This week, the AI community has been abuzz with discussions about DeepSeek V3, the latest model released by the Chinese AI lab, DeepSeek. Touted for its performance on several AI benchmarks, this model has shown an unusual quirk: it appears to believe it's ChatGPT, the renowned AI from OpenAI. According to various sources including X and TechCrunch's own evaluations, DeepSeek V3 has repeatedly identified itself as a version of OpenAI's GPT-4, a behavior that stirs both curiosity and concern in the realm of artificial intelligence development.
A Confusing Response from DeepSeek V3
Lucas Beyer of TechCrunch tweeted that in multiple instances, DeepSeek V3 identified itself as ChatGPT, suggesting a significant influence of OpenAI's GPT-4 in its training data. This occurrence raises questions about the integrity and originality of the datasets used in training DeepSeek V3. "It gives you a rough idea of some of their training data distribution," Beyer noted, highlighting the potential overlaps in the training material of these AI systems.The Perils of Cross-Contamination in AI Training
The training methodologies for AI models like DeepSeek V3 often involve massive datasets culled from the internet, including potentially large volumes of text generated by other AI models. Mike Cook, a research fellow at King's College London, suggests that such practices, while sometimes accidental, can lead to degraded model quality and even violate the terms of service of the original AI systems, like those of OpenAI. Cook elaborated on the dangers of this approach: "Like taking a photocopy of a photocopy, we lose more and more information and connection to reality," indicating that training on outputs from rival AI systems might not only be unethical but could also compromise the quality and reliability of the AI.
The Broader Implications for AI Development
DeepSeek V3's predicament exemplifies a larger issue within the AI industry—the challenge of ensuring clean and distinct training datasets amidst a web increasingly dominated by AI-generated content. Estimates suggest that by 2026, up to 90% of the web's content could be produced by AI, complicating efforts to filter out AI-generated material from training data. Heidy Khlaaf, chief AI scientist at the nonprofit AI Now Institute, pointed out the economic allure of distilling existing models' knowledge, despite the risks. "The cost savings from distilling an existing model’s knowledge can be attractive to developers, regardless of the risks," Khlaaf said. This technique, while cost-effective, can inadvertently perpetuate the biases and flaws of the original models, leading to compounded errors and misjudgments in AI behavior.What Lies Ahead for DeepSeek V3 and AI Ethics?
The incident with DeepSeek V3 serves as a critical reminder of the complexities involved in AI development, particularly in maintaining the purity of training data. It underscores the need for stricter guidelines and transparency in AI training practices to avoid the pitfalls of unintended model mimicry and the propagation of biases.