Saturday, December 28, 2024
HomeUncategorizedAnthropic’s Claude highlights safety gaps in Gemini AI

Anthropic’s Claude highlights safety gaps in Gemini AI

Contractors tasked with improving Google’s Gemini AI are reportedly comparing its responses against those generated by Anthropic’s competitor model, Claude.

Google has not confirmed whether it obtained permission from Anthropic to use Claude in these evaluations.

As technology companies race to develop advanced AI models, performance is often assessed through industry benchmarks rather than direct comparisons by contractors. However, contractors working on Gemini are reportedly required to evaluate responses based on criteria such as truthfulness and verbosity, spending up to 30 minutes per prompt to determine which model performs better.

Reports revealed that contractors began noticing references to Anthropic’s Claude within the internal Google platform used to compare Gemini against other unnamed models. In at least one instance, an output presented to contractors explicitly stated, “I am Claude, created by Anthropic.”

One chat among contractors reportedly highlighted Claude’s emphasis on safety. Contractors observed that Claude’s safety measures were stricter than Gemini’s, with the former declining to respond to prompts it deemed unsafe. For example, Claude avoided answering a role-playing prompt, while Gemini’s response to a similar query reportedly included content flagged as a ‘huge safety violation’ involving ‘nudity and bondage.’

Anthropic’s commercial terms of service prohibit customers from accessing Claude to build competing products or train rival AI models without approval. Google, a major investor in Anthropic, has not disclosed whether it received such approval.

Shira McNamara, a spokesperson for Google DeepMind, which oversees Gemini, stated that while model outputs are compared during evaluations, Anthropic’s models are not used to train Gemini.

Last week, reports indicated that Google contractors working on Gemini have been asked to rate AI responses in areas outside their expertise. Internal concerns were raised about the potential for Gemini to generate inaccurate information on sensitive topics, including healthcare.

RELATED ARTICLES

LEAVE A REPLY

Please enter your comment!
Please enter your name here

- Advertisment -

Most Popular

Recent Comments