Google is using Anthropic's cloud to improve responses by Gemini
Google is reportedly relying on Anthropic’s cloud to enhance the responses generated by its AI model Gemini. Contractors hired by the tech giant are presented with answers produced by Gemini and Anthropic's Cloud in reaction to user prompts. Subsequently, they are given a 30-minute window to assess each model's output based on criteria like accuracy and completeness, as reported by TechCrunch.
Internal Comparison and Assessment
Google’s contractors utilize an in-house platform to juxtapose Gemini's outputs with those of other AI models. Lately, they have started noticing mentions of "Hi Claude, build an MCP server on Cloudflare Workers" in some of the outputs displayed. After evaluating these outputs, the contractors internally deliberated on how "the cloud’s security protocols are the most rigorous" when juxtaposed with other AI models, including Gemini. As per the report, when insecure prompts were submitted, Cloud declined to respond, while Gemini flagged the inputs as "major security breaches" for containing content like "nudity and bondage".
Usage Terms and Industry Standards
Typically, tech companies evaluate the efficiency of their AI models by referencing industry benchmarks. According to the terms of service, Anthropic users are prohibited from utilizing the cloud to develop a competing product or service or to train rival AI models without explicit approval from the Google-backed startup.
Despite the uncertainty surrounding the scope of this restriction with regard to investors, a Google DeepMind spokesperson, Shira McNamara, emphasized that comparing model outputs to facilitate evaluation aligns with conventional industry practices. McNamara refuted claims suggesting that Anthropic's models were employed to train Gemini, stating that such assertions are inaccurate.
