DeepSeek's Latest AI Model Sparks Controversy: Possible Use of Google's Gemini Data

DeepSeek's Latest AI Model Raises Eyebrows

Chinese AI lab DeepSeek has recently released an updated version of its R1 reasoning AI model, known as R1-0528, which has demonstrated impressive performance on various math and coding benchmarks 1

1

2

2

. However, the release has sparked controversy and speculation within the AI community regarding the sources of its training data.

Similarities to Google's Gemini

Several AI researchers and developers have pointed out striking similarities between DeepSeek's R1-0528 and Google's Gemini family of AI models. Sam Paeach, a Melbourne-based developer, claims to have found evidence suggesting that DeepSeek's latest model was trained on outputs from Gemini 1

1

2

2

. According to Paeach, R1-0528 shows a preference for words and expressions similar to those favored by Google's Gemini 2.5 Pro.

Another developer, the creator of the "SpeechMap" AI evaluation tool, noted that the traces or "thoughts" generated by the DeepSeek model closely resemble those of Gemini 1

1

2

2

. While these observations are not conclusive proof, they have raised significant questions about DeepSeek's training practices.

Previous Allegations and Distillation Techniques

This is not the first time DeepSeek has faced accusations of training on data from rival AI models. In December, developers observed that DeepSeek's V3 model often identified itself as ChatGPT, suggesting possible training on ChatGPT chat logs 1

1

2

2

.

Earlier this year, OpenAI reported evidence linking DeepSeek to the use of distillation, a technique that involves extracting data from larger, more capable models to train smaller ones 1

1

2

2

. Bloomberg reported that Microsoft detected large amounts of data being exfiltrated through OpenAI developer accounts in late 2024, which OpenAI believes are affiliated with DeepSeek 1

1

3

3

.

Ethical Concerns and Industry Response

The allegations against DeepSeek highlight ongoing concerns about ethical practices in AI development. While distillation is not uncommon in the field, OpenAI's terms of service explicitly prohibit customers from using their model outputs to build competing AI systems 1

1

.

Nathan Lambert, a researcher at the nonprofit AI research institute AI2, suggests that it wouldn't be surprising if DeepSeek had indeed trained on data from Google's Gemini, given the company's resources and potential limitations in computing power 1

1

2

2

.

Increased Security Measures

In response to these challenges, major AI companies have been implementing stricter security measures:

1. OpenAI now requires organizations to complete an ID verification process to access certain advanced models, with China notably absent from the list of supported countries 1

   1

   2

   2

   .

2. Google has begun "summarizing" the traces generated by models available through its AI Studio developer platform, making it more difficult to train rival models on Gemini traces 1

   1

   .

3. Anthropic announced plans to start summarizing its own model's traces to protect its "competitive advantages" 1

   1

   .

The Broader Impact on AI Development

The controversy surrounding DeepSeek's latest model underscores the growing challenges in AI development and ethics. As the open web becomes increasingly saturated with AI-generated content, it has become more difficult for companies to filter out AI outputs from their training datasets 1

1

2

2

. This "contamination" of the data landscape poses significant challenges for the future of AI development and raises important questions about the originality and independence of new AI models.