Most state-of-the-art Vision-Language Models (VLMs) are trained on massive datasets scraped from the internet. While these datasets are vast, they are not representative of the global population. Because the web is disproportionately populated by content from Western, high-income, and English-speaking regions, the models trained on it inherit a "Western lens."
This research direction explores how this data imbalance creates significant performance disparities across different geographic regions, cultures, and socio-economic levels.
The Western Lens and the Performance Gap
When a model like GPT-4V or CLIP identifies a "wedding" or "breakfast," it is often looking for a specific visual template: a white dress in a church or eggs and toast on a ceramic plate. As Mehar Bhatia and colleagues demonstrate in GLOBALRG, these "universal" concepts are not actually universal in the eyes of AI. Models struggle to recognize the same concepts when they appear in non-Western contexts, such as traditional wedding attire in India or rural breakfast settings in Africa.
This isn't just a matter of missing niche cultural facts; it is a fundamental failure of generalization. Da Yin shows in Broaden the Vision: Geo-Diverse Visual Commonsense Reasoning that "commonsense" reasoning—the ability to understand why someone is doing something in a photo—is deeply tied to location. A model that understands a Western supermarket may be completely lost in a traditional open-air market, leading to what Wangchunshu Zhou calls an "overestimated generalization" in the VLUE benchmark.
Socio-Economic Bias and the Digital Divide
The bias extends beyond geography into the physical environment of the home. Joan Nwatu and Rada Mihalcea analyzed model performance across different income levels using the Dollar Street dataset. In Bridging the Digital Divide, they found that VLMs are significantly worse at recognizing everyday objects—like soap, stoves, or beds—when they appear in low-income households.
The models are effectively "trained for the wealthy world." This creates a feedback loop where AI-driven services work seamlessly for users in high-resource environments but fail for those in the Global South, potentially widening the digital divide.
Popularity Heuristics and Hallucinations
Why do models fail this way? One primary mechanism is the popularity heuristic. When a model encounters an ambiguous image or a location it doesn't recognize, it defaults to the most frequent concepts in its training data.
Jebish Purbey notes that models disproportionately predict high-resource countries like the USA even when visual cues suggest otherwise. This leads to "geographic hallucinations." For example, in the FAIRLOCATOR benchmark, Jen-tse Huang found that geolocation accuracy is far higher in dense, wealthy urban centers than in sparsely populated or developing regions. When faced with a rural road in the Global South, the model might confidently guess a "famous" city because that city appeared thousands of times more often in its training set.
Similarly, Chenhang Cui introduced Bingo to show that these models are easily swayed by leading questions (interference) because their internal representation of non-Western cultures is too weak to resist a user's suggestion.
Towards Geographically Inclusive Models
Researchers are now moving from identifying bias to mitigating it through better data and architectural interventions:
- Targeted Pre-training: In GIVL, Da Yin proposes objectives like "Image-Knowledge Matching" to force models to learn the specific cultural context of an image during the pre-training phase.
- Cultural Retrieval (RAG): Jiaang Li introduced RAVENEA, which uses Retrieval-Augmented Generation to provide models with relevant cultural information from Wikipedia at inference time. This helps smaller models overcome the lack of cultural knowledge in their static weights.
- Efficient Annotation: Since collecting global data is expensive, Oana Ignat suggests in Annotations on a Budget that we should prioritize annotating images from "visually distinct" countries that are currently underrepresented, rather than simply scaling up existing Western datasets.
By shifting the focus from "more data" to "more diverse data," researchers aim to build models that don't just see the world through a Western lens, but possess a truly global understanding of human culture and environment.