Contents
Removing the GPS from a photo is two jobs, not one: deleting the GPS coordinates stored in the EXIF metadata is easy and worth doing, but it does not remove the location the photo still shows, because the visible scene alone can be geolocated by modern models.
Strip the GPS tag first
Phones can write GPS latitude, longitude, and often altitude into a photo’s EXIF metadata block. That tag is separate structured text inside the file, so it can be deleted in seconds with a metadata strip. Do that first. It is worthwhile hygiene because it removes the most direct location channel. The same EXIF block usually stores a capture timestamp and the camera or phone model, so a proper strip should clear the whole block, not the GPS field alone, since a precise time paired with a recognisable place is more identifying than either on its own.
But do not confuse metadata removal with location removal. Most platforms strip EXIF on upload, and that is often mistaken for full protection. It is not full protection. It changes no pixels, and the platform may already have logged the coordinates on its own servers before stripping the public file. Re-saving or screenshotting can also drop the tag, but it leaves the image content untouched; what a clean metadata strip still leaves exposed is the subject of can you be identified after removing EXIF data.
The safe mental model is: EXIF is the label, pixels are the evidence. Removing the label helps. It does not erase the evidence.
The visible scene is the durable half
A storefront, skyline, road sign, licence plate, house number, mountain shape, vegetation, bus stop, reflection, or window view can narrow a place with no GPS tag at all. People have always geolocated images from clues. The newer problem is that models can automate the first pass.
Hays and Efros showed the foundation of this in IM2GPS at CVPR 2008. Their system estimated geographic information from a single image using a dataset of over 6 million GPS-tagged images, and the page-one abstract reports performance “up to 30 times better than chance.” That result is old by modern standards, but it proves the core privacy point: location can be inferred from pixels alone.
Since then, the models have improved and broadened. The GPS tag is no longer the only location signal worth worrying about.
Models can geolocate from pixels alone
Weyand, Kostrikov, and Philbin’s PlaNet work at ECCV 2016 moved geolocation into convolutional neural networks. They report that the model “outperforms previous approaches and even attains superhuman levels of accuracy in some cases.” That is the quote to keep in mind before posting a supposedly anonymous outdoor photo.
Vivanco Cepeda, Nayak, and Shah’s GeoCLIP at NeurIPS 2023 localises images worldwide against a learned, continuous gallery of coordinates. Haas, Alberti, and Skreta’s StreetCLIP work in 2023 shows a publicly available foundation model geolocating zero-shot and, in its page-one abstract, “outperforming supervised models trained on more than 4 million images.” The same work notes OSINT researchers and journalists geolocating social-media imagery, including war footage.
Haas, Skreta, Alberti, and Finn’s PIGEON paper at CVPR 2024 gives the consumer-facing scale of the risk: it places 40.4 percent of its predictions within 25 km of the true location, a city-scale radius. Qian, Chen, and Wang’s EarthWhere benchmark in 2025 shows the frontier shifting into general vision-language models, reporting that across 810 globally distributed images, “Gemini-2.5-Pro achieves the best average accuracy at 56.32%.”
The practical point is not that every photo gives away an exact address. It is that a location guess from pixels is now normal model behaviour.
What to remove beyond EXIF
After stripping metadata, inspect the scene. Crop or cover street signs, shop names, school names, transit signs, licence plates, house numbers, workplace badges, event boards, delivery labels, and distinctive windows. Watch reflections in glass, mirrors, car panels, and sunglasses. A reflection can preserve the one clue you thought you removed.
Indoor photos still leak. A window view can reveal a skyline or courtyard. A whiteboard, package, mail stack, uniform, badge, calendar, or local notice can narrow a place. Mundane backgrounds are often more identifying than people expect because the viewer or model only needs one clue.
If the photo is sensitive, do a self-check before posting. Strip the metadata, crop the scene, then reverse-search or ask a model for a location guess. Treat a plausible city, neighbourhood, venue, or institution guess as a sign that the image still carries location, and edit further until the guess breaks down.
Do it before posting
The upload is the hard boundary. Before posting, strip EXIF, edit the scene, and check the result. After posting, copies, caches, screenshots, platform logs, and reuploads make cleanup harder.
The rule is straightforward: remove the GPS tag because it is easy and direct, then treat the pixels as the real privacy problem. A photo without EXIF can still say exactly where it was taken. Editing the scene itself in full, clue by clue, is the routine this guide points to: how to make a photo untraceable.
Sources
- Hays, Efros (2008). IM2GPS: estimating geographic information from a single image. CVPR 2008.
- Weyand, Kostrikov, Philbin (2016). PlaNet: Photo Geolocation with Convolutional Neural Networks. ECCV 2016.
- Vivanco Cepeda, Nayak, Shah (2023). GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localization. NeurIPS 2023.
- Haas, Alberti, Skreta (2023). Learning Generalized Zero-Shot Learners for Open-Domain Image Geolocalization (StreetCLIP).
- Haas, Skreta, Alberti, Finn (2024). PIGEON: Predicting Image Geolocations. CVPR 2024.
- Qian, Chen, Wang (2025). Where on Earth? A Vision-Language Benchmark for Probing Model Geolocation Skills Across Scales (EarthWhere).