Chinese AI firms accused of using OpenAI data to train new models

The allegations of data copying

Months after the release of GPT-6 Astra, a new report has surfaced regarding how Chinese AI developers are building their systems. The report suggests that two major companies, Moonshot AI and DeepSeek, relied heavily on outputs from OpenAI's models to develop their own technology. This development has sparked a significant debate regarding the ethics and legality of how artificial intelligence is trained in a competitive global market.

According to reports from The Times of India, the specific models involved in this alleged activity are Moonshot AI’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017. These models were launched shortly after GPT-6 Astra debuted in September. The core of the issue lies in how these companies obtained the information necessary to make their systems perform at high levels without doing all the foundational work themselves.

The situation has become a major flashpoint between global tech labs as competition between the United States and China intensifies. American tech giants are now looking closely at how their proprietary data is being used by international rivals. This tension is creating a complex environment where the boundaries of fair competition and intellectual property are being tested by the rapid pace of AI development.

How model distillation works

The technique used in these cases is known as model distillation. In this process, a smaller or newer model is trained using the synthetic responses generated by a much larger and more sophisticated model. Essentially, the smaller model learns to mimic the behavior and intelligence of the larger one by studying its outputs rather than learning from raw data alone.

This method is highly efficient because it allows developers to achieve high-quality results much faster than traditional methods. Instead of spending months or years training a foundation model from scratch, a company can use a pre-existing advanced model to provide the 'answers' for its own training set. This shortcut can significantly reduce the time and capital expense required to bring a new product to market.

However, this practice is controversial because it allows competitors to capture advanced capabilities without incurring the steep costs of original research. Frontier labs argue that this systematic extraction of data undermines the value of their investments. By using distilled data, a company can effectively replicate cutting-edge breakthroughs that were originally developed through massive computational power and human effort.

OpenAI's response to harvesting

OpenAI has already begun sounding the alarm on unauthorized model scraping and data harvesting. The company stated in a blog post that it had directly traced a core section of a broad distillation operation to individuals affiliated with Moonshot AI. This operation had been running since July, just months before the public release of their newest model.

The lab is concerned that these unauthorized activities allow rivals to bypass the heavy lifting required to create intelligent systems. By identifying specific links to companies like Moonshot AI, OpenAI is attempting to create a clearer picture of who is participating in these data extraction campaigns. These efforts are part of a broader struggle to protect the integrity of the models they spend billions of dollars to create.

The situation reflects a growing trend where AI companies must actively defend their data from being harvested by others. As models become more capable, the temptation for other firms to use them as a shortcut for training becomes stronger. OpenAI's actions indicate that they are moving toward a more defensive posture to ensure their technological lead is not easily eroded by competitors.

Technical guardrails and defenses

In response to these automated extraction campaigns, AI developers are beginning to engineer technical guardrails. These are specific safety measures designed to detect and choke off wholesale data gathering by outside parties. The goal is to make it much harder for a rival company to systematically scrape the outputs of a model to train their own systems.

Anthropic recently disclosed that it has deployed defensive countermeasures specifically designed for this purpose. These protections are meant to prevent rival companies from replicating the capabilities of its Claude model family. By making the data less accessible or harder to use for distillation, Anthropic aims to preserve the unique value of its research and development.

The company reported that these updated security protections have been highly effective so far. According to Anthropic, the new measures have cut down the amount of data harvested through distillation attempts by approximately 90% to 99%. This suggests that the industry is entering a new phase where technical barriers will play a major role in protecting intellectual property from being copied by competitors.

The future of AI competition

The conflict over model distillation highlights a shift in how the AI industry views data ownership and competition. As the gap between the US and China in AI technology narrows, the methods used to bridge that gap will come under intense scrutiny. Companies will likely have to balance the need for open collaboration with the necessity of protecting their most valuable assets from unauthorized use.

We can expect to see more labs implementing sophisticated defenses to prevent their models from being used as training sets for rivals. This could lead to a more fragmented landscape where access to high-quality training data becomes more restricted. Developers may have to find new ways to innovate that do not rely on the outputs of existing frontier models.

Ultimately, the success of these defensive measures will determine how quickly new models can be developed by smaller or less-funded companies. If distillation remains easy, the cost of entry for new AI products will remain low. If the guardrails succeed, it may take much longer and more capital for new players to reach the level of intelligence seen in models like GPT-6 Astra.