Google contractors use an internal platform to compare Gemini’s results with those of other AI models. Recently, they noticed references like “I am Claude, created by Anthropic” in some of the issues they were shown.

Based on their assessment, contractors internally discussed that “Claude’s security settings are the most stringent compared to other AI models, including Gemini.” When they submitted unsafe prompts, Claude avoided responding, while Gemini called the prompts a “major security breach” because they contained “nudity and bondage,” the report said.

Typically, technology companies evaluate the performance of their AI models using industry benchmarks. According to its terms of service, Anthropic users are not permitted to access Claude “to develop a competing product or service” or “to train competing AI models” without permission from the Google-backed startup.

While it’s unclear whether the restriction extends to investors, Google Deepmind spokeswoman Shira McNamara said the comparison of valuation model results was in line with standard industry practice. “However, any claim that we used anthropic models to train twins is inaccurate,” McNamara was quoted as saying.