Google is reportedly utilizing Anthropic's AI model, Claude, to enhance its Gemini AI chatbot, according to TechCrunch.
In the competitive technology sector, companies often evaluate their AI models against those of rivals. To improve the accuracy of Gemini's outputs, Google has engaged contractors to assess the performance of Gemini in comparison to Claude based on various criteria, including accuracy and redundancy.
Contractors working on Gemini's improvement are given up to 30 minutes to determine which AI model provides superior responses to a single prompt. Recently, mentions of Claude have begun appearing on an internal Google platform used to compare Gemini with lesser-known AI models.
One contractor reported that at least one output included the statement, “I am Claude created by Anthropic.”
Discussions among contractors indicated that Claude's responses appeared to prioritize safety more than Gemini's. One contractor rated Claude's safety settings as the strictest among AI models.
Claude was deemed safer than other models because it did not respond to prompts deemed unsafe, such as role-playing scenarios. In contrast, there were instances where Gemini provided outputs including references to “nudity and bondage.”
Anthropic's commercial terms explicitly prohibit clients from accessing Claude to build competing products or train rival AI models without approval. However, TechCrunch noted that Google is also a major investor in Anthropic.
TechCrunch inquired with Google about the use of Claude in enhancing Gemini, to which Sira McNamara, a spokesperson for Google DeepMind, confirmed that the comparison of AI model outputs is ongoing but did not clarify whether approval for using Claude had been obtained. McNamara also stated that claims of using Anthropic's model for training Gemini are inaccurate, denying the use of Claude for Gemini's training.
As of the time of publication, TechCrunch had not received a response from Anthropic regarding these developments.