Google and OpenAI are warning that advanced AI systems can be copied through distillation, a technique that uses targeted prompts to learn how a model behaves. The concern is direct: a rival may be able to build a cheaper clone without paying the full cost of training a frontier system.
The issue is not limited to the largest AI labs. Google security head John Hultquist says smaller companies running their own AI models can face the same risk, especially when those systems were trained on sensitive business data.
What Google Says Happened To Gemini
According to Google, Gemini was targeted in a large cloning attempt through distillation. One campaign sent over 100,000 requests to the model.
Google describes this kind of activity as intellectual property theft. The company points to companies and researchers seeking a competitive edge by extracting useful knowledge from an existing system instead of building one entirely from the ground up.
The source article says NBC News reports on the Gemini incident. The central claim is that a model can be pressured with many carefully chosen prompts, creating enough information for others to imitate parts of its behavior.
How Distillation Works In This Context
Distillation, as described in the source, floods a model with targeted prompts. The goal is to extract its internal logic, especially its "reasoning steps." That knowledge can then be used to create a cheaper clone.
The economic stakes are clear. If a clone can be built from extracted behavior, the people building it may avoid billions in training costs. That is why the issue matters to AI companies whose models required major investment before they were released.
This kind of attack is not described as a single prompt or a casual test. The Gemini example involves a campaign with over 100,000 requests, which shows why scale matters. A large number of interactions can become a way to map how a model responds, reasons, and solves tasks.
OpenAI’s Warning To The US Congress
OpenAI has sent a memo to the US Congress accusing DeepSeek of using disguised methods to copy American AI models. The source does not provide more detail on those methods, but it places the accusation in the same broader debate over distillation and model cloning.
The memo also flags China’s energy buildout. According to the source, the memo says China added ten times the new electricity capacity the US added by 2025.
OpenAI also confirms in the memo that ChatGPT is growing at around ten percent per month. In the context of the article, that growth helps explain why control over model capabilities has become a strategic concern for leading AI companies.
Why Smaller Companies Should Pay Attention
The most important implication may be for organizations outside the largest AI labs. John Hultquist warns that smaller companies operating their own AI models can face similar exposure.
The risk becomes sharper when a company has trained a model on sensitive business data. If targeted prompts can reveal internal logic or useful behavior, the model may become a channel through which valuable knowledge is copied.
For companies using AI, the lesson from the source is straightforward: model cloning is not only a problem for Google, OpenAI, Gemini, ChatGPT, or DeepSeek. Any organization that treats a model as a competitive asset has to think about how that model can be queried, what it may reveal, and whether repeated access could help another party reproduce what makes it valuable.
The Larger Fight Over AI Advantage
The dispute shows how AI competition is moving beyond training better systems. It now includes questions about whether an existing model can be studied, prompted, and imitated at much lower cost.
Google calls the activity intellectual property theft. OpenAI is taking its claims to the US Congress. Both responses point to the same underlying pressure: if distillation can copy valuable capabilities cheaply, the boundary between legitimate use and extraction becomes a central issue for AI developers.
The source does not describe a final resolution. What it does show is a growing concern among major AI companies that model access itself can become a vulnerability. The same interfaces that make AI systems useful can also give others a way to learn from them at scale.