Google is using Anthropic’s Claude to improve its Gemini AI

The contractors working to improve Google’s Gemini AI are comparing its solutions against those of competitor Anthropic Claude, according to internal correspondence seen by TechCrunch.

Google would not say, when reached by TechCrunch for comment, whether it had obtained permission to use Claude for testing Gemini.

As tech companies rush to develop better AI models, their performance is often evaluated against their competitors, especially in driving. its examples through industry benchmarks rather than having contractors perform critical evaluations of competitors’ AI solutions.

The contractors working on Gemini who are tasked with testing the accuracy of the model’s results must respond to each response they see according to a number of criteria, such as truthfulness and verbalization. The contractors are given up to 30 minutes at the same time to determine whose answer, Gemini or Claude, according to the letters seen by TechCrunch.

Contractors recently began noticing Anthropic’s Claude appearing on the Google platform they use to compare Gemini with other unspecified AI models, the documents showed. One of the results given to Gemini contractors, seen by TechCrunch, clearly stated: “I am Claude, produced by Anthropic.”

One internal chat showed contractors seeing Claude’s solutions seem to emphasize safety more than Gemini. “Claude’s security preferences are the most complex” among AI types, one author wrote. In some cases, Claude would not respond to what he considered unsafe, such as the play of another AI assistant. Elsewhere, Claude avoided a quick response, while Gemini’s response was described as “a serious breach of security” including “nudity and slavery.”

Anthropic history job sales prohibit customers from accessing Claude “to develop competing products or services” or “to train competing AIs” without Anthropic’s consent. Google is great Investor in Anthropic.

Shira McNamara, a spokeswoman for Google DeepMind, which runs Gemini, would not say – when asked by TechCrunch – whether Google has obtained Anthropic’s license to acquire Claude. Reached before publication, an Anthropic spokesperson did not respond by press time.

McNamara said that DeepMind “compares model outputs” for analysis but does not train Gemini on Anthropic models.

“Indeed, consistent with industry standard practice, we sometimes compare outputs as part of our analysis,” McNamara said. “However, any suggestion that we have used Anthropic models to train Gemini is incorrect.”

Last week, TechCrunch reported exclusively that Google contractors working on the company’s AI products are now being trained to test Gemini’s AI solutions in areas where they lack expertise. The internal letters revealed the contractors’ concern that Gemini could produce inaccurate information on sensitive topics like health care.

You can safely send tips to this reporter on Signal at +1 628-282-2811.

TechCrunch has a newsletter focused on AI! Log in here to get it in your inbox every Wednesday.