From Beverly Hills to Refugee Camps: The Same AI Injustice

We recorded this episode while wars were being fought and AI was already part of the story. It shapes how conflict is waged, how it's reported, and how displaced people try to survive it.

A recent report from the Center for Strategic and International Studies found that AI systems used in foreign policy simulations show systematic bias toward escalation and conflict scenarios. These systems don't just reflect decisions. They shape how threats are understood and how force is justified.

Our guest is Anissa Abeytia, an AI governance and digital equity specialist with the UNHCR Innovation Service and the Center for AI and Digital Policy. She has worked in over 20 low and middle income countries, studying what happens when AI systems designed elsewhere land in crisis zones. We invited her because of her article in The New Humanitarian, Model Drift: How Subtle Shifts in AI Responses Could Undermine Crisis Response, which named something most AI governance conversations are missing entirely.

Information Is the Currency

In conflict zones and humanitarian emergencies, people need to know if an aid corridor is open, if a ceasefire is holding, if it is safe to go home. These things change by the hour.

Anissa's concept of model drift addresses this directly. It is not the technical version — predictive inaccuracy as data patterns shift over time. It is something more specific and more dangerous. AI systems have a cutoff date. They get lazier with each query. They were not trained on what is happening right now on the ground. And when they don't have the information they need, they don't say so. They generate something that sounds coherent.

In a crisis zone full of rumors and fear, a confident but wrong answer causes real harm.

"Hallucination," she says, "is a narrative failure." This reframes the governance question. It is not just about accuracy. It is about who bears the cost of inaccuracy, and in humanitarian settings, that cost lands on the people who are already most vulnerable.

The Language Problem Nobody Is Solving

Most large language models were trained by scraping English from the internet. Not even British English in some cases.

When a journalist in Ethiopia asks an AI who is responsible for an atrocity, the answer shifts significantly depending on whether the question is asked in Amharic or English. The framing, the associations, the historical context embedded in the training all change with the language. Languages like Arabic, Japanese, and Chinese have relational grammatical structures that the token-by-token architecture of most AI systems is not built to handle. Within a single language, dialects and regional variations carry legal and cultural significance that a human rights lawyer would catch and an AI almost certainly will not.

And then there is the most invisible problem. Some of the communities most affected by humanitarian crises do not have words in their languages for the technologies being built to serve them. Some have very limited written literacy at all. An AI system built on the assumption of digital literacy excludes the people it claims to help before anyone has opened the app.

Anissa's point is not that this is unsolvable. It is that nobody designing these systems is asking the question.

The Architecture Is the Policy

AI governance conversations focus on data, privacy, and algorithms. Anissa argues they are looking at the wrong level.

The architectural logic underneath those things is what actually shapes outcomes. And in the real world, architectural logic is called policy. Policy does exactly what it is designed to do. If the people designing AI systems come from particular backgrounds, hold particular assumptions, and optimize for particular outcomes, the system reflects all of that, not as a bug but as a feature.

Technology has never been neutral. It has always served power. The question is whether we are willing to say that out loud and design differently.

What Doing It Differently Actually Looks Like

Anissa is building Humma, a healthcare AI in Los Angeles. Instead of scraping data, the team pays people from different communities to sit down for conversations so the model can learn what it actually means to be Japanese American or Mexican American in LA, including everything that does not fit a compliance checklist.

One of the characters they developed is Carlos. He has heart disease. His father died of heart disease. But his mother does all the cooking and he cannot tell her to change. A healthcare system that cannot account for that will fail him. An AI that cannot account for that will recommend things he will never follow. This is community-centered design in practice. It is slower and more expensive. It is far more likely to work.

She also outlines three things that humanitarian diplomacy has spent decades getting right, and that AI governance has almost entirely ignored:

• Narrative consistency across languages and platforms. Not just translating words but understanding what a language can and cannot convey.

• Community-centered mechanisms for detecting misalignment. Ongoing, iterative, involving the people being affected, not a one-time audit.

• Local oversight and participation. You cannot see behind your own head. Bias is invisible to the people it doesn't affect.

  
  

The Beverly Hills Problem

A woman in South Central Los Angeles pays the same monthly rate for internet as someone in Beverly Hills. She gets slower speeds because her infrastructure is older and her neighborhood maps as lower value. This is digital redlining. It is the same structural logic as housing redlining, now running through digital infrastructure.

Anissa's career has moved between refugee camps, inner cities, and crisis zones, and her conclusion is that the injustice in all of these places runs through the same policy and legal architecture. The geography changes. The mechanism does not.

AI is not correcting this. In her view, it is being used to extend it. And the divestment of public funding from civil society happening right now is removing the organizations best positioned to push back, at precisely the moment when the stakes are highest.

The Question Underneath Everything

We are building the technology first and asking what it is for second. Anissa argues we have it backwards. Her PhD research asks who, in the future, will be able to think. Who will be able to ask questions and make decisions. That is not an abstract question. It is a governance question, and the answer is being shaped right now by choices that most people affected by them have no part in making. Her vision for AI comes down to a word that has largely disappeared from the technology conversation: democracy. Not as a value to be stated, but as a practice to be maintained.

She says, "Democracy dies when we don't participate. The way our current world is structured, AI is a way to get us to be apathetic. To not think. To not participate." "You're smart enough and you know enough. We need every single one of you to participate in AI discussions."

Between Us and the Machine is hosted by Juliet (Mission AI) and Margot van Brackel. Episode 4 features Anissa Abeytia, AI governance and digital equity specialist with the UNHCR Innovation Service and the Center for AI and Digital Policy. Listen and join the conversation.