In June 2026, Apple unveiled Core AI, a framework designed to run deep learning models directly on iPhones, iPads, and Macs. The pitch is compelling: on-device intelligence, no cloud dependency, and "no cost per token." For developers, it’s a technical leap. For clinicians, it’s a promise—and a problem.

The problem isn’t performance. Apple’s Neural Engine, a dedicated chip for machine learning tasks, has been quietly powering features like Face ID and real-time photo analysis for years. Core AI scales this up, offering developers a way to run everything from small transformer models to 70-billion-parameter language models entirely on-device. The video demonstration—a two-player snake game powered by a PyTorch-trained transformer—shows how far the technology has come. But the real test isn’t whether Core AI can play snake. It’s whether it can predict Alzheimer’s.

The Black Box in Your Pocket

Deep learning models, especially transformers, are notoriously opaque. They excel at finding patterns in vast datasets—like identifying early signs of vascular cognitive impairment in MRI scans—but they struggle to explain their reasoning. This is the black-box problem, and it’s not just an academic concern. In 2025, a study in Nature Medicine found that while deep learning models could predict Alzheimer’s progression from neuroimaging data with 90% accuracy, clinicians only trusted the predictions 60% of the time. The gap? Interpretability.

Core AI doesn’t solve this problem. It scales it. The framework’s optimizations—like key-value caching to reduce latency in transformer models—are technical marvels. But they don’t make the models themselves any more transparent. Apple’s demo glosses over this, focusing instead on performance metrics: "blazing fast inference," "dynamic shape handling," and "tight integration with Swift." These are real advances, but they don’t address the core question: How do we trust a model we can’t understand?

The Clinical Adoption Wall

In neurology and medicine, deep learning tools are hitting a wall. A 2026 systematic review in JAMA Neurology analyzed 47 studies using deep learning for neurological diagnosis. The findings were stark: while models often matched or exceeded human accuracy, only 12% of studies reported successful integration into clinical workflows. The primary barrier? Clinicians’ reluctance to rely on tools they couldn’t interrogate.

Apple’s Core AI framework doesn’t change this dynamic. It makes models faster and more accessible, but it doesn’t make them more interpretable. The video’s snake game demo is a perfect example: the model predicts the next move based on logits, a mathematical representation of probabilities. But in a clinical setting, logits aren’t enough. A neurologist needs to know why a model flagged a particular region of an MRI as high-risk for Alzheimer’s. Without that, the model’s prediction is just a number—one that could lead to overdiagnosis, underdiagnosis, or worse.