OpenAI Investigates If DeepSeek Used AI Distillation to Train Its Model

OpenAI Suspects DeepSeek Used AI Distillation to Train Its Model

A Chinese startup launches an open-source model that shakes up markets. Days later, OpenAI and Microsoft investigate whether this model was trained using ChatGPT responses. The technique is called distillation, and initial searches suggest accounts were blocked for using it. The episode raises an uncomfortable question: how much technological miracle and how much copying is behind this new rival?

Distillation, explained without hype

Distillation is a known method in the sector: a large, expensive model, the teacher, generates responses used to train a smaller one, the student. The goal is for the student to perform similarly while costing less. If the owner of both models is the same, the process is legitimate. The problem arises when a third party uses another model's API to create its own product. OpenAI's terms prohibit using its outputs to "develop models that compete with OpenAI."

According to the Financial Times, OpenAI detected signs that DeepSeek accessed its data via this channel. Bloomberg added that OpenAI and Microsoft were investigating whether the Chinese startup used the API to distill. Both companies had reportedly blocked suspicious accounts the previous year. Trump's designated AI czar, David Sacks, spoke of "substantial evidence" and anticipated that major US tech companies would begin taking measures to "prevent distillation."

The detail that betrays the student

The circulating tests are not just corporate. An analysis circulated in the sector notes that, when faced with a complex programming challenge, the Chinese model mimics ChatGPT's response, while other competitors' solutions diverge completely. The comparison is like two exams with the same rare errors: if two students fail identically on the same question, the suspicion of copying is reasonable. This is compounded by a striking incident: when asked about its differences with ChatGPT, the model replied that both had been developed by OpenAI. It later called it a misunderstanding.

The matter has a political spin. The White House confirmed that the National Security Council was reviewing the app's impact, and its spokesperson described it as a wake-up call for the US industry. Meanwhile, Italy's data protection authority requested details on personal data usage, and the app was removed from Apple and Google stores in that country. Ireland also requested information.

Why is it debated that there are no chips?

The most repeated argument against the Chinese miracle narrative is material: training a frontier model requires computing power that, given current export restrictions, is difficult to obtain. Hence, part of the analysis suggests the advancement is only explained by leveraging others' work. The figure circulating in some media points to development requiring $1.6 billion and over 50,000 NVIDIA GPUs, a cost far higher than suggested in the public presentation.

The Chinese model's defense relies on its open nature. Available code would allow auditing how it was built. The objection is immediate: no actor with significant technological advantage gives away its weapon. If the code is released, it is argued, it is because there was nothing to protect or because the starting material came from outside.

What censorship reveals about training

There is a terrain where the model's behavior becomes more eloquent than any statement. Questions about Xinjiang, internment camps, or Xi Jinping's very existence trigger blocks and evasive answers. For part of the sector, this layer of ideological filters suggests adjustment work on an external base: the effort focused on covering up, not building. The first rule of a credible AI is not to insult the user, and blunt filters do the opposite.

The episode has also left victims in the markets. The model's emergence caused a shock in NVIDIA shares, which does not live solely on AI: it is the main supplier of graphics cards for video games and is expected for Nintendo's next console. The most widespread reading among sector trinc is that the drop was an overreaction, because the Chinese model itself was trained on NVIDIA hardware.

---

The underlying question remains, unresolved by any statement: if the model is so open and cheap, why does no one with a real advantage give it away? The answer, today, remains in the air.