- OpenAI, Anthropic, and Google began coordinating in April 2026 to detect and block Chinese AI developers from systematically extracting outputs from their frontier models, according to Bloomberg.
- The technique at issue—model distillation—allows a competitor to collect API input-output pairs at scale and use them as supervised training data, substantially reducing original R&D costs.
- The collaboration marks an unusual alignment between three companies that otherwise compete directly for enterprise and consumer AI customers.
- All three companies already prohibit API output use for training competing models in their terms of service, but individual enforcement has proven difficult to sustain at scale.
What Happened
OpenAI, Anthropic, and Alphabet’s Google have begun working together to counter Chinese AI developers who systematically extract outputs from US frontier AI models and use those outputs to train competing systems, Bloomberg reported on April 6, 2026. The three companies, which compete directly across the frontier AI market for enterprise API contracts and consumer products, are coordinating specifically to address what Bloomberg describes as a structural practice through which Chinese competitors acquire capabilities by querying commercial US APIs at scale rather than independently developing comparable models.
The collaboration is described as already underway. No formal joint announcement has been made by any of the three companies.
Why It Matters
The practice of using a proprietary model’s API outputs as training data became a prominent point of contention in early 2025, when OpenAI alleged that Chinese AI startup DeepSeek had used outputs from OpenAI’s models in developing its R1 reasoning model—a system that demonstrated performance on coding and mathematics benchmarks comparable to OpenAI’s o1 at a fraction of the reported training cost. DeepSeek disputed the characterization, and no legal action followed, but the episode illustrated how difficult it is for a single company to police API misuse unilaterally against actors who can distribute queries across accounts and jurisdictions.
The challenge is compounded by US export controls that restrict Chinese developers’ access to Nvidia’s most advanced accelerators, raising the cost of frontier-scale training runs from scratch and increasing the relative value of output extraction as an alternative capability acquisition strategy.
Technical Details
Model output extraction works by systematically querying a commercial API—submitting large volumes of designed prompts—and recording the full model responses. Those input-output pairs are then used as supervised fine-tuning or knowledge distillation training data for a separately developed model, transferring behavioral patterns and apparent task performance from the frontier model to the derivative one without replicating the underlying architecture or training data.
OpenAI’s published usage policy explicitly bars customers from using its outputs “to develop any AI model that competes with OpenAI’s products and services.” Anthropic’s acceptable use policy contains parallel restrictions on training competing models using Claude outputs. Google’s Gemini API terms include equivalent language. Enforcement, however, has depended on detecting automated extraction signatures in API traffic—patterns that can be obscured by throttling request rates, rotating API keys, or routing queries through intermediary services.
Bloomberg’s reporting indicates the joint effort is focused on the detection layer, though the specific technical mechanisms being developed or shared across the three companies have not been publicly disclosed.
Who’s Affected
The primary targets of the enforcement effort are Chinese AI developers building commercial or state-adjacent products using extracted outputs from US frontier APIs. Both well-funded private labs and research organizations operating under China’s national AI development programs fall within the scope of the described activity.
Enterprise customers of OpenAI, Anthropic, and Google who rely on high-volume API access for legitimate applications—including data pipeline operators, application developers, and testing infrastructure operators—may face additional verification requirements or enhanced usage monitoring as the three companies deploy more aggressive detection systems. Developers outside China who generate repetitive or high-volume query patterns may also draw automated scrutiny.
What’s Next
Bloomberg’s report describes the collaboration as ongoing, suggesting coordinated enforcement mechanisms are being implemented ahead of any public announcement. Whether the effort will extend to policy advocacy—such as pushing for regulatory frameworks governing cross-border API misuse as a form of intellectual property or national security concern—has not been reported.
The US government has separately pursued export controls targeting advanced AI chips sold to China, most recently through restrictions on Nvidia’s H20 and equivalent products announced in early 2025. A private-sector enforcement layer targeting software-level capability transfer via API output extraction would represent a distinct and complementary approach to restricting Chinese competitors’ access to US frontier AI capabilities.
