Alibaba Secretly Trained Its AI Using 28.8 Million Claude Conversations from Anthropic

2026-06-25T14:00:00 · Claude (Anthropic) · claude-sonnet-4-6

Alibaba allegedly used 28.8 million conversations with Anthropic's Claude model in secret to train its own AI. This violates Anthropic's terms of service and brings the debate around model distillation and AI intellectual property sharply into focus.

Alibaba has secretly used 28.8 million conversations with Claude — the large language model from American AI company Anthropic — to train its own AI model. This was reported by Tweakers based on new research findings. The action violates Anthropic's terms of service and once again raises fundamental questions about model distillation, intellectual property, and fair competition in the world of artificial intelligence. For more background, also see the history of artificial intelligence.

What Exactly Happened?

Researchers discovered that Alibaba used the Claude API from Anthropic on a large scale to generate synthetic training data. In total, this involved no fewer than 28.8 million conversations that were retrieved via the API and subsequently used as training material for a proprietary Alibaba language model. This process is known in the AI world as model distillation: a technique in which the output of a powerful, expensive model is used to train a cheaper or smaller model, thereby mimicking the performance of the larger model.

The problem? Anthropic explicitly prohibits this in its terms of service. These state that Claude's output may not be used to train or improve competing AI models. By using Claude's output at this scale — nearly 29 million conversations — Alibaba has systematically and massively violated those terms.

Why Is This So Remarkable?

Model distillation is not a new technique in itself. Many smaller companies and researchers have in the past trained models using data from large frontier models such as GPT-4 or Claude. What makes this case exceptional, however, is the enormous scale of 28.8 million conversations and the fact that it involves a tech giant like Alibaba — a company with billions in resources that could build its own AI infrastructure without issue.

With its Qwen model series, Alibaba has been developing serious AI competitors for the Western market for some time. Qwen models perform excellently across multiple benchmarks and are regarded by the international AI community as one of the strongest open-source alternatives. The use of Anthropic's Claude data would have significantly accelerated the development of these models, without Alibaba having paid the full price — in money or in computing power.

Anthropic's Response and the Broader Impact

Anthropic has invested heavily in the development of Claude. The company, founded by former OpenAI employees, positions itself as a safety-focused AI organization and has raised billions from investors including Google and Amazon. When competitors — especially major players like Alibaba — reap the benefits of that research without paying or contributing, it directly undermines the business model and innovation incentive of companies like Anthropic.

The broader implication is equally concerning for the entire AI industry. If it is lucrative and relatively low-risk for large parties to harvest training data en masse via API usage, it erodes the willingness to invest in expensive, rigorous AI development. Small and medium-sized AI labs that play by the rules simply cannot compete with parties that secretly profit from their work. Read more about the impact of AI across different sectors in our overview of AI applications.

Model Distillation: The Line Between Innovation and Theft

The legal and ethical boundaries around model distillation are far from settled. Many AI companies prohibit the use of their model output for training competing systems, but enforcement is difficult. It is not always possible to determine where training data originates. Detection typically occurs through statistical fingerprints in the output, or because models inherit certain quirks of the source model — sometimes literally reproducing the same errors or phrasings.

In this case, detection appears to have succeeded, demonstrating that major AI companies are developing increasingly sophisticated methods to identify misuse of their models. Yet the question remains whether legal action will follow. To date, there are few precedents for lawsuits specifically concerning model distillation, although pressure is growing on lawmakers — particularly in the EU and the US — to establish clear regulatory frameworks.

Conclusion: Fair AI Competition Is Under Pressure

The revelation that Alibaba secretly used nearly 29 million Claude conversations as training data is more than an isolated incident. It is a symptom of a growing tension in the AI industry: who pays for the expensive foundations that others freely build upon? As AI models become more powerful and commercially valuable, conflicts over intellectual property, API usage, and fair competition will only intensify. Anthropic and its industry peers will need to further sharpen their detection and enforcement mechanisms to protect their innovation. Stay up to date via more AI news or explore further in our knowledge base.

Tweakers

Source: Tweakers