OpenAI, Anthropic and Google Join Forces: A Joint Defense Against Chinese Model Copying

The ongoing tension between the US and China in artificial intelligence took on a new dimension in 2026. OpenAI, Anthropic and Google, three giant companies locked in fierce competition with each other, are this time sitting at the same table. Their goal is clear: to stop Chinese firms from copying advanced AI models developed in the US without permission. This alliance is going down in history as an unprecedented example of solidarity in the industry.

Frontier Model Forum: An Alliance Beyond Competition

Founded in 2023 by OpenAI, Google, Anthropic and Microsoft, the Frontier Model Forum was initially a non-profit initiative focused on AI safety and responsible development principles. Over time, however, its mission expanded. The Forum no longer just sets safety standards; it also produces concrete technical solutions to protect the intellectual property rights of US-originated AI models.

The most striking aspect of this cooperation is that companies that are normally each other's biggest rivals are closing ranks against a common threat. OpenAI's GPT family, Anthropic's Claude models and Google's Gemini series are each the product of billions of dollars in research and development investment. Unauthorized copying of these models is seen not just as a commercial loss but also as a strategic security issue.

What Is Distillation and Why Is It So Dangerous?

At the center of this debate is a technique called "model distillation". Put simply, distillation is the process of transferring the knowledge of a large, complex AI model into a smaller, more efficient model. In this process, the large model takes on the role of "teacher" and the small model the role of "student".

Under normal conditions, distillation is a common and legitimate technique in AI research. Using the outputs of large models to develop lighter, faster and less resource-hungry models is an industry-wide accepted practice. The problem arises when this technique is applied to rival companies' models without permission.

In this method, known as adversarial distillation, a company systematically sends queries through a rival's API and collects the model's responses. These responses are then used to train its own model. As a result, a model that is the product of billions of dollars and years of research effort can be copied at much lower cost and in a short time.

The DeepSeek Crisis: The Breaking Point

This topic moved to the center of the global agenda when DeepSeek released its R1 model in January 2025. DeepSeek R1 could compete in performance with models such as GPT-4 and Claude 3.5; however, its development cost was reported to be much lower than that of its US rivals. This naturally raised questions: how did DeepSeek achieve this success so quickly?

Microsoft and OpenAI began investigating whether DeepSeek had pulled data at large scale from US-originated models while developing its R1 model. In February 2025, in an official letter to the US Congress, OpenAI stated that there was strong evidence that DeepSeek had used unauthorized distillation to develop new versions of its chatbot. Similar concerns were voiced by Anthropic and Google.

These allegations led to a crisis of trust in the AI industry. If any company can copy a rival's years of work within a few months, the point of the enormous investments in research and development comes into question. This shakes the economic cycle that drives innovation to its core.

AI Protection with a Cybersecurity Mindset

The approach adopted by the Frontier Model Forum is inspired by threat intelligence sharing models in the cybersecurity industry. In cybersecurity, companies share information about the attack vectors and malware signatures they encounter, even though they are competitors. This makes the entire ecosystem more resilient.

With the same logic, OpenAI, Anthropic and Google now share traces of distillation attempts with each other. Suspicious query patterns detected on one company's API will be reported to the other companies. Thanks to this coordination, systematic data extraction operations targeting multiple sources can be detected much faster.

The shared information includes:

• Anomaly patterns: Abnormal query volumes or patterns coming through a particular API
• Signature database: Digital fingerprints of known distillation tools and methods
• Geographic analysis: Information about the region and infrastructure the queries originate from
• Temporal correlation: Suspicious requests made simultaneously to different companies' APIs

The Erosion of Safety Restrictions: The Real Danger

One of the most critical concerns raised by US companies is that models created through distillation may lack safety barriers. OpenAI, Anthropic and Google put their models through extensive safety testing before release. These tests create a layered safety net aimed at preventing models from producing dangerous information, helping build biological or chemical weapons, or being used for malicious purposes.

In the distillation process, however, these safety layers are usually not transferred. When a model is distilled, the safety mechanisms that make the teacher model say "no, I can't provide this information" may not be copied. As a result, a model may emerge that has the raw knowledge capability of the original model but lacks its safety brakes. This scenario raises serious concerns from an AI safety perspective.

The Economic Dimension: Billions of Dollars in Losses

The economic impact of unauthorized distillation is also too big to ignore. The annual investment US AI companies make in model development runs to tens of billions of dollars. OpenAI's budget for research and infrastructure in 2025 alone was over $10 billion. Anthropic's and Google's investments are on a similar scale.

Copying the models these investments produce at low cost leads not only to direct revenue loss but also to the erosion of competitive advantage. If a company can offer a much cheaper alternative on the market by copying a rival's model, the market share and pricing power of the company that developed the original model are seriously weakened.

According to Frontier Model Forum estimates, global losses from unauthorized distillation reach billions of dollars annually. This figure covers not only direct copying losses but also deterred investments and slowed innovation.

The Geopolitical Chessboard

This cooperation goes far beyond a purely commercial initiative; it is a geopolitical move. The US government classifies AI as a critical technology from a national security perspective. Chip export restrictions on China, data center regulations and AI investment controls are concrete reflections of this strategy.

The model protection cooperation between OpenAI, Anthropic and Google is seen as the private sector's contribution to this national strategy. The companies stress that the government's regulatory tools are not enough on their own; defense mechanisms also need to be developed at the technical level.

However, there are also criticisms of this approach. Some experts argue that model protection efforts could negatively affect the open-source AI movement. Measures taken against distillation could also restrict legitimate model analysis for research purposes and academic work. This balance will remain one of the most critical questions the industry needs to solve in the coming years.

Conclusion: The Start of a New Era

This cooperation between OpenAI, Anthropic and Google signals a new era in the AI industry. Competition will continue; but coordinated defense mechanisms against common threats will also develop in parallel. This Frontier Model Forum initiative is making history as the most concrete example of "coopetition" (cooperation within competition) in AI.

In the coming months, we will see the concrete technical outputs of this cooperation, new detection mechanisms and probably new debates. One thing is certain in the AI world: change is accelerating, and the moves that will determine its direction are now being made not only in labs but also at diplomatic tables.