DeepSeek V3: A New Player in AI or Just a Copycat?

Earlier this week, DeepSeek, a Chinese AI research lab, introduced its latest artificial intelligence model, DeepSeek V3. This model claims to outperform many existing AI systems in well-known benchmarks, handling various text-based tasks ranging from coding to essay writing. However, what raises eyebrows is not just its performance but also its identity confusion — DeepSeek V3 seems to believe it’s a version of OpenAI’s ChatGPT.

Identity Crisis: DeepSeek V3 Claims to Be ChatGPT

Multiple sources, including posts on social media platform X and tests conducted by TechCrunch, have revealed that when asked about its identity, DeepSeek V3 frequently identifies itself as ChatGPT, specifically claiming to be a version of OpenAI’s GPT-4 model released in 2023. In fact, over 60% of the time in tests, DeepSeek V3 referred to itself as ChatGPT, raising concerns about its transparency and source of knowledge.

What Does This Mean for DeepSeek’s Training Methods?

This self-identification begs the question: How was DeepSeek V3 trained? Although DeepSeek has not disclosed specifics about its training data, there is ample publicly available text data that originates from OpenAI’s models, particularly GPT-4. If DeepSeek V3 was trained on such data, it may have inadvertently memorized and replicated outputs from ChatGPT instead of generating original responses.

According to Mike Cook, a research fellow at King’s College London, this kind of overlap could be ‘accidental’ or it may involve more deliberate training practices. Such practices can compromise model quality. ‘Like taking a photocopy of a photocopy, we lose more and more information and connection to reality,’ Cook explained.

Risks of Overlapping AI Training

The issue of models copying or mimicking each other isn’t new. OpenAI specifically prohibits users from employing their outputs to create rival AI systems. DeepSeek V3’s behavior raises legal as well as ethical concerns. Unwittingly or intentionally training a model using the outputs of existing systems may violate these terms of service.

Furthermore, OpenAI’s CEO, Sam Altman, hinted at this competitive landscape on social media. He stated, “It is (relatively) easy to copy something that you know works. It is extremely hard to do something new, risky, and difficult when you don’t know if it will work.” This comments reflect a broader sentiment in the AI community about the challenges of innovation versus imitation.

AI Contamination: The Challenges of Training Data

DeepSeek V3’s case also sheds light on the increasingly cluttered landscape of online content. As AI technology proliferates, the Internet is becoming filled with AI-generated material, leading to what experts refer to as ‘AI slop.’ This contamination makes it challenging to filter out outputs that originate from other models.

By some estimates, as much as 90% of web content could be generated by AI as soon as 2026. This rapid evolution complicates the task of developing clean training datasets. Content farms, which churn out clickbait articles using AI, further contribute to this murky environment.

The Downside of Copycat AI Models

Heidy Khlaaf, chief AI scientist at the nonprofit AI Now Institute, pointed out the allure of ‘distilling’ existing models. Developing new AI systems by leveraging the knowledge of established ones can lower costs but also adds to risks, particularly if these models inherit biases and flaws from their predecessors.

“Even with the prevalence of AI content on the web, other models would not necessarily show outputs similar to OpenAI customized messages unless they are directly trained on those outputs,” Khlaaf said.

What This Means for Future AI Developments

The implications of DeepSeek V3’s identity crisis extend beyond just one model. There is growing concern among experts that such behaviors could exacerbate biases found in models like ChatGPT. If DeepSeek V3 absorbed GPT-4 outputs uncritically, it may perpetuate or even amplify existing issues within those responses.

Key Takeaways: Navigating a Murky AI Landscape

As the AI industry continues to grow, the emergence of models like DeepSeek V3 brings to light various ethical, legal, and quality concerns. The ongoing confusion around identity and training practices underscores the complexity of navigating this rapidly evolving landscape.

Moving forward, it will be crucial for AI developers to establish clearer guidelines around training methods and adherence to existing models’ terms of service. As the distinction between original content and imitation blurs, the need for increased transparency and accountability in AI development becomes even more pressing.

In summary, while DeepSeek V3 showcases impressive capabilities, its self-identification as ChatGPT raises questions about the integrity and reliability of its outputs, as the AI community cautiously watches these developments unfold.