Aleph Alpha Study Finds Chinese AI Models Give Balanced Answers on Sensitive Topics Only 17-41% of the Time

Chinese large language models frequently repeat official state positions or decline to answer outright when asked about politically sensitive subjects, according to an evaluation published by German AI company Aleph Alpha. The study scored only 17 to 41 percent of the tested models' responses as balanced, meaning the large majority either echoed party doctrine or refused to engage. The results add hard numbers to a long-running anecdotal complaint about the output of models built and hosted inside China.

Aleph Alpha evaluated several Chinese models rather than a single system, and the wide 17-41 percent band suggests the problem varies substantially by vendor and by topic. Questions touching on Taiwan, Xinjiang, the 1989 Tiananmen Square crackdown, and the political status of Tibet are the likely pressure points, though the company's published framing groups them under the broader label of politically sensitive prompts. Refusals counted against a model in the scoring just as biased answers did, which is a meaningful methodological choice: a model that stays silent is not treated as neutral.

The study's authorship matters as much as its findings. Aleph Alpha is a Heidelberg-based firm that sells what it calls sovereign AI to European governments and public-sector buyers, a pitch built explicitly on the idea that model provenance and training data determine whose interests a system serves. A report showing Chinese models aligned with Beijing's positions gives the company a direct commercial argument for its own products. That does not make the numbers wrong, but readers should weigh them as the output of an interested party rather than a neutral academic lab.

The evaluation lands as European institutions wrestle with which AI systems to allow into public procurement. Several national and EU bodies have moved toward requiring transparency about training data and model governance, and dependence on US hyperscalers has become its own political concern in Brussels and Berlin. Aleph Alpha's framing tries to position European vendors as the third option: not American, not Chinese, and accountable to local law. Whether procurement officials accept that framing is a separate question from whether the bias measurements hold up.

There is no public evidence yet that Aleph Alpha's methodology has been independently replicated, and the study does not appear to name a confidence interval or sample size in the coverage so far. The obvious next step is third-party replication using the same prompt set on the same model versions, ideally by researchers with no stake in the sovereign-AI market. Until that happens, the 17-41 percent figure is a useful data point and a sales document at the same time.