Archive — US — techcrunch
DeepSeek may have used Google’s Gemini to train its latest model
Chinese lab DeepSeek has released its updated model R1-0528, which may have been trained using outputs from Google's Gemini AI family. Speculations arise as developers find similarities between DeepSeek's model and Gemini's outputs. DeepSeek has faced accusations of data sourcing practices including distillation from rival AI models, with OpenAI previously noting evidence of this behavior.
Key points
- DeepSeek released R1-0528, an updated reasoningAI model.
- The model reportedly has outputs resembling those of Google's Gemini 2.5 Pro.
- Evidence of training on Gemini outputs is suggested by developer Sam Paech.
- Previously, DeepSeek was accused of training on ChatGPT data.
- OpenAI has indicated that DeepSeek has engaged in distillation, which is prohibited by its terms.
- Security measures by AI companies have increased to prevent distillation.
Why it matters
- The controversy highlights ongoing issues of data sourcing and ethics in AI training.
- In 2025, companies ramping up security measures relate to rising competition among AI models.
- The trend of using distillation reflects broader practices as companies seek higher performance from AI.
DeepSeek's potential use of Google’s Gemini for training raises significant questions regarding ethics in AI development and data sourcing practices, emphasizing the ongoing challenges faced by companies like OpenAI and Google in securing their intellectual property.