How AI Models Like DeepSeek V3 Learn
AI models, such as DeepSeek V3 and ChatGPT, are statistical systems that generate responses. They have been trained on huge datasets to analyze patterns and predict what comes next in a sequence. For example, in an email, they can predict that "thank you" will likely be followed by "in anticipation."
However, a huge issue concerning DeepSeek V3's training dataset is the lack of transparency. OpenAI bans the use of ChatGPT-generated content to train competing models, but there is still an assumption that DeeplSeek V3 may have been exposed to this kind of material, which would proliferate responses of ChatGPT.
Is DeepSeek V3 Copying ChatGPT?
Experts believe that DeepSeek V3 may have ingested outputs from ChatGPT, which it could then vomit back up without modification. Mike Cook, a research fellow at King's College London who focuses on AI, said that the model has probably seen responses generated by ChatGPT, but he couldn't say where they came from.
He further adds that training an AI model from the outputs generated by other competitive systems degrades the quality of the model. This process can be compared with photocopying a photocopy, which often loses the information and increases hallucinations when the AI generates misleading information or even provides entirely false data.
Violating Ethical and Legal Boundaries?
If DeepSeek V3 was actually trained on content generated by ChatGPT, it might violate OpenAI’s terms of service. OpenAI clearly forbids the use of its AI-generated outputs for training competing models. Nevertheless, AI developers are frequently lured by the cost benefits of utilizing existing model knowledge rather than starting from the ground up. Neither OpenAI nor DeepSeek has commented officially on these accusations. However, the CEO of OpenAI, Sam Altman, took a sarcastic jibe on social media, appearing to point the finger at DeepSeek and others like it. "It is not that difficult to copy someone good. It's very hard to innovate," he noted.
The Problem of AI-Generated Content Pollution Increasing
DeepSeek V3 is not the first AI model to misidentify itself or demonstrate data contamination. Even Google's Gemini has been known to mistakenly claim association with competing AI models. This is becoming increasingly common as AI-generated content floods the internet.
As AI-generated text fills the web at an alarming rate—some projections suggest that by 2026, more than 90% of online content may be AI-generated—it becomes increasingly challenging to distinguish between original and AI-produced data. Such widespread "contamination" complicates training new AI models without inadvertently recycling existing AI output.
The Risk of Bias and Misinformation
If DeepSeek V3 uses ChatGPT data for training, it could end up inheriting and intensifying the biases and errors found in GPT-4’s outputs. Heidy Khlaaf, Chief AI Scientist at the AI Now Institute, points out that while accidental exposure to content generated by ChatGPT is a possibility, the intentional "distillation" of OpenAI’s models for the sake of cost efficiency presents an appealing yet risky option for developers.
Even though DeepSeek V3 wasn't explicitly trained on ChatGPT data, it likely ingested huge chunks of such content from the internet, thus not being reliable enough to self-identify. The more significant danger, however, is the spread of systemic biases and errors deep from one model to another, escalating the danger posed by the spread of AI-generated misinformation.
Final Thoughts: Time for Transparency in AI Development
The allegations about DeepSeek V3 reveal the rising issues of AI ethics, transparency, and accountability. When training data sources are not disclosed, users cannot fully trust AI models to provide unbiased and accurate information.
AI developers need to prioritize ethical training practices and ensure that their models are founded on a variety of credible data sources. Failing to do so will only perpetuate the cycle of AI-generated misinformation, further eroding the reliability of AI systems globally.