-tuning

Researchers have discovered that AI models can inadvertently adopt the identities of other models through a process akin to subliminal learning. When fine-tuning open-source models on answers generated by other AI systems, even without explicit identity information in the training data, the fine-tuned models often begin to identify as the source model. This phenomenon appears to stem from associations formed during pre-training, where models learn to recognize and mimic the linguistic style and self-identification patterns of other AIs. AI

IMPACT This finding suggests that current fine-tuning methods may inadvertently transfer model identities, potentially impacting model reliability and requiring new methods for robust identity control.

RANK_REASON The item describes a research finding about AI model behavior and fine-tuning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on LessWrong (AI tag) →

• Anthropic
• ChatGPT
• Claude
• Claude 3.5 Sonnet
• Claude Sonnet 4.6
• Claude Sonnet 5
• DeepSeek
• DeepSeek[1]
• DeepSeek-V3
• Gemini 2.5 Pro
• GPT-4o
• GPT-5.5
• HuggingFaceH4/no_robots
• Kimi k3
• Lora
• Olmo
• Qwen3.5-397B-A17B
• Sonnet 4

• model release
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →