Google contractors working on the Gemini AI model are reportedly comparing its responses to those generated by Claude, a rival AI developed by Anthropic. Google is using Anthropic’s Claude to improve its Gemini AI by comparing output quality across multiple criteria like truthfulness and verbosity. TechCrunch’s Internal correspondence reveals this practice’s details, raising questions about consent and industry norms.
Contractors involved in improving Gemini are tasked with assessing the model’s accuracy, truthfulness, and verbosity. According to reports, these contractors evaluate each prompt over 30 minutes, scoring the outputs based on multiple criteria. While such assessments are standard in AI development, the involvement of Claude’s outputs has drawn attention.
TechCrunch reported seeing explicit references to Claude in the internal systems used by Google contractors. One example included an output that identified itself as, “I am Claude, created by Anthropic.”
Claude’s Focus on Safety
Contractors noted distinct differences between the two AI models, particularly in their approach to safety. According to internal chats, Claude’s responses were more cautious, often refusing to engage with prompts it deemed unsafe. For instance, Claude avoided certain role-playing scenarios or prompts related to sensitive topics, while Gemini’s responses reportedly led to a “huge safety violation” in one instance.
Claude’s strict safety measures appeared to surpass those of Gemini, sparking discussions among contractors. One noted that Claude would not generate responses involving nudity or bondage, whereas Gemini failed to adhere to such restrictions in at least one case.
Anthropic’s Terms of Service Questioned
Anthropic’s terms of service prohibit using Claude to create competing AI models or products without prior approval. While Google is a significant investor in Anthropic, questions remain about whether it obtained permission to use Claude in this capacity.
Google DeepMind, which oversees Gemini, stated that its evaluations involve comparing model outputs, but denied using Claude to train Gemini. A spokesperson clarified that such comparisons align with “standard industry practice.”
Concerns Over Expertise and Accuracy
Separately, TechCrunch reported last week that contractors working on Gemini expressed concerns about being assigned tasks outside their areas of expertise. These tasks included evaluating Gemini’s responses to sensitive topics like healthcare, which raised fears of inaccuracies.
Concerns have arisen over whether Google is using Anthropic’s Claude to improve its Gemini AI without proper approval from Anthropic. When approached, Google did not confirm whether Anthropic’s consent was sought for using Claude in these evaluations. Anthropic, too, declined to comment before publication.
This development highlights ongoing challenges in AI development, including ethical concerns, safety measures, and competitive practices. As companies race to improve their AI models, scrutiny over these processes continues to grow.
Safety Versus Performance
Google is using Anthropic’s Claude to improve its Gemini AI by identifying areas where Claude excels, such as safety and cautious responses. The differences in safety measures between Gemini and Claude further highlight the tension between AI performance and ethical safeguards. Contractors noted that Claude’s strict safety protocols prevented it from generating responses to unsafe prompts. On the other hand, Gemini produced outputs flagged as significant safety violations, such as inappropriate content.
This comparison underscores the challenge of balancing innovation with responsibility. While safety protocols might limit an AI’s versatility, they are crucial for preventing harmful or unethical outputs. The fact that Gemini struggled with such violations suggests a need for stronger safety mechanisms during its development.
Involving contractors without relevant expertise in evaluating sensitive topics, like healthcare, adds another layer of concern. Such practices risk inaccuracies that could have far-reaching consequences, particularly when dealing with life-critical information.
Also Read: OpenAI’s Next Flagship Model is Running Late, Facing Major Delays.
