DeepSeek's R1 AI Model Sparks Debate Over Potential Training on Google Gemini Data

The world of artificial intelligence is a rapidly evolving landscape, marked by intense competition and groundbreaking advancements. Companies race to build larger, more capable models, pushing the boundaries of what AI can achieve. However, this pursuit of progress is not without its controversies, particularly concerning the source and nature of the data used to train these sophisticated systems. A recent development involving the Chinese AI lab DeepSeek and its latest reasoning model, R1, has brought these issues back into sharp focus, with researchers speculating that the model may have been trained on data originating from Google's powerful Gemini family of AI models.

DeepSeek recently unveiled an updated version of its R1 reasoning AI model, which has demonstrated impressive performance on various benchmarks, particularly in areas like mathematics and coding. While the company has remained tight-lipped about the specific datasets used for training R1, the model's behavior and output characteristics have led some AI researchers to hypothesize a controversial training source: synthetic data generated by rival AI models, specifically Google's Gemini.

The Evidence: Linguistic Fingerprints and Model Traces

The speculation isn't entirely unfounded. Several developers and researchers have pointed to intriguing clues within the DeepSeek R1-0528 model's outputs. Sam Paech, a developer based in Melbourne known for creating "emotional intelligence" evaluations for AI, presented what he believes is evidence supporting the theory that DeepSeek's model was trained on Gemini outputs. According to Paech, the R1-0528 model exhibits a preference for words and expressions that are notably similar to those favored by Google's Gemini 2.5 Pro model. This linguistic overlap, while not definitive proof, suggests a potential influence from Gemini's style and vocabulary.

If you're wondering why new deepseek r1 sounds a bit different, I think they probably switched from training on synthetic openai to synthetic gemini outputs. pic.twitter.com/Oex9roapNv

— Sam Paech (@sam_paech) May 29, 2025

Further adding to the speculation is the observation made by the pseudonymous creator of SpeechMap, an AI evaluation tool designed to assess a chatbot's willingness to discuss controversial topics. This developer noted that the internal "traces" generated by the DeepSeek model – the step-by-step reasoning process the model follows to arrive at a conclusion – bear a striking resemblance to the traces produced by Gemini models. These traces can offer insights into a model's underlying architecture and training methodology, and similarities here could indicate that DeepSeek's model learned its reasoning patterns from Gemini's examples.

A Pattern of Suspicion: Previous Accusations Against DeepSeek

This isn't the first time DeepSeek has faced accusations related to its training data sources. In December of the previous year, developers observed that DeepSeek's V3 model frequently identified itself as ChatGPT, the popular AI chatbot platform developed by OpenAI. This peculiar behavior strongly suggested that the V3 model might have been trained on conversational logs or outputs generated by ChatGPT, leading to it inheriting the rival model's identity.

More serious allegations emerged earlier this year. OpenAI informed the Financial Times that it had uncovered evidence linking DeepSeek to the use of distillation techniques. Distillation is a method where a smaller model is trained to replicate the behavior and outputs of a larger, more capable model. While distillation itself is a recognized technique, using it to train a competing model on the outputs of another company's proprietary AI raises significant intellectual property concerns.

According to a Bloomberg report, Microsoft, a major investor in and collaborator with OpenAI, detected a large volume of data being exfiltrated through OpenAI developer accounts in late 2024. OpenAI reportedly believes these accounts are affiliated with DeepSeek, suggesting a potential unauthorized access and data extraction operation aimed at obtaining training material from OpenAI's models.

The Murky Waters of AI Training Data and 'AI Slop'

It's important to contextualize these accusations within the broader challenges of AI model training. The internet, the primary source of data for training large language models, is becoming increasingly saturated with AI-generated content, often referred to as "AI slop." This proliferation of synthetic text makes it difficult for AI companies to curate clean, human-generated datasets.

Many AI models, regardless of their origin, can sometimes misidentify themselves or converge on similar linguistic patterns. This phenomenon can occur simply because they are trained on overlapping datasets scraped from the public web, which now includes a significant amount of AI-generated text. Content farms are leveraging AI to produce clickbait articles, and bots are flooding platforms like Reddit and X (formerly Twitter) with automated content. This "contamination" makes it increasingly challenging to filter out AI-generated outputs from training datasets effectively.

