Google is Utilizing Outputs from Claude to Enhance the Gemini Model, Unclear if Permission Was Granted
Brook.X
Currently, competitors such as Gemini, ChatGPT, and Claude are in a fierce race. Despite this, Google is leveraging Claude to refine the outputs of its Gemini model, though it's uncertain if they have obtained the necessary permissions.
Internal communications viewed by TechCrunch reveal that a Google-contracted team focused on enhancing the Gemini AI has been comparing the model's responses with those generated by Claude.
As tech companies vie to develop superior artificial intelligence models, their performances are typically benchmarked against industry standards, rather than through exhaustive evaluations of competitors’ responses by contracted teams.
Initially, the contractors were unaware of the data's origin, as Google only compared Gemini's outputs with those from an undisclosed model. They were tasked with scoring each response based on criteria such as authenticity and verbosity, spending up to 30 minutes to determine which answer was superior.
However, the team recently discovered references to Claude on Google's internal platform used for comparison, including at least one explicit output viewed by TechCrunch: "I am Claude, created by Anthropic."
Internal chats further revealed that the team noticed Claude’s responses seemed to prioritize safety more strongly than Gemini's. Some members mentioned, "Claude’s safety settings are the strictest among all AI models, refusing to respond to prompts it deems unsafe in certain instances." In one case, Claude avoided a prompt, whereas Gemini's response, which included nudity and bondage, was flagged for severe safety violations.
Anthropic’s terms of service explicitly prohibit clients from accessing Claude to develop competitive products or services, or train competitive AI models without permission. However, Google remains one of Anthropic's investors.
A spokesperson for Google DeepMind declined to confirm whether permission had been obtained from Anthropic, stating that comparing model outputs as part of the evaluation process is standard industry practice under certain circumstances, but any claims that we are using Anthropic’s model to train Gemini are inaccurate.
