Apple reportedly claimed that its recently released Open Source Efficient Language Model (OpenELM) AI models were not used in Apple Intelligence or the company's machine learning features. The Cupertino-based tech giant open-sourced the OpenELM models in April. The reported statement came just days after an investigation found that the AI models were trained on datasets that included subtitles for hundreds of thousands of YouTube videos. Notably, these datasets were also used to train AI models from Anthropic, Salesforce, Nvidia, and others.
Apple reportedly denies using OpenELM for Apple Intelligence
The iPhone maker told 9to5Mac that OpenELM isn't used for any of its Apple Intelligence-branded AI features or other machine learning capabilities. The company also added that OpenELM was created as a contribution to the research community to advance the development of large-scale language models.
The AI model was designed for research purposes and isn't used for Apple's internal AI innovation, the company told the publication. Of note, OpenELM is still open source and the model weights are accessible from the company's Hugging Face list.
At the time of release, Apple made the OpenELM family of AI models available in two variations: pre-trained and prescriptive. These models are further divided into 270 million, 450 million, 1.1 billion, and 3 billion different parameters.
AI model trained on YouTube data
A recent investigation accused Apple's OpenELM and several other AI models from major AI companies of using YouTube video caption data for training. The report found that companies like Anthropic, Salesforce, and Nvidia used a public data repository called Pile, which contained data from YouTube videos from major content creators like Marques Brownlee, MrBeast, and CarryMinati.
While Apple has denied using the AI models on its devices or operating system, Anthropic spokesperson Jennifer Martinez told Proof News, which conducted an investigation, “Pile contains a small portion of YouTube subtitles. YouTube's terms cover direct use of their platform, which is separate from use of the Pile dataset. Any potential violations of YouTube's terms of service should be directed to the Pile authors.”
