Did DeepSeek Train Its AI on Google’s Gemini Without Permission?

Chinese AI company DeepSeek released an updated reasoning model last week that performs well on math and coding tests. The problem? Researchers suspect the company trained it using data stolen from Google's Gemini AI.

Sam Paech, a Melbourne developer who tests AI systems, published what he calls evidence that DeepSeek's R1-0528 model learned from Gemini outputs. The model uses similar words and expressions that Google's Gemini 2.5 Pro favors, Paech noted on X.

Another developer who goes by a pseudonym found that DeepSeek's model traces read like Gemini traces. These traces are the "thoughts" AI models generate while working toward answers.

The Pattern Repeats

This isn't DeepSeek's first time facing such accusations. In December, developers noticed the company's V3 model often identified itself as ChatGPT, OpenAI's chatbot. This suggested it learned from ChatGPT conversation logs.

OpenAI told the Financial Times earlier this year it found evidence linking DeepSeek to distillation. That's a technique where companies extract data from bigger, more capable AI models to train their own systems.

Microsoft detected large amounts of data being pulled through OpenAI developer accounts in late 2024. OpenAI believes these accounts connect to DeepSeek.

The practice violates OpenAI's terms of service, which ban customers from using model outputs to build competing AI systems.

Why This Happens

Training AI models requires massive amounts of text data. Companies initially used books, articles, and websites written by humans. But AI-generated content now floods the internet. Content farms use AI to create clickbait. Bots spam Reddit and X with AI-written posts.

This contamination makes it hard to filter AI outputs from training datasets. Many models now misidentify themselves and use similar phrases because they learned from the same polluted data sources.

Still, AI researcher Nathan Lambert from nonprofit AI2 thinks DeepSeek probably did train on Gemini data. "If I was DeepSeek, I would definitely create synthetic data from the best API model out there," Lambert wrote on X. "They're short on GPUs and flush with cash. It's more compute for them."

The Arms Race Begins

AI companies now race to prevent data theft. OpenAI requires organizations to verify their identity before accessing advanced models. The process needs government-issued ID from supported countries. China isn't on the list.

Google recently started summarizing traces from models on its AI Studio platform. This makes it harder for rivals to train performant models on Gemini traces. Anthropic announced similar protections in May, citing competitive concerns.

Double Standards and Dirty Data

The accusations highlight a double standard in AI development. Western companies trained their models on copyrighted books, news articles, and creative works without permission. Publishers and authors sued over this practice.

Now those same companies cry foul when competitors might be using their AI outputs as training data. The hypocrisy is striking.

Meanwhile, the internet becomes less useful for training AI systems. As models generate more content, they risk learning from their own outputs. This creates a feedback loop that degrades quality over time.

Researchers call this "model collapse." AI trained on AI starts producing incoherent results. Accuracy drops. Models generate content that ranges from wrong to offensive.

The Censorship Factor

DeepSeek's latest model faces criticism beyond alleged data theft. The R1-0528 version shows increased censorship around Chinese political topics.

A developer using the handle "xlr8harder" tested the model with a tool called SpeechMap. The results show this version as "the most censored DeepSeek model yet for criticism of the Chinese government."

The model refuses to discuss internment camps in Xinjiang, even when prompted with documented human rights cases. It sometimes acknowledges violations occurred but stops short of assigning responsibility.

This fits China's 2023 AI regulations. Systems must not produce content that challenges government narratives or undermines state unity. Previous research found the first DeepSeek R1 model refused to answer 85% of questions on state-designated taboo topics.

What Companies Can Do

The situation forces AI companies to get creative about data sources. Some pay publishers for access to clean, human-written content. Others develop partnerships with news organizations and book publishers.

Retrieval-augmented generation offers another approach. This lets AI models search the internet in real-time instead of relying only on pre-trained data. But tests show these systems produce more unsafe responses, from privacy violations to outright misinformation.

The fundamental problem remains: AI models need human creativity to evolve. As one researcher put it, these systems cannot replace humans because they depend entirely on human input to improve.

Companies may need to pay people to create content specifically for AI training. This reverses the current model where AI companies take human work without compensation.

Why this matters:

• The AI industry built its foundation on free human content but now faces a data shortage as the internet fills with AI-generated material, creating a feedback loop that could collapse the quality of future models.
• Chinese AI companies operate under different rules than Western rivals, leading to models that excel technically but face severe political restrictions, showing how geopolitics shapes artificial intelligence development.