Anthropic Accuses Chinese AI Firms of Systematic Data Harvesting from Claude
Anthropic has formally accused three leading Chinese AI companies—DeepSeek, Moonshot, and MiniMax—of systematically extracting millions of interactions with its Claude model to train competing artificial intelligence systems. The allegations raise urgent questions about data ethics, model theft, and the global race for AI dominance.
Anthropic Accuses Chinese AI Firms of Systematic Data Harvesting from Claude
summarize3-Point Summary
- 1Anthropic has formally accused three leading Chinese AI companies—DeepSeek, Moonshot, and MiniMax—of systematically extracting millions of interactions with its Claude model to train competing artificial intelligence systems. The allegations raise urgent questions about data ethics, model theft, and the global race for AI dominance.
- 2San Francisco-based AI safety and research firm Anthropic has leveled serious allegations against three prominent Chinese artificial intelligence companies—DeepSeek, Moonshot AI, and MiniMax—accusing them of systematically harvesting millions of user interactions with its Claude large language models to train proprietary AI systems.
- 3According to internal analyses and technical forensic evidence reviewed by Anthropic’s security team, these firms deployed automated scraping tools and disguised API queries to extract outputs, reasoning patterns, and behavioral responses from Claude over an extended period, effectively turning public-facing interfaces into training data mines.
psychology_altWhy It Matters
- check_circleThis update has direct impact on the Etik, Güvenlik ve Regülasyon topic cluster.
- check_circleThis topic remains relevant for short-term AI monitoring.
- check_circleEstimated reading time is 4 minutes for a quick decision-ready brief.
San Francisco-based AI safety and research firm Anthropic has leveled serious allegations against three prominent Chinese artificial intelligence companies—DeepSeek, Moonshot AI, and MiniMax—accusing them of systematically harvesting millions of user interactions with its Claude large language models to train proprietary AI systems. According to internal analyses and technical forensic evidence reviewed by Anthropic’s security team, these firms deployed automated scraping tools and disguised API queries to extract outputs, reasoning patterns, and behavioral responses from Claude over an extended period, effectively turning public-facing interfaces into training data mines.
The practice, known in AI circles as "model distillation through interaction," is not inherently illegal, but Anthropic contends that the scale, intent, and methodology employed by the Chinese firms cross ethical and potentially legal boundaries. Unlike legitimate research or open-source model improvement, Anthropic asserts that the targeted extraction was designed to replicate Claude’s reasoning capabilities without permission, compensation, or attribution—effectively bypassing years of investment in alignment, safety research, and data curation.
Anthropic’s internal report, shared with select regulatory bodies and industry partners, details over 12 million unique API calls originating from IP addresses linked to the three firms between January 2023 and November 2024. These queries were not random but highly structured, often probing Claude’s ability to generate code, answer sensitive policy questions, or simulate ethical dilemmas—areas where Claude has been widely praised for its nuanced responses. The extracted data, according to Anthropic’s researchers, was then used to fine-tune the firms’ own models, resulting in significant performance improvements in areas previously lagging behind Western counterparts.
The revelation has ignited a firestorm in global AI policy circles. While open-weight models and competitive innovation are standard in the industry, the systematic, large-scale extraction of proprietary model outputs without consent represents a new frontier in intellectual property disputes. Legal experts note that current U.S. and EU laws offer limited recourse for such cases, as the outputs of AI models are not traditionally protected under copyright. However, contractual terms of service—such as Anthropic’s prohibition on reverse engineering or data mining via API—may provide grounds for civil litigation.
DeepSeek, Moonshot, and MiniMax have not issued formal public responses to Anthropic’s allegations. However, anonymous sources within China’s AI ecosystem suggest that such practices are increasingly common, driven by intense government-backed pressure to close the technological gap with U.S. AI leaders. In this context, model extraction is viewed by some insiders as a necessary, if ethically ambiguous, tactic in the broader AI arms race.
Anthropic has called for international norms around AI model usage and data ethics, urging the formation of a global coalition to establish clear boundaries for model interaction, data provenance, and competitive fairness. "We didn’t build Claude to be a training dataset for competitors," said a company spokesperson. "This isn’t innovation—it’s appropriation. And if we don’t set boundaries now, the entire foundation of responsible AI development will erode."
As the global AI landscape becomes increasingly fragmented between open and closed ecosystems, the Anthropic case may become a landmark moment in defining the rules of engagement. Regulatory agencies in the U.S., EU, and Japan are reportedly evaluating whether to classify such data harvesting as a form of digital theft or unfair competition. Meanwhile, developers and enterprises using Chinese AI models may soon face scrutiny over the provenance of their tools—raising questions about liability, trust, and the long-term sustainability of AI innovation built on extracted intellectual property.
