Apple's Mac Chips Quietly Give Wrong AI Answers, Study Finds

A researcher named Junichiro Niimi tested the core arithmetic inside AI models — matrix multiplication, the repeated number-crunching that every chatbot, image generator and language model depends on — running it on Apple's Mac chips. His finding: when a "tensor" (a giant grid of numbers, the basic unit of AI math) grows past 2^32, or roughly 4.3 billion entries, PyTorch quietly returns the wrong result. No crash, no warning, no red text. Just numbers that look fine and aren't.

This is easier to hit than it sounds. New Macs with 192GB of shared memory make those enormous grids routine, which is exactly why the problem surfaced now. Niimi ran 1,584 experiments across different data types, memory layouts, shapes and batch sizes, checking every answer against a slow but reliable CPU calculation. Three simple rules explained all the failures on the newest version. One particularly nasty detail: slightly smaller inputs throw a proper error, while bigger ones fail silently — so a bigger job can turn a loud alarm into a quiet lie.

The practical damage is real. In a public sentiment classifier — software that reads text and labels it positive or negative — one oversized batch corrupted a third of the outputs and squashed them all onto the same label. If you're a researcher, a startup, or a hospital running AI locally on a Mac to save on cloud bills, your results could be wrong in a way you'd never spot by eyeballing them. The same class of silent error also appeared on Nvidia's CUDA hardware, though only in a different operation.

The fix, for now, is a free "guard" the author published that stops any operation touching that 4.3-billion-number threshold before it runs. Until Apple or PyTorch patch it, anyone running large AI workloads on Apple hardware should either install that guard or keep an eye on how big their calculations get. The scariest bugs aren't the ones that break things — they're the ones that don't.