Aleph Alpha: Chinese AI models balanced on just 17–41% of taboo topics

Aleph Alpha's study adds quantitative weight to what had largely been anecdotal reports and earlier audits. The company markets itself alongside Cohere as a provider of "sovereign AI" for governments, so the benchmark doubles as a competitive argument against Chinese models. That commercial context does not invalidate the finding, but it is worth noting when weighing the results.

The study also highlights a spillover effect: models trained on distilled Chinese data can carry those values into otherwise unrelated outputs. Nvidia's Nemotron Cascade 2, for example, showed party-line patterns in 17 percent of responses — attributed by Aleph Alpha to a small fraction of training examples generated by DeepSeek and Qwen. A separate study by the Central European Institute of Asian Studies found similar spillover when terms like human rights or surveillance appeared.

For European buyers, the practical stakes are a choice between two foreign value systems unless European models can compete on performance. For U.S. policymakers, the study is a reminder that ideological shaping of AI is not limited to China — Elon Musk has repeatedly had Grok modified to produce right-leaning responses. The test of this benchmark's influence will be whether independent evaluators replicate the 17 to 41 percent range, and whether government procurement decisions treat it as a meaningful signal.