AI Photo Geolocation
An AI-based photo geolocation tool that estimates where images were taken is drawing mixed reactions for being occasionally impressive with landmarks and familiar locations, yet often wildly inaccurate and prone to confident hallucinations. Users report technical issues, training-set bias toward certain regions, and uncertainty about whether EXIF data or visual similarity drives its guesses, alongside speculation about the underlying models. Many are alarmed by the privacy and ethical implications—especially given evidence of insecure image storage and the potential for mass doxxing or OSINT abuse—while others see early signs of a useful capability if accuracy and safeguards improve.
Perceived Accuracy and Behavior
- Reported performance is highly mixed. Some users see “scary accurate” results, even down to specific hiking trails, island identification, or interiors roughly localized by regional architecture.
- Others report large errors: wrong countries or cities, misplacing NYC as Detroit, Chicago as NYC, Montevideo as Buenos Aires, Turks and Caicos as the Bahamas, etc.
- It seems stronger on famous landmarks and highly photographed tourist spots, weaker on generic streets, interiors, and rural or nondescript landscapes.
- Distance errors range from “same city, wrong building” to “wrong continent,” though sometimes it’s directionally close (e.g., off by a few hundred kilometers or neighboring country).
Hallucinations and Explanations
- Explanatory text often mentions non‑existent features: houses in lake photos, palm trees in indoor scenes, signs and cowboy hats that aren’t there, or “English road signs” where no signs exist.
- Some descriptions are eerily specific and correct (e.g., identifying a fire tower or helicopter), but many are generic stories stitched to probabilistic guesses.
Biases and Limitations
- Strong bias toward the US and certain regions; many non‑US photos are mislocated to US cities or to major hubs like Moscow.
- Users note it can be fooled by architectural style (e.g., Venetian or French style misplacing Italian locations).
- Some question whether EXIF data is used; others report stripping metadata and still seeing good/bad performance, so this remains unclear.
Technical Implementation Speculation
- Several commenters infer a visual-embedding + nearest‑neighbors approach, potentially combined with a multimodal model and an LLM for explanations.
- Others argue a simpler classifier might be more appropriate than an LLM, given the hallucinations.
UX and Reliability Issues
- Many report the web app as “broken”: heavy flickering, lockups, constant errors, expired API keys, especially on Firefox/Linux and some Chrome setups.
Ethical and Privacy Concerns
- Strong worries about doxxing, OSINT misuse, and normalizing tools that circumvent intentional metadata removal.
- A serious issue surfaced: user uploads were stored in a Google bucket that was publicly accessible, and the same bucket reportedly contained scraped dating-app images, raising data-handling and consent concerns.