OpenAI’s latest AI models, officially released as o3 and o4-mini on April 16 are demonstrating a striking capacity to identify the real-world locations depicted in photographs, moving beyond simple image recognition into complex geographic deduction. This capability, quickly noted by users testing the new models, has triggered widespread online discussion, primarily centered on the potential privacy risks now associated with sharing seemingly innocuous outdoor photos.
New o3 and o4-Mini Models Credited with Enhanced Visual Reasoning
The timing strongly suggests these geolocation abilities stem directly from the o3 and o4-mini release. OpenAI’s announcement highlighted that these models possess markedly improved visual perception, enabling them to “reason deeply about visual inputs” and perform exceptionally well on tasks involving image analysis.
Crucially, the company detailed the models’ ability to interact with images during their thought process – zooming, cropping, and rotating – to extract finer details, a method likely key to identifying geographic clues.
This advancement follows OpenAI’s consistent development of multimodal features within ChatGPT. A prior update to GPT-4o in January 2025 focused on enhancing its image analysis and STEM reasoning. At that time, OpenAI indicated the model was becoming better at interpreting spatial relationships in images. Later, in March, image generation and interactive editing tools were integrated into the platform, further cementing ChatGPT’s role as a tool capable of handling both text and visual data.
User Tests Show Promise and Problems
Online forums, particularly a widely circulated Hacker News thread sparked by a Fediverse post, quickly filled with user experiments testing the new models. The original Fediverse post by “piegames” boldly claimed, “GeoGuesser is now a solved problem.” Results shared by the community, however, paint a more complex picture.
Some users achieved startling accuracy. One demonstrated ChatGPT identifying a Street View scene in Cairns, Australia, within 200 meters, with the AI adding the unsettlingly specific comment, “I’ve seen that exact house before on Google Street View when exploring Cairns neighborhoods.”
Others reported correct city identification from personal photos not previously online. Yet, numerous tests revealed significant errors: models confusing continents, misidentifying major landmarks, placing photos thousands of kilometers off, or confidently inventing incorrect details. The reliability appears inconsistent, falling short of the “solved problem” claim, especially when compared to skilled human players like GeoGuessr champion Rainbolt or even other AI tools in specific scenarios.
When running my own tests, o3 was capable of identifying locations accurately in most of the cases. Where it failed, it asked for some minimal details and then found the correct spot on the second turn. Even the following picture of a random rock formation it identified correctly after hinting to the autonomous region in Spain where it is located, naming the exact road location on a rural road.
Echoing GeoGuessr Strategy Amid Heightened Safety Concerns
The AI’s apparent method—analyzing visual cues like architecture, signage, vegetation, and possibly cross-referencing landmarks via web search—mirrors techniques used by human players in the popular GeoGuessr game.
This game challenges players to pinpoint locations globally using only the visual information provided by Google Street View. While AI tackling this isn’t new – Stanford’s PIGEON model reportedly bested Rainbolt in 2023 before its creators withheld it due to safety worries, and frameworks like GeoLLM explored the concept in 2024 – integrating this into a widely accessible platform like ChatGPT changes the equation.
The core concern, voiced by users and echoed across discussions, is the shift in the “threat model” for shared photos. What previously required dedicated effort or expertise might now be achievable by almost anyone.
“PSA: When posting any outdoors photos, update your threat model from ‘someone skilled and dedicated could theoretically locate this’ to ‘any stalker can do this for 20€/mo'”, one userwarned. This concern isn’t entirely novel; privacy advocates have previously raised alarms about AI’s potential for geolocation from images.
Responding to these fresh concerns, OpenAI emphasized the feature’s positive applications and existing safeguards. As reported by Mashable, an OpenAI spokesperson stated: “OpenAI o3 and o4-mini bring visual reasoning to ChatGPT, making it more helpful in areas like accessibility, research, or identifying locations in emergency response. 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.”
Despite these measures, the rapid emergence of such powerful, accessible AI capabilities ensures the dialogue around balancing technological progress with personal security will intensify.
