DeepSeek’s latest AI model, R1 0528, has raised eyebrows for a further regression on free speech and what users can discuss.