Google Reveals Over 100,000 Cloning Attempts on Gemini Platform
Google recently reported that over 100,000 attempts to clone the Gemini AI chatbot have been detected. These attempts, described as “commercially motivated,” primarily involved non-English languages. The activity was characterized as “model extraction,” which Google views as a form of intellectual property theft.
Understanding Model Extraction
Model extraction entails using another company’s AI model to create a cheaper or alternative version. Google released this information in a quarterly assessment that frames the company both as a victim and a guardian of its technology.
Recent Cloning Attempts
- Cloning attempts exceeded 100,000 prompts.
- Many sessions were conducted in various non-English languages.
- The activity is seen as a threat from mostly private companies and researchers seeking a competitive advantage.
Google’s Historical Context
In 2023, Google faced accusations from The Information regarding its Bard team. They allegedly utilized outputs from ChatGPT shared on ShareGPT to train their own chatbot. Jacob Devlin, a senior Google AI researcher and creator of the BERT model, warned that this practice could breach OpenAI’s terms of service.
Industry Standards and Practices
Despite Google denying these allegations, footnotes of this incident highlight the complexities of AI ethics. Generally, the practice of training a new model using the outputs of an existing one is known as “distillation.” This method allows companies to shortcut the extensive resources required to build their own large language models (LLMs).
Global Implications
Google asserts that these cloning attempts originate from various parts of the world. Specific suspects have not been identified. The ongoing challenge of model extraction underscores the need for better regulatory frameworks within the AI industry.
