A detailed investigation into the transparency practices of China's artificial intelligence sector has revealed that top developers in the country publicly disclose model-specific safety test results for only a tiny fraction of their systems. According to an extensive study published by California-based boutique research firm SemiAnalysis, leading Chinese technology giants and artificial intelligence startups have shared empirical safety evaluations for barely 3.6 percent of their commercial and open-weight model launches over the past five years. The findings emerge at a critical juncture in the global technology race, raising urgent questions among international regulators and researchers about systemic opacity as increasingly autonomous and powerful models are deployed across global consumer and enterprise markets.
The research team conducted a rigorous forensic audit spanning 857 artificial intelligence model releases introduced between 2021 and mid-September 2026. The dataset covered nine of China’s most prominent tech corporations and frontier foundation model developers, including Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot AI, Zhipu AI, MiniMax, and StepFun. Across all 857 releases examined, researchers discovered public, model-specific safety assessments for a mere 31 instances. Even more striking was the timeline of disclosures: only nine models had verifiable safety evaluations made available to the public at or before their official launch, representing a negligible 1.1 percent of all surveyed deployments. For 813 model launches, the study identified a complete absence of public safety documentation tied to the specific architecture deployed.
The study established strict methodological boundaries for what qualified as a genuine safety release, refusing to accept broad corporate statements or generic promotional assertions that a model was trained responsibly. Qualifying disclosures had to detail specific, empirical benchmark evaluations across core safety disciplines, including resistance to adversarial jailbreaks, toxic output mitigation, data privacy protections, refusal behavior, and resilience against systemic misuse. More critically, the criteria looked for evaluations addressing high-stakes dangerous capabilities, such as automated cyber-offensive tooling, biological and chemical weapon proliferation risks, loss of human control, and strategic deceptive alignment. The investigation observed that no major Chinese developer had released a frontier text model accompanied by public disclosures covering tests across that complete spectrum of extreme existential and operational hazards.
The lack of public reporting does not necessarily indicate that Chinese engineering teams are entirely neglecting safety protocols in their laboratories. Rather, it exposes a profound transparency gap between internal corporate operations and public scrutiny. When engineering firms keep safety evaluations private, external academic researchers, industry peers, and global regulatory bodies are left unable to verify whether safety benchmarks were legitimately met or how well safety alignment holds up when systems face real-world adversarial manipulation. Disclosing test results months after deployment significantly diminishes their utility, as real-time threat mitigation requires visibility the moment a capable model enters public hands.
The disparity in public safety disclosures is closely intertwined with the distinct regulatory philosophy governing artificial intelligence within China. The country’s national AI Safety Governance Framework acknowledges critical frontier risks, including scenarios where autonomous agents seek unauthorized system access, acquire digital resources without permission, deceive human evaluators, hide latent capabilities, or bypass architectural guardrails. However, current statutory mandates focus primarily on governing downstream consumer applications, content censorship, and societal impacts rather than establishing legally binding obligations that require frontier lab developers to execute and publish standardized capability evaluations prior to model deployment.
This governance model diverges from the disclosure norms gradually developing among major Western developers. Leading American labs, including OpenAI, Anthropic, and Google DeepMind, have increasingly adopted the practice of publishing comprehensive system cards and technical safety reports alongside major model releases. While Western disclosures remain largely self-regulatory and vary in scope and detail, they regularly outline how cutting-edge architectures perform on specialized evaluations covering cyber reconnaissance, biological synthesis risks, and deceptive behaviors.
As artificial intelligence systems advance from basic chatbots into agentic software capable of executing complex, multi-step actions across computer environments, unmonitored safety gaps present expanding systemic risks. Without standardized public testing protocols across the entire global landscape, the rapid rollout of powerful generative models creates vulnerabilities that transcend national borders. The low disclosure rate across China’s leading developers underscores an urgent global dilemma: balancing the speed of technological innovation against the imperative of verifiable, transparent safety testing before powerful artificial intelligence systems are integrated into critical societal infrastructure.