OpenAI’s latest models, released this week, can determine the location of photos using context clues, which OpenAI calls a “significant breakthrough in visual perception.”
Early testers are uploading photos and asking ChatGPT to “geoguess” where they were taken. The results of this “reverse location search” are surprisingly accurate.
The AI analyzes the photo and spends a few seconds “thinking,” or analyzing the context clues within it. It then provides an answer along with “a long internal chain of thought” about why it came to that conclusion. In the example below, a user took a photo of a library book, and it correctly guessed that it was taken at the University of Melbourne based on a code on the label.
This Tweet is currently unavailable. It might be loading or has been removed.
Another user uploaded a generic photo of a home in Suriname, which looks like it may have been pulled from Google Earth, and ChatGPT gets it right. It’s like the AI version of this guy.
AI-powered photo locators have been around for some time, but the version using OpenAI’s o3 model seems to have popularized it for the masses. That said, the previous model and current flagship, GPT-4o, has the same capability, albeit with lower accuracy.
We tested it by asking both models to guess the location of an image we took this week at the New York Auto Show of Subaru’s newest EV, which debuted at the show.
GPT-4o wasn’t able to pinpoint a specific location, but it correctly said that it might be at an auto show, in either Chicago, New York, or Los Angeles. It based that on “the setting: polished display environment, multiple vehicles in close proximity, informational signage, and people walking around observing the cars.” It incorrectly read the name on the vehicle and called it the “Trailspeed,” not “Trailseeker.
GPT-4o guesses that the photo was taken at an auto show, but doesn’t know which one, and gets the name of the car wrong. (Credit: ChatGPT)
Get Our Best Stories!
Your Daily Dose of Our Top Tech News
By clicking Sign Me Up, you confirm you are 16+ and agree to our Terms of Use and Privacy Policy.
Thanks for signing up!
Your subscription has been confirmed. Keep an eye on your inbox!
The newer o3 model got it right. It “thought for 1m 40s” and then explained that the “blue crossover is Subaru’s new 2026 Trailseeker EV, a model that was first revealed to the public on the show stand at the 2025 New York International Auto Show (NYIAS) inside Manhattan’s Jacob K. Javits Convention Center.”
It crawled Subaru’s vehicle launch page to confirm the Trailseeker debuted at the show, and paired images of Subaru’s booth design with the photo, finding a match with the “lighting, carpeted ‘forest‑floor’ motif.”
OpenAI’s o3 model correctly guesses where the photo may have been taken. (Credit: ChatGPT)
ChatGPT can also combine image recognition with image manipulation. If users upload an imperfect image, ChatGPT can move elements around to answer questions about it. In the example below, it deciphers somewhat illegible writing written upside down on a notebook.
Recommended by Our Editors
“I’ll need to load the image so I can inspect the text. Once I view it, I realize the text is upside down, so I’ll rotate it so it’s readable,” ChatGPT explains as part of its ‘thought’ process. “From there, I can check what’s written and share my findings clearly with the user.” It determines the writing says, “4th February – finish roadmap.”
(Credit: OpenAI)
OpenAI says the model “can still make basic perception mistakes” and that “even when tool calls correctly advance the reasoning process, visual misinterpretations may lead to incorrect final answers.”
Other apps also use AI to identify the location of photos. Geospy, for example, uses context clues such as vegetation and architecture to determine location. It made headlines earlier this year when 404 Media reported that it could be exploited by law enforcement and stalkers since users could ask ChatGPT to geolocate a photos posted on social media.
OpenAI imagines the tech will be “helpful in areas like accessibility, research, or identifying locations in emergency response,” a company spokesperson tells us. “We’ve worked to train our models to refuse requests for private or sensitive information, added safeguards intended to prohibit the model from identifying private individuals in images, and actively monitor for and take action against abuse of our usage policies on privacy.”
About Emily Forlini
Senior Reporter
