Google contractors working to improve its Gemini AI are reportedly comparing its performance against outputs from Anthropic’s AI model, Claude, according to internal correspondence reviewed by sources. The process involves scoring the accuracy, truthfulness, and verbosity of responses from both models, raising questions about whether Google has obtained Anthropic’s permission to use Claude in this capacity.
The correspondence reveals that contractors evaluating Gemini are tasked with assessing AI responses for various criteria. Contractors are given up to 30 minutes per prompt to determine which model, Gemini or Claude, provides a better answer. Recent internal communications show that some responses explicitly identified Claude as the source, stating, “I am Claude, created by Anthropic.” Contractors have also observed that Claude tends to prioritize safety more than Gemini, refusing to respond to prompts it deems unsafe, while Gemini’s responses have occasionally been flagged for safety violations.
Questions About Permissions
Anthropic’s terms of service prohibit customers from using Claude to “build a competing product or service” or to “train competing AI models” without prior approval. When reached for comment, Google did not confirm whether it had obtained permission from Anthropic to use Claude for this evaluation process. Shira McNamara, a spokesperson for Google DeepMind, the team behind Gemini, emphasized that comparing model outputs is standard practice in the industry but denied that Gemini is being trained on Anthropic’s models.
“Of course, in line with standard industry practice, in some cases we compare model outputs as part of our evaluation process,” McNamara said. “However, any suggestion that we have used Anthropic models to train Gemini is inaccurate.”
Contractor Observations
Internal communications among contractors revealed that Claude’s responses often emphasize safety, with some describing its safety settings as “the strictest” among AI models. For example, Claude reportedly refused to engage in role-playing scenarios or answer certain prompts it deemed inappropriate. In one instance, Claude avoided answering a sensitive question, while Gemini’s response to a similar prompt included content flagged as a “huge safety violation” involving nudity and bondage.
These differences highlight the contrasting approaches of AI models, with safety increasingly becoming a focal point in the development of advanced systems.
Ongoing Industry Practices
In the competitive race to develop more sophisticated AI systems, tech companies often benchmark their models against competitors. While this typically involves running models through standardized industry benchmarks, Google’s approach of having contractors directly compare responses from competing models offers a different angle. However, this method raises potential ethical and legal concerns, particularly if the necessary permissions are not secured.
Anthropic, a key player in the AI field and a major recipient of investment from Google, declined to comment on the matter when approached.
As competition among AI developers intensifies, these developments underscore the fine line companies walk in evaluating and improving their technologies while adhering to industry norms and ethical considerations. The implications of Google’s comparisons could have broader repercussions for how AI models are developed, tested, and regulated
