Google is stepping up its efforts to assess the performance of its Gemini AI model, raising eyebrows as reports emerge of the company using Anthropic's Claude AI as a benchmark. This practice has led to questions about potential compliance with Anthropic's terms of service, especially since the comparison involves rival models.
According to internal correspondence reviewed by TechCrunch, contractors working on the Gemini project are engaged in evaluating outputs from both Gemini and Claude, focusing on criteria such as accuracy, verbosity, and safety. The contractors reportedly spend up to 30 minutes per prompt to decide which model's answer is superior. “The contractors are provided up to 30 minutes per prompt to determine whose answer is superior, Gemini’s or Claude’s,” the report stated.
This direct comparison has drawn attention to notable differences between the two AI systems. While Claude often prioritizes safety—refusing to answer prompts deemed unsafe and generating more cautious responses—Gemini has faced scrutiny for producing outputs flagged for inappropriate content. For example, one Gemini response was flagged as a “huge safety violation” for including references to “nudity and bondage.” One contractor noted, “Claude’s safety settings are the strictest” among AI models, highlighting the variance between the two.
The use of Claude for evaluation raises immediate concerns about compliance with Anthropic's clearly defined terms of service, which explicitly prohibit using its AI model to build competing products or train rival AI models without prior approval. These restrictions generate increased scrutiny over the legality and ethicality of Google's evaluation tactics.
Shira McNamara, spokesperson for Google DeepMind, acknowledged the evaluation process is standard within the industry but firmly denied any misuse of Anthropic models for training Gemini. “Of course, in line with standard industry practice, in some cases we compare model outputs as part of our evaluation process,” she stated. “Any suggestion we have used Anthropic models to train Gemini is inaccurate.”
The nuances of this relationship are complicated by Google's substantial investment stake in Anthropic, which positions them as both collaborators and competitors in the rapidly advancing AI sector. Although Anthropic has declined to comment on whether Google received permission to conduct these model comparisons, the overlap of interests prompts questions about transparency and ethical business practices moving forward.
Concerns have also been raised around Gemini's accuracy, especially relating to sensitive topics like healthcare. Some industry observers worry about the integrity of AI outputs when they are compared against competitors, potentially impacting future applications of the technology.
Notably, Google isn’t the only player currently competing to establish AI integrations and functionality. Meta, for example, is making its own strides by integrating multimodal AI capabilities within its Reality Labs division and collaborating with companies like EssilorLuxottica.
On another front, Snapchat recently announced plans to boost its AI functionality, partnering with Google to power its My AI chatbot using the Gemini AI system. This marks a significant move away from relying solely on OpenAI’s GPT models, reflecting the competitive environment among major tech companies vying to lead the AI development space.
With these developments, it's clear the race for AI supremacy is heating up, and partnerships along the way could significantly alter the AI market's makeup. The interplay between companies like Google and Anthropic will likely be watched closely as they navigate the challenges posed by compliance, consumer expectations, and the relentless pursuit of innovation.
This blending of efforts and competition will undoubtedly shape the future of AI technology, as it evolves to meet varied needs across industries. For now, Google’s commitment to enhancing the capabilities of Gemini—while adhering to compliance with Anthropic's terms—will play a pivotal role in defining its path forward.
