Apple has clarified that its suite of artificial intelligence features, collectively known as Apple Intelligence, is not powered by the company’s OpenELM AI model. In a statement to 9To5Mac, the Cupertino-based tech giant confirmed that “OpenELM doesn’t power any of its AI or machine learning features – including Apple Intelligence.”

This clarification follows a Wired report that claimed major tech companies, including Apple, Nvidia, and Amazon-backed Anthropic, used content from thousands of YouTube videos, including subtitles, to train their AI models. The report suggested that Apple utilized plain text from video subtitles and their translations to train its OpenELM model.

Google’s policies prohibit the use of YouTube videos for applications that are “independent” of the platform.

Apple explained to 9To5Mac that OpenELM was developed to support the research community and advance open-source large language model (LLM) development. The company emphasized that OpenELM was created solely for research purposes and not for powering AI features in its products and devices.

Apple Intelligence Training

In a research paper published on June 10, Apple stated that it does not use “users’ private personal data or user interactions” to train its AI models. Instead, the company uses “publicly available data” from the web, gathered through its web-crawler AppleBot. Web publishers must opt out if they do not want Apple to use their content for Apple Intelligence training.

In April, Apple released its OpenELM AI model on the Hugging Face model library. OpenELM, which stands for “Open-source Efficient Language Models,” includes four small language models capable of running on devices such as phones and PCs. The models have 270 million, 450 million, 1.1 billion, and 3 billion parameters, respectively. These parameters indicate the number of variables an AI model can understand from its training data for decision-making.

For comparison, Microsoft’s Phi-3 model can handle up to 3.8 billion parameters, while Google’s open model Gemma, launched earlier this year, supports up to 2 billion parameters.

Source