Apple's Sensitive Content Warning is supposed to do one thing: catch nude images before they land on a screen that didn't ask for them. On Tuesday, it blurred a video of a dog.
That is not a metaphor. A user received a video of a friend's dog. The on-device classifier — the actual piece of software doing the work — scored it above whatever threshold Apple set for nudity. The screen went gray. The warning appeared.
The classifier runs locally, on Neural Engine silicon in recent iPhones and Macs. Apple does not publish the model architecture, the training data composition, or the threshold value. It ships as a binary: blocked or not blocked. The user never sees a confidence number. There is no appeal. There is no log entry readable by the person whose content was flagged.
This is the Sensitive Content Warning's real specification, measured against a dog video: a classifier with an unknown false-positive rate, no transparency layer, and no correction path. What Apple said it built and what the classifier actually does are two different things. The dog video is evidence.
A false positive on a pet clip is not the dangerous failure mode here. The dangerous failure mode is the same architecture running in silence, making the same opaque calls, on content that matters more. The dog surfaces the mechanism. The mechanism is the story.
Apple has not commented on the specific incident. The feature remains enabled by default on supported devices running current OS versions.