When Does AI Distillation Become Misappropriation?

The recent allegations by Anthropic against Alibaba raise a legal question that may soon become one of the most important issues in AI law:

Can the capabilities of an AI model be legally protected, even when the model weights and source code remain inaccessible?

Anthropic alleges that operators linked to Alibaba’s Qwen AI lab created thousands of fraudulent accounts and conducted millions of interactions with Claude in order to extract and replicate valuable model capabilities through large-scale distillation. If proven, this case goes far beyond a simple breach of terms of service. It touches upon the emerging legal boundaries between legitimate AI training, competitive intelligence, and unlawful capability extraction.

From a legal perspective, several potential claims may arise:

1. Breach of Contract and Terms of Service

The strongest immediate claim is likely contractual. If access restrictions, geographic limitations, anti-scraping clauses, or anti-distillation provisions existed within Anthropic’s Terms of Service, the alleged use of fake accounts could constitute a direct contractual violation.

2. Fraud and Circumvention of Access Controls

The reported use of approximately 25,000 fraudulent accounts could transform the matter from a simple contractual dispute into a potential fraud and unauthorized access case. Courts may increasingly view large-scale account evasion as deliberate circumvention of technological protection measures.

3. Trade Secret Misappropriation

Perhaps the most important legal question concerns trade secrets.

Traditional trade secret law protects confidential information that derives economic value from secrecy. The challenge here is that no source code or model weights were allegedly stolen. Instead, the dispute concerns extraction of behavioural capabilities through repeated querying.

Future courts may need to determine whether systematically reproducing a frontier model’s reasoning abilities can amount to trade secret misappropriation even when only outputs are observed.

4. Copyright Law Remains Uncertain

Copyright may offer limited protection. In many jurisdictions, individual AI outputs may not qualify for copyright protection, and copyright traditionally protects expression rather than functionality.

This case therefore highlights the growing mismatch between classical copyright doctrines and modern AI systems.

5. The Rise of “AI Capability Rights”

The most significant long-term consequence may be regulatory.

Governments may eventually introduce new legal categories specifically designed to protect frontier AI capabilities from large-scale extraction and replication. Such rules would resemble a hybrid of trade secret protection, database rights, and cybersecurity regulations.

The Anthropic-Alibaba dispute may therefore become an early test case in defining whether AI capabilities themselves constitute a protectable legal asset.

Ultimately, the dispute is not merely about one company allegedly copying another. It is about whether the law will recognize a new form of intellectual asset: the emergent capabilities of advanced AI systems.

References:

• Reuters, Anthropic says Alibaba illicitly extracted Claude AI model capabilities (24 June 2026)
• CNBC/Bloomberg reporting regarding Anthropic’s allegations of large-scale model distillation involving approximately 28.8 million interactions and 25,000 accounts.
• Anthropic, Detecting and Preventing Distillation Attacks (February 2026).

CEO of Anthropic Dario Amodei attends a working lunch with G7 leaders, G7 outreach partners and global tech CEOs on innovation and AI, during the G7 Summit on June 17, 2026, in Évian-les-Bains, France.

Anna Moneymaker | Getty Images News | Getty Images