Google is using Anthropic’s AI model, Claude, to improve its own Gemini AI, according to internal documents seen by TechCrunch.
Contractors working on the Gemini project are comparing its answers to those produced by Claude.
Google has not confirmed whether it has permission from Anthropic to use Claude for this testing.
As AI companies race to build better models, they often compare their models to others. Normally, this is done through industry benchmarks, not by manually comparing responses from different AI systems.
The contractors evaluating Gemini’s performance are scoring its responses based on factors like truthfulness and clarity.
They have up to 30 minutes per prompt to decide whether Gemini or Claude gives the better answer.
Internal documents show that contractors recently noticed references to Claude in the platform used to compare Gemini to other AI models. One output explicitly stated, “I am Claude, created by Anthropic.”
The contractors also noted that Claude’s responses seemed to focus more on safety compared to Gemini.
In some cases, Claude refused to answer prompts it considered unsafe, while Gemini gave responses that were flagged for safety violations.
Anthropic’s terms of service forbid using Claude to build competing AI models without approval. Google is a major investor in Anthropic.
Leave a comment