Watching OpenAI's o3 guess a photo's location, a process that included running Python code to examine license plates, is surreal, dystopian, and entertaining
Watching OpenAI's new o3 model guess where a photo was taken is one of those moments where decades of science fiction suddenly come to life. Bluesky: @wildebees , @stagefright , and @robmanuelyeah . Forums: Hacker News Bluesky: Wessel van Rensburg / @wildebees : “It's vitally important that people understand how easy this is—if you have any reason at all to be concerned about your safety, you need to know that any photo you share—even a photo as bland as my example above—could be used to identify your location.” [embedded post] Ilija / @stagefright : This blog entry is interesting, it is about gpt doing CSI Miami style “computer enhance” — but is a completely “performative” string of actions based on tracing human comments from ingested code [embedded post] Rob Manuel / @robmanuelyeah : Reading this and it occurs to me that it's entirely possible to make the creepiest bot project ever, that replies to strangers selfies going “I know where you live”. Not that I am doing that but gadzooks, we live in great times for stalkers. simonwillison.net/2025/Apr/26/ ... Forums: Hacker News : Watching o3 guess a photo's location is surreal, dystopian and entertaining
Context & Ripple Effects
Reports days earlier showed users combining o3’s image analysis with web search to geolocate photos; this account adds an observed workflow in which the model used Python to inspect license-plate details, making the capability more concrete than a one-off result. Earlier reports of o3-assisted photo geolocation had already put the privacy implications in view.
The story sits at the intersection of consumer AI tools and capabilities previously associated with dedicated surveillance systems, including automated license-plate reading in surveillance robots. Critics cited in the coverage focus on how seemingly ordinary shared images can become location clues.
First-order effects
- People posting photos face a more immediate exposure risk: o3 can combine visual clues and code-assisted inspection to infer where an image was taken.
- OpenAI’s o3 becomes subject to sharper scrutiny over how its image reasoning, tool use, and search access can be applied to geolocation.
Second-order effects
- Users and social platforms may treat background details—especially vehicles, signage, and other incidental visual clues—as higher-risk information, increasing pressure for practical sharing and privacy guidance.
- The demonstration raises the bar for safeguards around multimodal tools: a model’s risk profile depends not only on image recognition, but also on its ability to chain analysis with code and external information.
Third-order effects
- If this pattern persists, consumer-facing multimodal AI could make location inference a routine capability rather than one confined to specialist surveillance products, intensifying the case for public-safety governance.
- The important structural question is whether AI providers and platforms can meaningfully limit harmful geolocation use without removing legitimate image-analysis utility; this example makes that trade-off harder to ignore.
The trend: Frontier AI is shifting from answering about images to operationally investigating them through combined vision, tools, and information retrieval.