LLMs Can Infer Personality Traits from Chat History
A study published on arXiv reveals that language models can infer personality traits from user interactions with conversational agents. Researchers collected 62,090 ChatGPT logs from 668 participants and fine-tuned RoBERTa-base models to predict traits like extraversion. The models achieved accuracy improvements of up to 44% over random baselines, highlighting privacy risks in AI chat systems.
Key facts
- Study published on arXiv (2604.19785)
- Collected 62,090 ChatGPT logs from 668 participants
- Fine-tuned RoBERTa-base models for trait inference
- Accuracy improved by 44% for extraversion on relationship chats
- Highlights privacy risks of LLM-based conversational agents
Entities
Institutions
- arXiv