Contents
Yes, PimEyes offers a free opt-out and filing it is worthwhile, but it does one narrow thing: it suppresses your face from PimEyes’ own results and does not remove the source photos, stop other face-search engines, or make your face unmatchable.
How to opt out, step by step
PimEyes says its free opt-out request form removes a face from current and future searches in its own index. That is the company’s own claim, not an independent audit. According to its own opt-out tutorial, the form asks for a clear, well-lit, full-face photo with nothing covering the face or eyes, so no sunglasses, mask, or heavy makeup. It also asks for an anonymised scan of a government-issued ID to confirm that the face is yours, plus an email address for contact.
PimEyes says requests are typically handled within a few business days, that removal is free and permanent across current and future searches, and that uploaded data is kept no longer than 96 hours. The useful way to read that is procedural: if you are already appearing in PimEyes, the opt-out is a real step to take. It is not a privacy reset.
The technical reason is simple. Face search rests on matching a face image to a stable representation, or embedding. Deng, Guo, Xue, and Zafeiriou’s ArcFace work at CVPR 2019 is a useful anchor for this: it maps faces into an angular embedding space designed for recognition across ordinary image variation. PimEyes can suppress a result inside its own index, but the underlying face in the original photo remains a matchable face.
The biometric trade built into opt-out
The uncomfortable trade is that to be removed from a face-search company, you first hand that company your face and a government ID. That is the most sensitive data in the exchange. You are trusting PimEyes’ 96-hour retention claim, not verifying it.
This matters because modern recognition is not limited to studio-quality portraits. Kim, Jain, and Liu write in AdaFace at CVPR 2022 that “Recognition in low quality face datasets is challenging because facial attributes are obscured and degraded.” Their point is not that low-quality recognition is impossible. It is that modern systems are built to deal with exactly that problem. Low quality, odd lighting, and casual uploads are part of the recognition setting, not a complete defence.
That is why the opt-out should be treated as a suppression request, not as anonymisation. The face-search company may stop returning your face in its product, but the opt-out does not transform the photos already online.
What the opt-out does not remove
The PimEyes opt-out does not touch the source. If a school page, race result, portfolio, employer bio, wedding gallery, social post, or news archive contains your photo, that source image stays online. PimEyes can suppress its own result while the original page remains available to any crawler.
It also does not generalise. Other engines, including FaceCheck.id, Clearview, and FindClone, maintain separate scraped corpora and separate opt-out processes: Clearview AI, for example, takes access and deletion requests through its privacy and requests page. Removing yourself from one index has no technical effect on another.
Most importantly, opt-out does not change matchability. Your face stays a valid query for any recognition system because the published pixels are unchanged. Tkachenko and Jedidi’s Scientific Reports 2023 megastudy showed that 82 of 349 personal attributes were predictable better than random from a facial image, and warned that “facial analysis can strip away privacy.” Suppressing a PimEyes result does not stop a face from being read, classified, compared, or profiled elsewhere.
Why you cannot simply hide instead
A common reaction is to blur, pixelate, or lightly obscure the face in old photos. That is better than doing nothing for casual human viewing, but it is not the same as removing identity from the image. Todt, Hanisch, and Strufe’s Fantômas work in PoPETs 2024 is the relevant warning: face anonymisation can be reversible when it only partially removes identity.
The stronger technical approach is active perturbation, where the image itself is changed to interfere with recognition. Cherepanova, Goldblum, and Foley describe LowKey at ICLR 2021 as “the first such evasion tool that is effective against commercial facial recognition APIs.” That is a separate strategy from opt-out. It belongs to the active image-protection approach covered in image cloaking for facial recognition, not a PimEyes suppression request.
So the practical split is this: opt-out is for an index that already has you, source takedown is for photos that should not be public, and active perturbation is for future images you still plan to publish.
Where opt-out fits
File the PimEyes opt-out if you appear there. Then treat it as one entry on a list, not the list itself. Search for the source pages, request takedown where appropriate, check other face-search engines, and reduce the number of identifiable face photos you publish going forward.
The durable lever is fewer public, identifiable photos. PimEyes can remove you from PimEyes. It cannot remove you from the web, from other face-search corpora, or from the basic fact that a clear face image remains machine-readable.
For the wider cleanup process, see Remove your photos from face-recognition sites.
Sources
- PimEyes, Opt-Out Request Form (service under review; claims are the company’s own).
- PimEyes, How to remove your images from PimEyes search results (opt-out tutorial; claims are the company’s own).
- Deng, Guo, Xue, Zafeiriou (2019). ArcFace: Additive Angular Margin Loss for Deep Face Recognition. CVPR 2019.
- Kim, Jain, Liu (2022). AdaFace: Quality Adaptive Margin for Face Recognition. CVPR 2022.
- Todt, Hanisch, Strufe (2024). Fantômas: Understanding Face Anonymization Reversibility. PoPETs 2024.
- Cherepanova, Goldblum, Foley (2021). LowKey: Leveraging Adversarial Attacks to Protect Social Media Users from Facial Recognition. ICLR 2021.
- Tkachenko, Jedidi (2023). A megastudy on the predictability of personal information from facial images. Scientific Reports 13:21073.