Qwen3.5 Hallucinates Tiananmen Safety Context

This looks less like grounded reasoning and more like a model surfacing a brittle, overlearned censorship/safety pattern around a sensitive topic.

• –The model appears to anchor on a politically sensitive prior and then rationalize it with safety language, which is a classic hallucination failure mode.
• –The thread does not prove a hidden chain-of-thought leak or a policy violation; it’s a single user-reported interaction, so reproducibility matters.
• –For developers, the takeaway is that “safe” refusals can still be factually wrong, especially in sensitive-history prompts.
• –If you ship models into production, you need adversarial evals for sensitive topics, not just generic safety filters.
• –This is more about trust calibration than raw capability: users will notice when the model sounds certain while inventing context.