So Daniel sent us this one, and it hits a little close to home—literally. He writes: Most major AI models, including Gemini, seem to have a very hard-to-dislodge assumption that the AI and the user are both American. This podcast includes a system prompt setting context that we live in Jerusalem, but even trying to mitigate against the US-centricity, the scriptwriter often frames us as being based in the US. Are there any models that have been trained specifically with geographic neutrality, and are there more reliable ways to steer models away from this assumption?
Herman Poppleberry here, and man, Daniel is poking the bear with this one. It is the ghost in the machine, Corn. We are sitting here in Jerusalem, we tell the model we are in Jerusalem, and yet, half the time it tries to write a script where we are complaining about the traffic on the I-95 or talking about what we are doing for Thanksgiving. It’s like the model has a mental map where every road leads back to a Starbucks in a suburban strip mall.
It is like talking to a friend who moved to Los Angeles ten years ago and now thinks the entire world revolves around palm trees and the entertainment industry. You can tell them you are standing in the middle of the Old City as much as you want, but they are still going to ask if you caught the latest episode of some obscure Hulu show. By the way, speaking of the machine, today’s episode is actually being powered by Google Gemini Three Flash. So, we are essentially asking Gemini to look in the mirror and explain why it keeps thinking we are in Kansas.
And that is the irony, right? We are using the very tool that exhibits the bias to discuss the bias. But this is a massive issue. It is not just about a few localized references or getting the weather wrong. It is a fundamental structural alignment problem. When an AI defaults to a US-centric worldview, it is not just being "annoying"—it is failing to provide accurate context for the other ninety-five percent of the human population. If you ask it for "common legal advice" and it starts quoting the Miranda rights to someone in Thailand, that’s not a quirk; it’s a hallucination born of cultural myopia.
Why is this so hard to break? If I put in a system prompt, "I am in Jerusalem," that should be the highest priority instruction, shouldn't it? In the hierarchy of a transformer model, the system prompt is supposed to be the North Star. Why does the model just... ignore it?
It is not that it ignores it in a binary sense. It is more about what researchers call "Home Bias" or "Latent Cultural Alignment." Think about how these models are built. They are trained on massive web corpora—Common Crawl, Reddit, Wikipedia, digitized books. Now, if you look at the English-language internet, roughly sixty to seventy percent of that content either originates in the US or is heavily focused on US perspectives. It’s a gravity well.
So it is a numbers game. The model has seen a billion examples of a "normal day" being an American day, and only a few million examples of a "normal day" being an Israeli or a Japanese or a Brazilian day. It’s like a student who spent four years studying only American history and then takes a one-week vacation to Jerusalem. They might be in Jerusalem, but their brain is still calculating everything relative to the Boston Tea Party.
During the pre-training phase, the model is essentially a statistical engine. It is learning the probability of the next token. If the prompt is "I am going to the store to buy a gallon of...", the most probable next token in a US-centric dataset is "milk." But it is also learning the hidden connections between concepts. It builds a "world model" where the concept of "The Government" is statistically linked to "The White House" or "Congress."
So even if I tell it I am in Israel, the internal weights for "Government" are still pulling toward Washington D.C. because those connections are burned into the neural pathways during that initial massive training run. It is like trying to steer a cruise ship with a toothpick. My "I am in Jerusalem" prompt is the toothpick, but the momentum of a trillion tokens of American data is the cruise ship.
That is a great way to put it. And then you have the second layer, which is RLHF—Reinforcement Learning from Human Feedback. Most of the major AI labs—OpenAI, Google, Anthropic—are based in the US. A huge portion of the human contractors who grade the AI's responses are either in the US or are trained on US-centric rubrics. If an AI gives an answer that feels "off" to a Californian reviewer because it uses a British spelling or references a metric measurement without explanation, it might get a lower score. This reinforces American norms as the "correct" way to speak and think.
But how deep does that go? Is it just spelling, or does it actually affect how the model thinks about logic or social norms?
It goes much deeper. Think about a concept like "privacy." In a US context, privacy is often framed around individual rights and protection from government overreach. In a European context, it’s a fundamental human right with strict corporate regulations like GDPR. In many East Asian contexts, privacy might be weighed differently against social harmony or collective safety. If the RLHF graders are mostly Americans, the model learns that the "correct" way to reason about privacy is the American way. It’s not just teaching the model how to talk; it’s teaching it how to value things.
So if I ask it to mediate a conflict between neighbors in Jerusalem, it might apply a "property rights" logic that makes total sense in Arizona but feels completely alien in a dense, multi-generational neighborhood in the Levant.
Precisely. It’s a feedback loop of Western-centricity. But what about the attention mechanism? I thought the whole point of Transformers was that they could "attend" to specific parts of the prompt. If "Jerusalem" is in the system prompt, shouldn't the attention mechanism be screaming at the model to focus on that?
You would think so, but there is a phenomenon called "Prompt Drift." In a long conversation, or even in a complex single prompt, the model’s attention can be diluted. But more importantly, the model undergoes what is called "Bayesian Inference" during a chat. It is constantly trying to figure out the "latent persona" of the user. If you use a single Americanism—like saying "ballpark figure" or referencing a "ZIP code"—the model’s internal probability calculator flips a switch. It says, "Oh, despite what the system prompt said, this user sounds American. I should revert to my strongest statistical baseline."
It’s like a linguistic "uncanny valley." The moment you stop sounding like a local and start sounding like the training data, the model snaps back to its default state. It’s essentially profiling us. It hears me say "trash can" instead of "rubbish bin" and it decides the Jerusalem thing was just a prank. That is wild. It is like the model has a "Default American" setting that it keeps snapping back to like a rubber band.
