Comparing LLMs and Geolocation: GPT-5 Underperforms Against Other AI Models

In June, Bellingcat conducted 500 geolocation tests comparing various AI models, including Google Lens, to assess their accuracy in identifying locations in photographs. Initially, ChatGPT o4-mini-high was the standout performer, with Google Lens excelling as well. When the tests were rerun two months later, including new tools like Google AI Mode, GPT-5, and Grok 4, results shifted. Google AI Mode emerged as the most effective geolocation tool, accurately identifying locations like Noordwijk, Netherlands, while GPT-5’s accuracy declined, often misidentifying locations. Despite improvements in speed, GPT-5 faced criticism for errors. As models continue to evolve, it’s evident that AI’s geolocation capabilities may fluctuate, necessitating ongoing evaluations by Bellingcat to ensure accuracy. Users should remain cautious, as even top-performing models can sometimes provide incorrect results, highlighting the importance of verifying AI output. To support Bellingcat’s investigative work, donations are encouraged.

Source link