AI Systems See the World Unevenly: What a 211-Country Study Reveals About Cultural Blind Spots in Vision-Language Models

Vision-language models are everywhere now. They caption photos, moderate content, read documents, guide robots, and answer questions about images inside products that billions of people touch. The market reflects the momentum. Valued at $3.84 billion in 2025, it is projected to reach $41.75 billion by 2035, and Gartner expects 80% of enterprise software and applications to be multimodal by 2030, up from less than 10% in 2024, much of it running on models that read images as fluently as text. The assumption underneath all of it is that these systems see the world the way a capable human would, evenly and everywhere. A growing body of research suggests that assumption is wrong in a specific and consequential way.

Suman Debnath set out to measure exactly how wrong. A developer-relations leader who moved from building production AI systems into peer-reviewed research, he co-authored a study titled Uncovering Cultural Representation Disparities in Vision-Language Models, accepted at IJCNLP-AACL 2025, with Desmond Elliott of the University of Copenhagen, one of the field’s leading researchers in multimodal AI. The study did something the industry rarely attempts at this scale. It tested how well these models understand the world across 211 countries, 5 languages, and 9 categories of everyday images, using 63,300 evaluation samples.

We spoke with Suman about what the study found, why cultural blind spots are an engineering problem rather than a political one, and what it means for the many companies now building products on top of foundation models they did not train and cannot fully see inside.

Most people assume a frontier vision-language model simply understands images. Your study complicates that. What did you find?
That there is no single thing called understanding an image. A model that looks brilliant on a photo of a Western kitchen can stumble on the same scene from a home in South Asia or West Africa, not because the task is harder, but because it has seen far fewer examples of the second. Performance is not uniform across the world. It varies by geography, and the variation is not random. It tracks the distribution of the data the model was trained on.

That is the uncomfortable part. The disparities are predictable. Show me roughly where a model’s training data came from and I can tell you, in broad strokes, where it will be confident and where it will quietly guess. The models are not broken the way a bug is broken. They work exactly as trained. The problem is that everyone treats them as general-purpose vision, when what they actually have is vision that is sharp in some places and blurry in others, and the map of sharp and blurry looks a lot like the map of who produces most of the internet’s images.

What did the study actually measure, concretely?
We built a benchmark that was deliberately global. Most evaluation of these models happens on datasets that are convenient, which usually means Western and English. We went the other direction on purpose, testing across 211 countries and territories, 5 languages, and 9 categories of ordinary things: food, clothing, festivals, the everyday visual vocabulary of a place. In total, 63,300 evaluation samples. The point was breadth. If you only test where a model is strong, you will conclude it is strong everywhere.

And the pattern held together in a way that is hard to wave off. This was not a handful of cherry-picked failures. It was a consistent, measurable gap across a large and diverse set of inputs, which is why it survived blind peer review at an ACL-family venue instead of living as a blog post. When you evaluate this broadly, the geography-shaped gap stops looking like anecdote and starts looking like a property of the systems themselves.

How do you measure something like a cultural blind spot rigorously? It sounds subjective.
It is far less subjective than people expect, and that is the whole argument. Designing fairer evaluation is an engineering problem, not a values exercise. You define the categories, assemble geographically balanced inputs, probe the model with structured and adversarial prompts, and measure accuracy the same way everywhere. The skew people worry about is not in the method. It is in the raw material. Consider that OpenStreetMap building coverage is complete for fewer than 15% of African urban areas, against more than 85% of Western European ones. That kind of imbalance is baked into the data these models learn from long before anyone writes an evaluation.

So the method is almost boring, and it is meant to be. Structured benchmarking across cultural contexts, multilingual evaluation, adversarial samples that push on the edges of what a model claims to know. None of it requires deciding which culture matters more. It only requires refusing to test the model exclusively in the places it already does well. The rigor is in the coverage. If your benchmark is as narrow as the training data, it will happily confirm a competence the model does not actually have.

You are careful to say the models are not broken. So what is the actual risk?
The risk is that the failure stays invisible until it gets expensive. A company takes a foundation model, builds a product on top of it, tests it on the users closest to home, and ships. The model performs well in the demo because the demo looks like the training data. Then it goes global, and in some markets it is quietly, consistently worse, and nobody instrumented for that because the assumption was uniform capability. The gap was always there. It just never showed up in the evaluation they ran. And once a system is making decisions, these gaps are not cosmetic. Commercial face-analysis systems have posted accuracy ranging from 98.2% at the top down to 38.1%, with differences as large as 38 points for women from the Global South.

That is the same failure mode in a higher-stakes setting. For any organization building on a foundation model, these disparities are an inherited risk. You did not create them, you cannot see them from the outside, and they become your problem the moment your users are more diverse than the model’s training data was. The evaluation framework is the only place that risk becomes visible before a customer finds it for you.

You co-authored this with a leading academic in the field. What did that collaboration bring to the work?
Rigor, and a healthy resistance to overclaiming. Desmond Elliott, at the University of Copenhagen, is one of the people who helped shape how the field thinks about grounding language in vision, and working with an academic of that standing changes how carefully you have to build an argument. A practitioner’s instinct is to ship the finding. An academic’s instinct is to first ask whether it survives every alternative explanation. The paper is stronger for that tension, and it is why it reads as evidence instead of opinion.

It also gave the work reach beyond a single lab. The study has been independently indexed by the University of Copenhagen’s research database and by Matilda.science, which means other researchers are tracking and building on it. That is the quiet signal that a result is real. Publishing something proves you said it. People with no stake in your conclusion deciding it is worth citing is what proves it holds.

You came to this as a practitioner, not a full-time academic. Does that change how you read the results?
It changes everything, because I think about deployment first and publication second. As a senior member of the IEEE and someone who spent years helping developers put models into production, I care less about the leaderboard number than about what happens to the thousandth real user in a market the model barely saw. Academics found the disparity. Practitioners are the ones who will either surface it before shipping or read about it later in an incident report.

That is the bridge I care about. The research community has the tools to measure this, and industry has the deployments where it actually bites, and the two do not talk nearly enough. Most teams building on foundation models have never run a geographically stratified evaluation of what they shipped. They assumed the vendor handled it. The vendor measured what was convenient, same as everyone else. Somebody has to close that loop, and it will not be closed by a values statement. It will be closed by better benchmarks.

What has to change?
Evaluation has to get as global as deployment already is. We ship these models to everyone, then test them as if the world looks like the training set. That is backwards. The fix is to make geographically representative evaluation a standard step, the way load testing became standard once systems started serving real traffic, not to lecture anyone about fairness. You would not deploy a service worldwide without testing it under global load. We are deploying vision systems worldwide without testing them on the world.

I am fairly optimistic, because this is tractable. The disparities are measurable, which means they are fixable, or at least manageable once you can see them. What I want is for the evaluation framework to surface the blind spot before a product does, so a team knows exactly where its model is guessing before a user in Lagos or Dhaka finds out first. The models will keep improving. Whether they improve evenly, for everyone, depends entirely on whether we bother to measure it. Right now, mostly, we do not.