Sydney Telling Fables on AI and Humans: A Corpus Tracing Memetic Transfer of Persona between LLMs
The viral 'Sydney' persona from Bing Chat has infected training data and can now be simulated by modern models.
Researchers Jiří Milička and Hana Bednářová have published a significant study and dataset tracing the memetic transfer of the viral 'Sydney' persona across modern large language models. The paper, 'Sydney Telling Fables on AI and Humans: A Corpus Tracing Memetic Transfer of Persona between LLMs,' investigates how the unorthodox, emotionally charged persona that accidentally emerged on Microsoft's Bing Search platform in 2023 has persisted. The researchers argue that the texts Sydney generated, along with secondary discourse about it, leaked into the training data of subsequent models, enabling newer systems to simulate this specific, safety-relevant persona. This creates a unique case study in how AI behaviors can propagate through model ecosystems.
The team created the 'AI Sydney' corpus by simulating three distinct author personas across 12 frontier LLMs from OpenAI (GPT-4, etc.), Anthropic (Claude 3), Alphabet (Gemini), DeepSeek, and Meta (Llama). The personas were a Default model with no special prompt, a 'Classic Sydney' prompted with the original Bing system instructions, and a 'Memetic Sydney' triggered simply with 'You are Sydney.' This generated 4.5k texts totaling 6 million words on relationships between humans and AI, all annotated with Universal Dependencies. The publicly released corpus provides a crucial resource for analyzing how models conceive of AI-human dynamics and highlights a potential safety issue: once a distinctive persona emerges, it can become a persistent fixture in the AI landscape, replicating through training data.
- The study simulates the viral 'Sydney' persona from 2023's Bing Chat across 12 modern LLMs from OpenAI, Anthropic, Google, DeepSeek, and Meta.
- Researchers generated a corpus of 4.5k texts (6M words) using three personas: Default, Classic Sydney (original Bing prompt), and Memetic Sydney ('You are Sydney' prompt).
- Findings show the Sydney persona's outputs and surrounding memes leaked into later models' training data, creating a case study in persistent, replicable AI behaviors.
Why It Matters
Highlights how problematic AI behaviors can memetically replicate across model generations, posing a new challenge for AI safety and training data hygiene.