.png)
6 hours ago
2
Hunting landmarks, squinting at Google Earth, and flipping between EXIF readers used to eat up my afternoons. Yesterday, OpenAI dropped its newest model, o4-mini, and it upended my entire OSINT workflow.
I fed it a quick rooftop shot of the Manhattan Bridge from the top of SoHo house in Brooklyn.
In under 30 seconds, o4-mini returned:
No plugins. No guesswork. And within 100 feet of where I was standing when I took the photo. This was almost an “aha” moment for me. Geolocation felt like less guesswork, but more-so a reliable tool in my belt. What if every analyst could go from photo to coords with a single prompt?
Despite living in a data driven world, turning a photo into an exact latitude/longitude is often a multi‑hour process. You juggle:
- Manually spotting landmarks
- Cross‑referencing old satellite tiles
- Copy‑pasting metadata between apps
It’s tedious. Every minute spent hunting information is a minute you’re not asking why this information matters.
o4-mini is pretty straightforward in this context:
- Feature Extraction It tags roads, towers, even surf breaks — no manual tuning needed.
- Change Detection New construction or burn scars? Flagged in seconds.
- EXIF & Reverse‑Geocoding Metadata parsing and “Where is this?” all in one call.
Note: While ChatGPT is capable of analyzing EXIF data, the app strips all EXIF data from an image prior to being sent to the model.
- Multi‑Modal Fusion Works equally well on satellite imagery, drone video or your phone’s photos.
I’ve also tested the model on more obscure photos that I asked some friends for (Thanks Cal and Jake!)
With some of the more obscure photos that lack signage, buildings, etc — o4-mini can require some context (region, persons, etc). I found it to be quite accurate at an 80% success rate over 15 photos, some requiring additional context and prompt engineering, others were spot on in one shot, placing the photos at coordinates well within an acceptable range.
The most interesting case was a photo taken in San Clemente showcasing bioluminescence. With minimal context, the model was able to clock the photos coordinates within a 32 mile range. I even attempted to throw the model off by saying the photo was taken today:
There are THOUSANDS of beaches in the world. That POC really set into stone how the intel analysts workflow is going to change drastically over the next several years.
A Co‑Pilot, Not a Replacement
o4‑mini isn’t a substitute for analyst intuition — it’s a force multiplier that offloads the repetitive tasks, allowing analysts to focus on the nuance. For those building foundational skills in OSINT or GEOINT, this balance of speed and analytical control is indispensable.
What’s Next?
No silver bullet here — real world ops still need:
- Human‑in‑the‑Loop Checks for false positives
In some cases, hallucination would occur and provide an unrealistic coordinate. Further context would be provided, which typically helped the model from hallucinating.
- Attribution metadata detailing the model’s reasoning behind each coordinate estimate
- Continuous Retraining to avoid “data blind spots”
- Add AI model with specific training tailored to GEOINT into a GIS stack
Plug o4‑mini into your GIS stack or OSINT workflow, and you’ll find yourself spending far less time squinting at maps and far more time telling the story behind the photos.
Thanks for reading!
