404media.co web signal

Apple's iMessage Nudity Scanner Flagged a Video of a Dog

TL;DR

  • Apple's on-device Sensitive Content Warning, introduced in iOS 17, blurred a video of a dog on her back with paws in the air as suspected nudity.
  • Reporter Joseph Cox found similar Apple-forum complaints going back more than a year, including a dog photo in May 2025 and a trail camera deer image in October 2024.
  • Apple did not respond to 404 Media's request for comment; the scan runs locally and users can disable the feature in settings.

A video of a friend's dog on her back, her little paws in the air, being petted got blurred by iMessage before it would play, flagged as possible nudity. That is the story Joseph Cox at 404 Media reports, and it is a small, specific window into how on-device safety classifiers actually behave in the wild.

The feature is Apple's Sensitive Content Warning, which Apple introduced in iOS 17 a few years ago. It uses on-device machine learning to analyze photos and videos, and, per Apple's own description, "Because they're analyzed on your device, Apple doesn't receive an indication that nudity was detected." A user has to tap through the blur to see the flagged content, and the whole feature can be switched off in settings.

Why the dog case matters is less about one funny misfire and more about the shape of the failure. Cox found similar complaints on Apple support forums going back more than a year: a dog photo flagged in May 2025, and a trail camera picture of a deer flagged in October 2024. Same class of false positive, on a feature Apple has had time to retrain. The reporting is single-anecdote by design, so take it as an existence proof rather than a measured rate.

The honest caveats are worth naming. Apple did not respond to 404 Media's request for comment, so there is no company-side number for how often this happens, no confirmation of whether the underlying model has been updated since launch, and no visibility into whether users even have a proper way to report a misclassification back to Apple. The scan does run locally, which is the design point Apple leans on for privacy, and that same locality is what makes the feedback loop hard.

For anyone building consumer safety filters, whether that is parental controls, workplace DLP, or moderation stacks, the useful takeaway is smaller than the headline. Ship the classifier if you must, but ship a friction-free "this was wrong" path with it, and assume the false-positive tail on animals and other non-human bodies is wider than your test set will tell you.

Shared on Bluesky by 3 AI experts