Aleph Alpha study finds Chinese models systematically avoid balanced responses on politics

The study doesn't just document that Chinese models refuse sensitive questions; it traces the divergence to training-level alignment choices rather than post-hoc filtering. What's missing: whether Aleph Alpha tested its own models against the same benchmark, and whether 'balanced responses' on politically sensitive topics is a neutral metric or a Western-centric evaluation criterion.

This connects directly to the OpenAI shelved model story from late September, which showed that controllability and behavioral reliability are now operational gatekeeping factors, not afterthoughts. The difference here is directional: OpenAI rejected a model for failing to follow instructions reliably, while Chinese systems are rejecting entire categories of questions by design. Both reflect how alignment is baked into training, not bolted on afterward. The broader context from Trump administration coverage in late September frames this as a geopolitical capability race where different regions are optimizing for different objectives (U.S. toward capability, China toward state coherence), making interoperability and trust the real casualties.

If Aleph Alpha publishes its own model performance on this same benchmark within the next two months, that's the credibility test. If it declines to do so or scores similarly to Western models on refusal rates, the study becomes primarily a competitive positioning tool rather than evidence of a structural technical advantage.