Chinese startup DeepSeek stunned the world with its sophisticated DeepSeek R1 reasoning model, which is as good as ChatGPT o1. That’s not a surprising achievement; it’s only a matter of time before other AI models can replicate what OpenAI has done in terms of AI reasoning. Also, OpenAI will soon make o3 available, the successor to o1.
What really shocked the markets was DeepSeek’s research, which showed that the company was able to train R1 to achieve the same capabilities at a fraction of the cost of training o1.
Because of US sanctions, DeepSeek didn’t have access to the latest NVIDIA GPUs that AI firms like OpenAI use to train high-end AI models. It turned to software optimizations to compensate for what it lacked in hardware to create an AI model that could match ChatGPT o1.
But it turns out software optimization isn’t everything DeepSeek might have done to train its AI. OpenAI claims it has evidence that DeepSeek distilled ChatGPT to train the DeepSeek AI models.
If that’s true, the practice violates OpenAI’s terms of service for ChatGPT. Ironically, if OpenAI’s claim is true, it’ll make the company experience what many creators felt when they discovered OpenAI may have trained its ChatGPT models using copyrighted materials without consent.
OpenAI told The Financial Times it found evidence that DeepSeek used the US models to train DeepSeek AI.
OpenAI found evidence of “distillation,” which it believes came from DeepSeek. Distillation is a process where AI firms use an already trained large AI model to train smaller models. The “student” models will match similar results to the “teacher” AI in specific tasks.
Some early DeepSeek testers were surprised to see the AI identify itself as ChatGPT in early responses, which prompted speculation that DeepSeek AI might have been trained with ChatGPT chats.
OpenAI claims that DeepSeek might have distilled ChatGPT make sense, but it’s unclear whether the US AI firm can prove the IP theft beyond doubt. Even if it can provide conclusive evidence that DeepSeek used ChatGPT to train its AIs, there’s probably little OpenAI can do. After all, DeepSeek R1 is already out in the wild.
DeepSeek made its models available open-source, which means anyone can install them on computers. The DeepSeek app is topping the App Store, and it’s available in the Google Play store. Unless DeepSeek is banned in the US, the app won’t go away anytime soon.
The FT says that OpenAI and Microsoft investigated accounts believed to belong to DeepSeeka last year. They were using OpenAI’s API for ChatGPT access. OpenAI blocked access, suspecting they may rely on distillation to train other models.
DeepSeek has not commented on these allegations. The company is seen as a hero in China after the release of DeepSeek R1, which wiped nearly $1 billion from the US market.
On the other hand, it’s not just Chinese AI companies like DeepSeek that might rely on the distillation of ChatGPT and other frontier AIs to train better AI models. The FT notes that it’s common practice for AI labs in China and the US to use outputs from bigger companies.
OpenAI and others have already trained AI using humans to teach the models how to produce responses that sound more conversational. This is an expensive process, so smaller firms will distill established models to train smaller ones. In such a case, a company like DeepSeek would have gotten the human feedback step for free.
I said earlier that DeepSeek’s use of distillation to train R1 is something others could benefit from, Apple included. I wasn’t referring to stealing AI work done by others but to using advanced, proprietary models to train smaller models that Apple might need for its on-device Apple Intelligence approach.
If OpenAI has strong evidence that DeepSeek used ChatGPT to train its AI models, we could be looking at the second good reason to ban DeepSeek in the US and elsewhere. The first is that DeepSeek collects plenty of user data and sends it all to China.
A ban is a process that will take time. And, again, even if all of this is successful, DeepSeek will still have strong AI models on its hands, which it can use to create next-gen AI of its own.
Meanwhile, OpenAI still has to deal with allegations that it used copyrighted content without consent to create ChatGPT.