However, the specific similarities observed between DeepSeek R1 and Gemini, particularly in their reasoning traces, go beyond mere convergence due to shared web data. They suggest a more direct influence, potentially through the deliberate use of Gemini's outputs as training material.

The Strategic Advantage of Synthetic Data

From a strategic perspective, training on synthetic data generated by a state-of-the-art model like Gemini could offer significant advantages, especially for companies with limited access to the vast computational resources required to train foundational models from scratch. Nathan Lambert, a researcher at the nonprofit AI research institute AI2, articulated this viewpoint in a post on X.

If I was DeepSeek I would definitely create a ton of synthetic data from the best API model out there. Theyre short on GPUs and flush with cash. It’s literally effectively more compute for them. yes on the Gemini distill question.

— Nathan Lambert (@natolambert) June 3, 2025

Lambert suggested that if he were in DeepSeek's position, he would "definitely create a ton of synthetic data from the best API model out there." He reasoned that companies like DeepSeek, which might be "short on GPUs and flush with cash," could effectively gain more computational power by leveraging the outputs of existing powerful models like Gemini. This practice, essentially a form of distillation, allows them to build capable models without the immense compute resources needed for pre-training on massive raw datasets.

While distillation is a known technique, using a competitor's API outputs to train a directly competing model typically violates the terms of service of most AI providers, including OpenAI. This is precisely the core of the controversy: whether DeepSeek obtained and used Gemini's outputs in a manner that infringes upon Google's intellectual property rights and terms of service.

Industry Response: Bolstering Security Measures

In response to the increasing threat of data exfiltration and distillation, major AI companies are implementing more stringent security measures to protect their proprietary models and the data they generate. These measures aim to make it harder for malicious actors or competitors to scrape or extract valuable information from their APIs and services.

In April, OpenAI began requiring organizations accessing certain advanced models through its API to complete an ID verification process. This process typically involves providing a government-issued ID from a country supported by OpenAI's API. Notably, China is not currently on this list, which could potentially restrict access for Chinese entities like DeepSeek to OpenAI's most advanced models.

Google has also taken steps to protect its Gemini models. The company recently started "summarizing" the detailed traces generated by models available through its AI Studio developer platform. By providing a less granular view of the model's internal reasoning process, Google makes it more challenging for others to reverse-engineer or train performant rival models based on these traces. This move is a direct response to the potential for distillation using trace data.

Anthropic, another leading AI lab and developer of the Claude models, announced in May that it would also begin summarizing its own model's traces. Citing the need to protect its "competitive advantages," Anthropic's decision underscores the industry-wide concern about the vulnerability of detailed model outputs to distillation and replication by competitors.

The Broader Implications for the AI Ecosystem

The DeepSeek-Gemini situation, if the allegations prove true, highlights several critical issues facing the AI industry:

• Intellectual Property and Data Rights: The use of synthetic data derived from competitor models raises complex legal and ethical questions about intellectual property ownership in the age of AI. Is a model's output protected? Can it be used freely for training competing systems?
• Competitive Dynamics: If companies can quickly train competitive models by distilling knowledge from market leaders, it could disrupt the competitive landscape, potentially undermining the significant investments made by pioneers in developing foundational models.
• Data Sourcing Challenges: The increasing prevalence of AI-generated content on the web makes it harder and more expensive to acquire clean, diverse, and high-quality human-generated data for training. This could push more companies towards synthetic data, exacerbating the issues of attribution and potential contamination.
• Security and Access Control: The incident involving alleged data exfiltration from OpenAI accounts underscores the need for robust security measures to protect access to valuable AI models and their outputs.

The AI community is grappling with these challenges. While open research and knowledge sharing have historically driven progress, the commercial realities of building and deploying powerful AI models are leading companies to protect their assets more fiercely. The balance between fostering innovation through openness and protecting proprietary investments is a delicate one.

DeepSeek has not publicly addressed the specific allegations regarding the use of Gemini data for training its R1 model. Google has been reached out to for comment, but no official statement has been provided as of the time of the original report. The lack of transparency regarding training data sources across the industry makes it difficult to definitively confirm or deny such claims, relying instead on empirical observations and behavioral analysis of the models themselves.

As AI models become more sophisticated and the competition intensifies, the methods used for training data acquisition and the ethical boundaries surrounding the use of synthetic data derived from other models will likely remain subjects of significant debate and scrutiny. The DeepSeek R1 case serves as a potent reminder of the complex challenges inherent in building the next generation of artificial intelligence.