What the term means
In research and product docs, undesired content detection usually means a classifier: a model scores text or an image against policy buckets such as sexual content, hate, violence, self-harm, and harassment. The best-known write-up is OpenAI’s 2022 paper A Holistic Approach to Undesired Content Detection in the Real World, later sold as the Moderation API. Other vendors ship similar “moderation” or “guardrail” APIs. The job is to guess “this looks like category X,” not “this is the same file as yours.”
Classifiers, not fingerprints
A classifier does not need a copy of your original photo. It can flag a brand-new deepfake or a prompt that asks for one. It also does not know whether the person depicted consented. “Sexual” is not the same as NCII. Consensual adult nudes, medical images, and non-consensual leaks can all look “sexual” to a model.
How it differs from hashing
- Hashing (PDQ, PhotoDNA, MD5) matches a known fingerprint. AntiNCII creates PDQ for photos and MD5 for videos in your browser. The file stays on your device.
- Classifiers score new material against a taxonomy. They can catch unseen or AI-generated nudes that no hash list holds yet — and they can miss, or flag the wrong thing.
- Hash lists are strongest against later copies of the same file. Classifiers are strongest as a coarse filter at platform scale.
- Neither scans end-to-end encrypted chats in a way we can promise. Neither wipes Google Images.
Industry CSAM systems often combine both: hash lists of known child-abuse material plus classifiers for new files, then human review. OpenAI’s own CSAM guidance says its Moderation API is not a substitute for dedicated CSAM detection. The sexual/minors bucket is text-only on that API. If a minor is involved, use official reporting, not this site.
What it cannot do for NCII
- It cannot prove you are the person in the image
- It cannot tell a platform to delete every copy
- It cannot see inside Telegram, WhatsApp, or other locked chats
- A “not sexual” score does not mean the abuse did not happen
What AntiNCII does instead
We do not send your photo or video to OpenAI, PhotoDNA, or any undesired-content API. We do not use Microsoft PhotoDNA. You still report in-app. A fingerprint may help a participating public platform catch a matching upload later. Removal is never guaranteed. More on that workflow: How will AntiNCII work? EU rules that prohibit some intimate-deepfake systems are a different question: Unlawful AI.