Chinese AI Firms Allegedly Used OpenAI Data for Model Training
DeepSeek and Moonshot AI reportedly utilized OpenAI's GPT-6 Astra outputs for model distillation.
1Distillation Practices Revealed
Chinese developers DeepSeek and Moonshot AI reportedly utilized outputs from OpenAI’s GPT-6 Astra to train their latest large language models. This practice, known as distillation, allows firms to replicate advanced AI capabilities more efficiently. The report highlights a significant intersection between proprietary model development and competitive data harvesting strategies in global AI.
2Specific Models Impacted
Moonshot AI’s recently released Kimi K3.1 Pro and DeepSeek’s upgraded V4.1-Flash 1017, launched on October 17, are identified as beneficiaries of this distillation process. These models represent the latest iterations from both companies, aiming to compete with Western counterparts by leveraging high-quality output data from established, powerful foundation models like Astra.
3OpenAI’s Prior Identification
OpenAI publicly addressed this issue in a September 30 blog post, explicitly linking a distillation campaign that began in July to individuals associated with Moonshot AI. This confirmation suggests that the practice has been under scrutiny by OpenAI for several months, marking a clear escalation in the monitoring of model usage.
4Industry-Wide Security Concerns
Leading AI firms including Anthropic, Google, and OpenAI argue that distillation enables rivals to reproduce sophisticated capabilities at a fraction of the original cost. This creates a strategic disadvantage for companies investing heavily in primary research, as competitors can bypass the expensive training phase by harvesting outputs from existing models.
5Anthropic’s Defensive Measures
Anthropic recently implemented new security measures aimed at preventing rivals from copying Claude AI capabilities. The company reported that these interventions have successfully reduced data harvested through distillation attacks by approximately 90% to 99%. This demonstrates the technical difficulty of securing model outputs against persistent scraping and unauthorized training data acquisition.
6Geopolitical Technology Competition
The practice of distillation has become a flashpoint as US-China technology competition intensifies. With restrictions on hardware and software exports, Chinese firms are increasingly focused on optimizing their development cycles. Distilling high-tier models provides a shortcut to achieving competitive performance levels without requiring the same massive compute infrastructure as Western leaders.
7Lack of Official Responses
OpenAI, DeepSeek, and Moonshot AI did not provide immediate comments regarding the reports. The silence from these organizations reflects the sensitivity of the allegations, which touch upon intellectual property rights and the ethical boundaries of training data procurement in an era where model outputs are increasingly treated as proprietary assets.
8Future Regulatory Implications
The reliance on distillation highlights a critical vulnerability in current AI development ecosystems. As companies implement stricter defensive measures, the industry may see a shift toward more closed-loop systems. Future regulatory frameworks will likely need to address whether using model outputs for training constitutes a violation of terms of service or copyright.