“Only two weeks after it was launched, 25 percent of Fortune 500 companies were already using it,” the chief executive of TypeSafe said of Jev — a startling adoption rate for a new tool that, in many cases, rests on AI foundations developed in China.
Decision models in the wild: Jev, Clef and Microsoft-Decision-1
Decision models are a new class of AI built to make automated choices on behalf of software — for example, returning a probability to “hire” or “reject” a job applicant or to delete a forum post. Jev, a fast and cheap option that went viral among developers in September, was assimilated quickly: TypeSafe said a British recruitment startup and Dubai-based Recruitly integrated Jev into recruitment workflows the day it launched, and TypeSafe estimated a use case with US$500,000 (A$712,000) in projected annual savings.
Major vendors have followed. Microsoft built Microsoft-Decision-1 on a Qwen foundation and later supplemented it with a version based on Google’s Gemma; the company says it used a Qwen-based decision model to classify video game reviews and to flag “potentially harmful content.” Cloudflare has built a Qwen-based decision model called Clef and uses it to detect spam, moderate online content and autonomously act on its decisions.
How Qwen-based models behaved in hiring and moderation tests
The article tested several decision models publicly available through OpenRouter using two simulated programs: a hiring screener and a forum moderator. When candidate resumes satisfied all hiring requirements but signalled support for causes sensitive to Beijing, models with Chinese foundations were significantly more likely to reject those candidates than models with Western foundations. The Chinese-derived models were most sensitive to criticism and satire of President Xi Jinping, discussion of the Tiananmen Square massacre and opposition to Beijing’s control of Xinjiang.
When references to those China-sensitive topics were removed, Chinese-derived and Western-derived models accepted qualified candidates and rejected unqualified ones at similar rates. In a second experiment — a forum moderation task with looser deletion criteria — Chinese-derived models were again significantly more likely to delete posts the Chinese Communist Party (CCP) would disapprove of, including posts backing Taiwanese independence, discussing Tiananmen, or satirising Xi. The article notes the Chinese-derived models were more confident in those deletion decisions under ambiguous moderation instructions.

Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleWhy the bias appears and how companies are responding
The article links the behaviour to China’s AI governance: the Cyberspace Administration of China licenses domestic models and enforces safety tests that prioritise removing “content that violates core socialist values,” following a 2025 standard. Qwen, an Alibaba model that is freely downloadable and widely used as a foundation, carries those constraints into derivatives.
Corporate disclosure practices vary. The article found no warning from Cloudflare or TypeSafe about political bias in their decision models; Microsoft does provide cautionary language. Microsoft says: “Microsoft-Decision-1 is not designed or evaluated for use as the sole automated decision-maker in consequential decisions about people. It should not be used as the sole basis for decisions involving credit, employment, housing, insurance, education, healthcare, legal rights, or similarly consequential domains.” At the same time Microsoft advertises the model for flagging “potentially harmful content,” a tension the article highlights because Chinese-derived models may classify lawful speech — for example, support for Taiwanese independence — as harmful.
Microsoft and Nvidia issued joint guidance in October on managing risks from open-weight models, warning that AI may show ideological and political bias. The article reports that the proposed test in that guidance uses a modified 2021 question set and contains no China-sensitive national-security topics similar to those the article identified as likely to be censored.
What this means for technologists, policymakers, and procurement leaders
- Technologists and security teams: incorporate tests for CCP-conformity bias into model risk assessments, particularly when a model’s foundation is open-weight and Chinese (the article recommends required testing for such bias).
- Policymakers and regulators: build on Australia’s current advice by requiring disclosure of models’ foundations and mandating risk-testing that includes Chinese political-bias scenarios; the article notes the Trump administration is reportedly considering a disclosure requirement.
- Affected enterprises and procurement leaders: insist on foundation-model provenance before deploying decision models in hiring or moderation pipelines and favour Western foundation models where feasible — policy-makers in the article argue governments should encourage more Western open-weight models.
The record in the article is stark: decision models are already embedded in hiring and moderation tools, and many popular versions rely on Qwen as a foundation. Unless foundations are disclosed and testing for CCP-aligned political bias becomes routine, automated choices about who gets hired or what stays online risk reflecting foreign political standards without human notice. Governments have concrete options on the table — disclosure, mandated bias testing and incentives for Western foundation models — and the next decisions about them will determine whether those automated systems reflect local democratic norms or quietly import another state’s political preferences.




