Google reportedly using Anthropic's Claude to improve Gemini

Technology media

https://techcrunch.com/2024/12/24/google-is-using-anthropics-claude-to-improve-its-gemini-ai/

In the tech industry, where there is a race to build better AI models, AI companies often evaluate the output of their AI models against those of their competitors. To improve the accuracy of Gemini's output, Google uses contractors to have the output evaluated by others, scoring Gemini according to multiple criteria, including the veracity and redundancy of the output.

Contractors working on improving Gemini are giving people up to 30 minutes per prompt to decide whether Gemini's or Claude's answers are better, according to information obtained exclusively by TechCrunch. The contractors said references to Anthropic's Claude recently began appearing on an internal Google platform used to compare Gemini to other unnamed AI models.

One contractor who spoke to TechCrunch said at least one of the outputs included the caption, 'I am Claude created by Anthropic.'

Internal Google chats also show that contractors noticed that Claude's responses seemed to prioritize safety over Gemini, with one contractor rating 'Claude's safety settings as the strictest of any AI model' within Google's internal platform.

The reason why Claude was rated safer than other AI models is that it did not respond to prompts that it deemed unsafe, such as role-playing. In addition, there have been reports of Gemini outputting answers including 'nudity and bondage' to prompts that Claude avoided answering.

Anthropic's terms of service specifically prohibit customers from accessing Claude 'to build competing products or services' or 'to train competing AI models' without Anthropic's approval, although Google is also a major investor in Anthropic, TechCrunch noted.

TechCrunch asked Google about using Claude to improve Gemini, and Shira McNamara, a spokesperson for Google DeepMind, which operates Gemini, acknowledged that 'Google DeepMind compares the output of AI models,' but did not answer whether Google had obtained approval for the use of Claude. She also denied using Claude to train Gemini, saying, 'The suggestion that we used Anthropic's models to train Gemini is inaccurate.'

TechCrunch has reached out to Anthropic for comment, but has not received a response at the time of writing.