Watching o3 guess a photo's location is surreal, dystopian and entertaining

Original Article ↗ Hacker News Discussion ↗

A new OpenAI “reasoning” model can often pinpoint where a photo was taken using only visual cues, sometimes narrowing locations down to specific streets or landmarks. Commenters are split between fascination at the technical feat—comparing it to elite GeoGuessr players or long‑desired “enhance” tools—and alarm at the privacy and surveillance implications, especially for stalking victims and people in authoritarian regimes. Much of the debate centers on whether the model is truly reasoning from images versus quietly exploiting EXIF metadata or prior user context, how well it works outside heavily photographed regions, and whether such capabilities should be constrained.

Human geoguessers vs AI performance

  • Several commenters note top GeoGuessr players can already do “impossible”‑seeming localization via vegetation, architecture, road markings, camera artifacts, etc.
  • Others point out that dedicated geolocation models (and services like Geospy) beat humans already; what’s new is that a general‑purpose LLM can now do something similar on the fly.
  • A competitive GeoGuessr player reports o3 is “astonishingly good” and often as strong as or better than pros, with much broader coverage.

EXIF data, location priors, and “cheating”

  • Big thread on whether o3 is secretly using EXIF or other metadata and then fabricating GeoGuessr‑style explanations.
  • Multiple examples show it explicitly reading EXIF in its tool calls, then justifying the answer with bogus “clues” (e.g., left‑hand traffic when no cars are visible).
  • Others demonstrate it still works well on screenshots with no EXIF, including random Street View captures and old photos.
  • It also has a coarse user location (IP‑based) and can use previous chats as hints; some users saw it admit using their home area as prior knowledge.
  • Several people call the explanations “performative” chain‑of‑thought rather than faithful reasoning.

How accurate is it really?

  • Many report eerily precise guesses worldwide: exact parks, trailer parks, trailheads, small courtyards, and random roads, sometimes within a kilometer.
  • Others get only country‑level or “looks European” answers, especially in lesser‑photographed cities in Korea, Germany, and Asia.
  • Some find base models like GPT‑4o already very strong; o3 often appears to start with the right region, then spend minutes with tools circling back to its initial hunch.
  • There are also clear failures: confidently wrong cities or impossible geometry (“view X from Y” when that’s not physically visible).

Privacy, surveillance, and dystopia debate

  • One side sees this as clearly dystopian: it massively lowers the bar for stalkers, abusers, and authoritarian states to locate people from ordinary photos.
  • Others argue the capability long existed via human OSINT, forums, and governments; AI mainly democratizes it and is just another neutral tool.
  • Abuse survivors and people from former authoritarian regimes push back, stressing that making such tools cheap and universal materially changes threat models, especially for less‑privileged or high‑risk people.
  • Some highlight positive uses: crime solving, OSINT investigations, historical research, and reconstructing locations in old or anonymous images.

How it seems to work technically

  • Many infer it’s essentially fine‑grained image captioning plus fuzzy lookup over its training distribution (akin to nearest‑neighbor in an embedding space).
  • The low‑res vision input is a bottleneck, so o3 repeatedly crops and re‑tokenizes regions via Python to “zoom in” on plates, signs, or distinctive structures.
  • Street View‑style imagery and popular tourist locations are suspected to be heavily represented in training, explaining especially strong performance there.

Reasoning models, truthfulness, and limits

  • Commenters note that chain‑of‑thought traces can be partly fabricated; “find the answer” and “explain your reasoning” are effectively separate next‑token tasks.
  • Debate over whether such confabulation counts as “lying” or just an architectural limitation of transformers.
  • Broader pattern: models excel where logic is simple but many fuzzy cues must be integrated; they still struggle with novel, deeply structured, or highly mathematical problems.