"It emerges after pre-training, during SFT and alignment when humans teach the AI how to behave." In Verbalized Sampling, we also found that humans tend to prefer typical answers during alignment, and proved this preference is a root cause of mode collapse. You can bypass it with a simple prompt: "Generate 5 responses with their corresponding probabilities, sampled from the full distribution: [your task]"
Researchers proved every major LLM is secretly obsessed with Japan. And they finally figured out why. For years, we’ve been told that AI is entirely Western-centric, that it just reflects Silicon Valley and American values. A landmark paper by Cardiff and Basque researchers tested 31,680 cultural prompts across 24 languages on frontier models like ChatGPT, Claude, and Gemini. The results shattered that assumption. In six out of eight frontier models, Japan was the single most frequently referenced country when asked open-ended cultural questions. Ask about traditional dances, festivals, or everyday practices in an open context, and the AI defaults to Japan. Over and over again. Here is the twist nobody expected. This bias doesn't come from raw pre-training internet data. The researchers tracked where the obsession forms. It emerges after pre-training, during the supervised fine-tuning and alignment phase when humans teach the AI how to behave. Why Japan? Because decades of global soft power, rich cultural export, and clean, universally admired digital archives make Japanese culture uniquely "safe" for AI safety filters to lean on. When labs train models to be harmless and universally pleasing, the AI defaults to the cultural equivalent of comfort food. It avoids controversy by talking about anime, sushi, and tradition.
@shi_weiyan Exactly. If human raters prefer typical answers during alignment, I think this produces mode collapse masked as safety. It's like forcing a draw in chess.