Google collaborates with Anthropic to improve Gemini AI using Claude

Google is reportedly using Anthropic's AI model Claude to compare and improve the responses of its own AI, Gemini. Contractors noticed that Claude has stricter safety settings, avoiding unsafe prompts, while Gemini flags such inputs more explicitly.

In Short

• Google is reportedly using Claude to compare and improve Gemini AI’s responses
• The company has hired contractors to compare responses generated by Gemini and Claude
• The job of these contractors is to evaluate the quality of the output based on different factors

Google is reportedly relying on Claude, an AI model developed by Anthropic, to help improve the performance of its own AI system, Gemini. According to a report from TechCrunch, contractors hired by Google are tasked with comparing responses generated by Gemini and Claude to the same user prompts. Their job is to evaluate the quality of the output based on factors such as truthfulness, clarity, and verbosity.

The process works like this: Contractors are shown responses from both Gemini and Claude for a specific prompt. They then have up to 30 minutes to carefully assess each answer and rate how well it meets the evaluation criteria. This feedback helps Google identify areas where Gemini may need improvement. Interestingly, the contractors started noticing some unusual patterns during this evaluation process. On occasions, Gemini’s output included references like, “I am Claude, created by Anthropic,” sparking curiosity about how closely the two systems might be connected.

One thing that stood out in the comparison was Claude’s strict approach to safety. Contractors reportedly observed that Claude tends to avoid replying to prompts it deems unsafe, staying firm on maintaining high ethical standards. On the other hand, Gemini sometimes flagged these unsafe prompts but did so in a way that contractors found more detailed yet possibly less strict. For instance, in one case involving prompts related to nudity and bondage, Claude refused to engage, while Gemini identified the input as a huge safety violation.

Google has a system in place for these comparisons, using an internal platform that makes it easier for contractors to test and review multiple AI models side by side. However, the use of Claude for this purpose has raised some questions. Anthropic’s terms of service explicitly state that users are not allowed to access Claude to create or train competing AI products without approval. This rule applies to other companies using the model, but it’s unclear if it extends to investors like Google, which backs Anthropic financially.

In response to the speculation, Shira McNamara, a spokesperson for Google DeepMind, clarified the situation. She emphasised that comparing AI models is standard practice in the industry and essential for improving performance. McNamara firmly denied any claims that Google has used Anthropic’s Claude to train Gemini, calling such suggestions inaccurate.

For now, the collaboration highlights how competitive the AI industry has become, with major players like Google leveraging a mix of internal development and external partnerships to enhance their models. At the same time, it also raises questions about ethical boundaries and how companies balance collaboration with competition.