DeepSeek likely utilized Google’s Gemini for training its newest AI model

DeepSeek has introduced its updated R1-0528 AI model, stirring speculation that its training data may partly originate from Google’s Gemini. While these claims come from developers pointing to similarities in language and structure, the true sources remain undisclosed.

Short Summary:

• DeepSeek’s latest model, R1-0528, shows impressive performance on math and coding benchmarks.
• Speculation surrounds the use of Google’s Gemini outputs in training this model.
• DeepSeek has a history of facing allegations regarding data source transparency and distillation practices.

In a significant move that has captivated the AI community, the Chinese lab DeepSeek has rolled out an updated version of its R1 reasoning AI model. The newest iteration, named R1-0528, boasts improved performance on various math and coding tasks. However, the release hasn’t come without its controversies; the company has attracted scrutiny regarding the origins of its training data. Interestingly, multiple developers are now suggesting that DeepSeek may have utilized outputs from Google’s Gemini, one of the tech giant’s leading AI families, during the training process.

One of the foremost voices in this conversation is Sam Paech, a developer based in Melbourne who specializes in “emotional intelligence” evaluations for AI technologies. Paech conducted analyses that indicated a close correlation between the expressions and word choices favored by DeepSeek’s R1-0528 and those found in Google’s Gemini 2.5 Pro model. He shared his findings on X, stating,

“If you’re wondering why the new DeepSeek R1 sounds a bit different, I think they probably switched from training on synthetic OpenAI to synthetic Gemini outputs.”

While Paech’s observations are compelling, they stop short of serving as definitive evidence that DeepSeek has trained its model on Gemini outputs. Another developer, who chose to remain anonymous, echoed this sentiment, highlighting that the “thoughts” generated by R1-0528 resemble those of Gemini, further fueling discussions on the potential overlap between the two models.

DeepSeek’s reputation for using data from rival AI models is not new; the lab has faced allegations before pertaining to the use of outputs from OpenAI’s ChatGPT. Back in December, industry observers noted that DeepSeek’s V3 model frequently identified itself as ChatGPT, which prompted speculation regarding the potential use of ChatGPT chat logs in its training dataset. This overlap raises questions about transparency in AI data sourcing.

Moreover, in an earlier revelation this year, OpenAI informed the Financial Times about potential evidence that insinuated DeepSeek was engaging in distillation practices. Distillation is a method whereby a smaller model is trained using the outputs of a larger, more advanced counterpart. According to reports from Bloomberg, Microsoft—an important ally and investor of OpenAI—detected substantial data exfiltration from OpenAI developer accounts in late 2024, which they suspect are linked to DeepSeek’s operations.

While distillation may not, in and of itself, constitute wrongdoing, it is crucial to note that OpenAI’s terms of service explicitly prohibit the training of competing AI models using their outputs. Thus, accusations surrounding DeepSeek’s methods carry significant weight in the industry and could have larger ramifications for the company’s legitimacy.

Interestingly, many experts in artificial intelligence caution against jumping to conclusions. The reality is that a myriad of AI models today tends to converge in terms of language use and phrasing. This is largely due to the oversaturation of AI-generated content across the internet—an ecosystem now plagued by what some refer to as “AI slop.” Content farms are churning out low-quality clickbait, while bots are inundating social media platforms like Reddit and X, creating a murky training environment for new AI models.

In the midst of this, AI specialists such as Nathan Lambert from the nonprofit AI research institute AI2 believe there may indeed be merit to the speculation. Lambert remarked,

“If I was DeepSeek, I would definitely create a ton of synthetic data from the best API model out there. They’re short on GPUs and flush with cash. It’s literally effectively more compute for them.”

The suggestion here is both a practical strategy for DeepSeek in navigating resource constraints and an acknowledgment of how firms like Google and OpenAI have built extensive datasets over the years.

In response to the increasing concerns surrounding data sharing and distillation, several top AI companies have initiated stringent security measures. In April of this year, OpenAI began mandating that organizations undergo an identity verification process before accessing more advanced models. This process requires a government-issued ID from one of the nations supported by OpenAI’s API—China, notably, is excluded from this list.

Simultaneously, Google has rolled out efforts to summarize the internal workings of its AI models accessible via their AI Studio developer platform. This approach adds a layer of complexity for developers attempting to train comparable competing models based on Gemini’s internal data. Likewise, Anthropic, another significant player in the AI field, announced plans to summarize its own models’ traces to protect its competitive edge.

The implications of these developments extend beyond a simple rivalry among AI organizations. As the landscape of synthetic data generation, model training, and competitive practices continues to evolve, what is at stake for companies like DeepSeek is nothing less than survival amidst scrutiny and intensive competition. AI models, particularly sophisticated ones like Gemini, have required years of development and substantial investment. As a result, any upstart company, under suspicion of borrowing from existing models, risks jeopardizing its reputation and future prospects.

For now, the details surrounding the potential overlap between DeepSeek and Google’s Gemini remain speculative. However, the ongoing discourse hints at larger themes in the industry context. Whether DeepSeek truly harnessed Gemini’s outputs or not may not be as important as the fundamental issues of data sourcing, transparency, and competitive ethics within the AI realms at play.

As the controversies unfold, it’s worth noting that while models like R1-0528 indicate a significant technological leap, vulnerabilities remain in the form of ethical practices concerning IP and data ownership. The broader AI community, including startups, established firms, and regulatory bodies, will need to navigate these turbulent waters as advancements continue at an unprecedented pace.

This situation serves as a reminder for organizations engaged in AI development to reevaluate their training data sources and to consider implementing robust operating models that honor ethical practices while maintaining fierce competitiveness. After all, at the intersection of AI innovation, there lies a complex web of accountability, transparency, and ethical considerations that will ultimately define the future landscape.

So, fellow enthusiasts, keep your eyes peeled for the latest updates in the AI industry. You can stay informed through platforms like Latest AI News and enhance your own content creation with tools like Autoblogging.ai to stay ahead in this rapidly changing environment.