DeepSeek could have used Google’s Gemini to coach its newest mannequin
Final week, Chinese language lab DeepSeek launched an up to date model of its R1 reasoning AI mannequin that performs nicely on quite a lot of math and coding benchmarks. The corporate didn’t reveal the supply of the information it used to coach the mannequin, however some AI researchers speculate that no less than a portion got here from Google’s Gemini household of AI.
Sam Paech, a Melbourne-based developer who creates “emotional intelligence” evaluations for AI, revealed what he claims is proof that DeepSeek’s newest mannequin was skilled on outputs from Gemini. DeepSeek’s mannequin, referred to as R1-0528, prefers phrases and expressions related to those who Google’s Gemini 2.5 Professional favors, stated Paech in an X publish.
That’s not a smoking gun. However one other developer, the pseudonymous creator of a “free speech eval” for AI referred to as SpeechMap, famous the DeepSeek mannequin’s traces — the “ideas” the mannequin generates as it really works towards a conclusion — “learn like Gemini traces.”
DeepSeek has been accused of coaching on information from rival AI fashions earlier than. In December, builders noticed that DeepSeek’s V3 mannequin typically recognized itself as ChatGPT, OpenAI’s AI-powered chatbot platform, suggesting that it might’ve been skilled on ChatGPT chat logs.
Earlier this 12 months, OpenAI informed the Monetary Instances it discovered proof linking DeepSeek to the usage of distillation, a way to coach AI fashions by extracting information from greater, extra succesful ones. In accordance with Bloomberg, Microsoft, a detailed OpenAI collaborator and investor, detected that giant quantities of information had been being exfiltrated by means of OpenAI developer accounts in late 2024 — accounts OpenAI believes are affiliated with DeepSeek.
Distillation isn’t an unusual apply, however OpenAI’s phrases of service prohibit prospects from utilizing the corporate’s mannequin outputs to construct competing AI.
To be clear, many fashions misidentify themselves and converge on the identical phrases and turns of phrases. That’s as a result of the open net, which is the place AI corporations supply the majority of their coaching information, is turning into littered with AI slop. Content material farms are utilizing AI to create clickbait, and bots are flooding Reddit and X.
This “contamination,” if you’ll, has made it fairly tough to completely filter AI outputs from coaching datasets.
Nonetheless, AI specialists like Nathan Lambert, a researcher on the nonprofit AI analysis institute AI2, don’t suppose it’s out of the query that DeepSeek skilled on information from Google’s Gemini.
“If I used to be DeepSeek, I’d positively create a ton of artificial information from the perfect API mannequin on the market,” Lambert wrote in a publish on X. “[DeepSeek is] quick on GPUs and flush with money. It’s actually successfully extra compute for them.”
Partly in an effort to stop distillation, AI corporations have been ramping up safety measures.
In April, OpenAI started requiring organizations to finish an ID verification course of with the intention to entry sure superior fashions. The method requires a government-issued ID from one of many international locations supported by OpenAI’s API; China isn’t on the listing.
Elsewhere, Google lately started “summarizing” the traces generated by fashions obtainable by means of its AI Studio developer platform, a step that makes it more difficult to coach performant rival fashions on Gemini traces. Anthropic in Could stated it might begin to summarize its personal mannequin’s traces, citing a necessity to guard its “aggressive benefits.”
We’ve reached out to Google for remark and can replace this piece if we hear again.