According to a report from TechCrunch, it seems Google is training its Gemini AI in a rather eyebrow-raising way. As part of Gemini's development, Google employs contractors whose main task is to rate the AI's responses based on qualities like accuracy, clarity, and safety. Here’s where it gets interesting: TechCrunch reports that these contractors compare Gemini’s answers with those from Claude, Anthropic’s large language model (LLM), for the same prompts.
Recently, some of these contractors started noticing something odd. In the internal Google platform used for these comparisons, one of the responses presented to Gemini contractors literally claimed to be from Claude. And it didn’t stop there. Other outputs showed Claude-like tendencies, with a stronger emphasis on safety. For example, in one instance, Claude refused to role-play a different AI assistant.
Now, here’s the catch: Anthropic's terms of service explicitly state that its chatbot cannot be used to build a competing product:
Use Restrictions. Customer may not and must not attempt to (a) access the Services to build a competing product or service, including to train competing AI models except as expressly approved by Anthropic; (b) reverse engineer or duplicate the Services; or (c) support any third party’s attempt at any of the conduct restricted in this sentence. Customer and its Users may only use the Services in the countries and regions Anthropic currently supports.
When TechCrunch pressed Google about whether they had Anthropic's permission to use Claude's outputs in this way, Google didn’t exactly give a clear answer. Shira McNamara, a spokesperson for DeepMind, told TechCrunch the following:
Of course, in line with standard industry practice, in some cases we compare model outputs as part of our evaluation process.
Cross-model comparisons aren’t uncommon in AI development. Companies routinely benchmark their models against competitors to measure performance and identify areas for improvement. But directly integrating or mimicking outputs without consent crosses into murky territory.
This all comes amid growing competition in the AI space. Google recently unveiled an experimental Gemini model, which reportedly outperformed OpenAI's GPT-4o in several benchmarks. Meanwhile, Anthropic has been enhancing Claude with notable updates. One recent improvement allows Claude to respond in various conversational styles, catering to different user preferences. Another upgrade introduced an analysis tool that can write and execute JavaScript code directly within the chatbot.
0 Comments - Add comment