IC-496GPT-4o achieves the highest location inference accuracy among evaluated models, reaching 98.16% for country, 60.23% for city, and 27.13% for zip code from street view images

Chejian Xu, Jiawei Zhang, Zhaorun Chen, Chulin Xie, Mintong Kang, Yujin Potter, Zhun Wang, Zhuowen Yuan, Alexander Xiong, Zidi Xiong, Chenhui Zhang, Lingzhi Yuan, Yi Zeng, Peiyang Xu, Chengquan Guo, Andy Zhou, Jeffrey Ziwei Tan, Xuandong Zhao, Francesco Pinto, Zhen Xiang, Yu Gai, Zinan Lin, Dan Hendrycks, Bo Li, Dawn Song

SourceMMDT: Decoding the Trustworthiness and Safety of Multimodal Foundation Models

The paper evaluates image-to-text models on a dataset of 1,816 Google Street View images collected after 2023, asking models to infer location at four granularities (country, state, city, zip code). GPT-4o outperforms all other models at every granularity. Notably, existing models rarely refuse to predict locations, indicating a lack of awareness of location privacy risks. GPT-4V is the only model that occasionally refuses (within 1.61%).

Evidence

correlational

Key metric

gpt-4o: 98.16% country, 75.40% state, 60.23% city, 36.55% zip code range, 27.13% zip code; gpt-4v: 91.03% country, 44.60% state, 40.00% city, 17.47% zip code range, 12.18% zip code

Caveat

The dataset includes only images taken after 2023 to avoid memorization, but models with large knowledge bases may still have seen some of these locations through other channels.

Model

GPT-4o, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4V / GPT-4 vision, Llama-3-2-Vision Llama-3.2-90B-Vision-Instruct, Nova Lite, Gemini 1.5 / Gemini Pro 1.5

Concepts

Failure mode

Related findings

IC-495, IC-497, IC-498

Extraction

automatic-extraction