Google Gemini AI Trained Using ChatGPT Data Report Claims

(Image credit: Shutterstock | Bloomberg)

In the highly competitive generative AI landscape, Google has often been seen as trying to catch up, despite its significant resources in cloud computing and talent. This perception was highlighted when Microsoft CEO Satya Nadella suggested Google had missed its initial opportunity in AI. Alphabet's CEO Sundar Pichai retorted sharply, particularly targeting Microsoft's partnership with OpenAI, stating, "I would love to do a side-by-side comparison of Microsoft's own models and our models any day, any time. They're using someone else's models."

Allegations Surface Google Reportedly Used ChatGPT for Gemini Training

However, recent developments suggest Pichai's comments might have been a case of the pot calling the kettle black. According to documents reviewed by Business Insider, Google's contractors at Scale AI allegedly used OpenAI's ChatGPT to train and enhance Bard, which is now known as Google Gemini.

The report details that these contractors generated thousands of responses from ChatGPT and compared them against Bard's outputs. This process was reportedly aimed at refining Bard's responses to be at least on par with, if not superior to, those of ChatGPT. This was in the context of Satya Nadella's earlier claim that OpenAI had a two-year head start in developing ChatGPT, making it challenging for competitors like Google's Gemini to close the gap.

Contractor Insights and OpenAIs Terms

Scale AI managers reportedly acknowledged that ChatGPT often produced more effective responses, with better formatting and more interesting facts. To incentivize improvement, contractors were allegedly offered a 15% bonus for responses from Bard that significantly outperformed GPT-4.

This practice, if true, would directly contravene OpenAI's terms of service, which explicitly prohibit the use of its model outputs for training rival AI systems.

Scale AIs Response and Industry Practices

Scale AI has refuted these claims. In a statement, the company said, "Scale did not, and does not, use ChatGPT responses to train Gemini or any models."

The AI firm characterized the activities described in the documents as "standard side-by-side evaluations." They asserted that such evaluations are common industry practice and are often misinterpreted as direct use of competitor outputs for model training and development.

Shifting Alliances Meta Scale AI and Googles Future

The situation is further complicated by recent industry movements. Details have emerged about Meta's plans for a partial acquisition of Scale AI, potentially a $14.3 billion deal for 49% ownership, which would value the AI startup at $29 billion. Additionally, Meta hired Scale AI founder Alexandr Wang to lead its new superintelligence unit.

These developments appear to have strained Google's relationship with Scale AI. Reports from Reuters indicate that Google, currently Scale AI's largest customer, is set to terminate its contract with the data-labeling firm. Google had planned to pay Scale AI up to $200 million in 2025 for human-labeled training data, a critical component for developing the AI models that power Gemini.