A photo of a washer and dryer, with no GPS tag attached, was enough for an AI tool called Raven to point to a specific Airbnb listing in Charlotte, North Carolina. The demonstration came from Graylark Technologies, the company that sells Raven to police, and the Los Angeles Police Department and the Miami-Dade Sheriff’s Office are among the agencies reported to have bought access.
Graylark chief executive Daniel Heinen announced the indoor feature on LinkedIn earlier this month, describing “the ability to geolocate interior images across millions of short-term rental units worldwide.” Raven was previously sold as GeoSpy, a tool built to place outdoor photos from clues such as buildings and terrain. Other demos matched a bathroom corner to a Vrbo rental in Montecito, California, and an apartment to images on Zillow in San Francisco. A separate HotelID feature is pitched at hotel rooms, reportedly using the toiletries on the counter.
How a room gives itself away
Outdoor geolocation and indoor geolocation are two different tricks, and the difference decides who is exposed.
The outdoor version reasons from general clues. A Stanford team led by Lukas Haas and Chelsea Finn showed how far that can go with PIGEON, a model presented at the CVPR computer vision conference in 2024: on Street View images it put 40.4% of its guesses within 25 kilometers of the true spot. A 2025 study by O. Grainge and four co-authors tested 25 vision-language chatbots and found the reverse pattern. They did poorly on generic street scenes, but the best ones placed 61% of social-media-style photos within one kilometer.
Raven’s interior feature does not work out a city from the shape of a radiator. It matches. The software picks out distinctive details, such as tile patterns, fixtures and appliances, and looks for the same room among photos it already holds from listing sites, even when the furniture has moved. That makes it more powerful than a reverse image search, because the two pictures do not need to be identical. It also sets a hard limit: a room that has never been photographed and published online gives the software nothing to match against. Rentals, hotels and homes recently listed for sale are the obvious targets. A private bedroom that never appeared on a listing site is far harder to place this way.
What the companies and police say
Heinen presents the product as a tool against crime. “Short-term rentals can be used as locations for human trafficking and exploitation,” he wrote. Whether it delivers is less clear. The Miami-Dade Sheriff’s Office has said its earlier use of GeoSpy was limited and had not led to significant arrests.
The listing platforms did not sign up for this. Airbnb said it has no connection to Graylark, strips location metadata from listing photos, shows only an approximate location for each host, and has safeguards against automated data collection; it said it was looking into the matter. Zillow said scraping its site breaks its terms of use and that it pursues legal options when needed. Vrbo did not comment.
There is a record that explains the unease. After GeoSpy’s public launch, users in the company’s Discord server asked for help stalking women, and Graylark then limited the tool to government and enterprise customers. The Grainge study names the same risk in general terms: “Precise geo-localization from publicly shared images can be exploited for stalking, surveillance or discrimination.”
What it means if you post photos from home
Old advice still covers the easy leak. The U.S. Army warned its soldiers back in 2012 that most smartphones write GPS coordinates into every photo, and that a month of tagged posts can sketch out a daily routine. Turning off location tagging in the camera app fixes that part. It does nothing against a matcher like Raven, which never reads the metadata at all. A few practical points follow:
- If you host a rental, your listing photos are the reference images. Anyone who later posts from that room, guest or host, can be tied back to the address.
- If your home was listed for sale or rent in recent years, those pictures may still sit on property sites. Ask the agent or platform whether old galleries can be taken down.
- Distinctive fixtures carry more signal than clutter. A plain white wall is harder to match than a patterned tile splashback or an unusual window.
- Strip location data anyway. It stops the cheaper, more common forms of tracking even if it cannot stop this one.
None of this is a reason to stop sharing pictures. It is a reason to know which pictures already exist. Grainge and his co-authors expect the accuracy of these models “is likely to improve in the future,” and Heinen has written, “We’re only beginning to scratch the surface of what’s possible.”
Further reading
- arXiv: PIGEON: Predicting Image Geolocations (Haas et al.)
- arXiv: Assessing the Geolocation Capabilities, Limitations and Societal Risks of Generative Vision-Language Models (Grainge et al.)
Image: Tony Hisgett, CC BY 2.0 via Wikimedia Commons
This article was produced with AI assistance and reviewed by our Editorial Desk. How we work.