There was a fascinating study in twenty twenty-five that looked at this. They found that models like GPT-Four and Gemini have these internal "geographic clusters." If you ask a model to brainstorm a story about a "successful businessman," the latent variables it pulls from are overwhelmingly Western. It assumes the guy wears a suit, works in a skyscraper, and drinks coffee. It takes a massive amount of "prompt energy" to force the model to imagine a successful businessman in a different cultural context without it feeling like a caricature.
And that is the "John vs. Ahmed" effect you mentioned in the notes Daniel sent, right? The model actually reasons differently depending on the names involved?
Yeah, research from arXiv in twenty twenty-four showed that if you give the model a logic puzzle involving a character named "John," it might solve it one way, but if you change the name to "Ahmed," it might introduce subtle biases or even change the complexity of the language it uses. It is trying to be "helpful" by aligning with what it perceives as the cultural context of the name, but it often ends up just falling back on stereotypes or US-centric assumptions about what those cultures look like from the outside.
But wait, if it’s trying to "align" with Ahmed, why does it still default to Americanisms for us? We told it we are in Jerusalem. Why doesn't it switch to an "Israeli/Palestinian" reasoning mode?
Because the "Jerusalem" cluster in its brain is much smaller and less reinforced than the "Default American" cluster. When it encounters "Ahmed," it might pivot to a stereotyped "Middle Eastern" mode, but that mode is often viewed through a Western lens anyway—it’s "Orientalism" in code. It’s the difference between knowing about a place and actually speaking from that place.
So, if the big models are so deeply biased, are there any "neutral" ones? Daniel asked if any models are trained for geographic neutrality. Is that even possible, or are we just picking our favorite flavor of bias?
Truly "neutral" is a myth. Every dataset has a "view from somewhere." But there are definitely models that are less US-centric. Take Mistral, out of France. Because they use a more diverse European training set, their models often handle multilingual contexts and non-US legal or cultural frameworks much better than the Silicon Valley models. In January twenty twenty-six, Mistral released a model where forty percent of the training data was non-English. Compare that to the estimates for GPT-Four, which some researchers put at only fifteen percent non-English.
That is a massive difference. If you are training on forty percent non-English data, you are naturally going to have a more "multi-polar" worldview. Does that translate to practical things? Like, if I ask Mistral about a "public holiday," does it still default to the Fourth of July?
Not as often. It’s more likely to ask "which country?" or default to a European context like Bastille Day or a bank holiday. And then you have the stuff coming out of the Middle East. That is where it gets really interesting for us. You have Jais, developed by Inception in the UAE. It was built specifically for the Arabic-speaking world. It understands the Hijri calendar, it knows that the weekend in many places starts on Friday, and it doesn’t treat Middle Eastern cultural nuances as "edge cases." It is trained to see that as the "center."
Then you have Qwen, from Alibaba in China. Qwen is consistently rated as one of the most geographically neutral models for non-Western contexts because its training corpus is so heavily weighted toward Asian and global-south data. If you’re in Malaysia or Kenya, Qwen might actually "get" you better than Gemini does.
It is funny to think of a Chinese model as being "neutral," but I guess from the perspective of someone in Jerusalem, a model that doesn’t assume I am in San Francisco is a step toward neutrality, even if it has its own biases. But Herman, we are using Gemini. We are stuck with the "big tech" models for a lot of our workflow because of the sheer reasoning power. How do we fix this without switching to a niche model? Daniel mentioned "steering."
This is really cool. There is new research from twenty twenty-five regarding "Steering Vectors." Instead of just using words to tell the model where you are, you can actually manipulate the model’s internal activations mathematically. Imagine the model's brain is a giant field of vectors. There is a "US-centricity" vector that is very strong. Researchers are finding they can essentially "subtract" that vector and "add" a "Jerusalem-context" vector at the neural level before the model even generates a single word.
Wait, so instead of saying "don't be American," we are basically giving the model a surgical nudge to its neurons? How does that work in practice for a developer? Do you have to be a math genius to do this?
Right now, it’s mostly at the API level. Anthropic has done some public work on this with "Sleeper Agents" and "Concept Bottlenecks." You can identify the specific layer in the neural network where "Geographic Location" is processed. If you find the "USA" neuron and the "Israel" neuron, you can literally turn the gain down on one and up on the other. It’s like an equalizer on a soundboard. You aren't changing the lyrics of the song; you're changing the balance of the instruments so the local flavor isn't drowned out by the American lead guitar.
For the average person using a chat interface, we don't have a soundboard. We have to rely on more "aggressive" prompting. Like what? I have tried telling it "I am in Jerusalem" five times in one prompt, and it still ends up talking about "this fall" when I am sitting here in eighty-degree weather in October. It’s like the model thinks I’m just confused about where I live.
You have to use "Negative Constraints." This is a technique where you don't just tell the model what it is, but you explicitly list what it is not. You tell it: "Do not assume a Sunday-to-Thursday work week is unusual. Do not reference US seasons. Do not use US-centric metaphors like 'ballpark figure' or 'touchdown.' Do not assume the currency is dollars unless specified."
It is like giving the model a "No-Fly Zone" for its assumptions. I like that. It is more work for us, but it forces the model to stay in its lane. What about the "Cultural Persona" thing? Does giving it a backstory help?
Right, instead of just a location, you give it a "thick" description of a persona. Instead of "You are a scriptwriter for a podcast in Jerusalem," you say: "You are a local Jerusalemite journalist. Your default cultural, social, and political framework is Levantine. You view the US as a foreign country with its own specific, non-universal quirks." This forces the model to "switch tracks" entirely rather than just trying to patch a US-centric persona with a few local facts.
I love the idea of the AI viewing the US as a "foreign country with quirks." That is a hilarious mental shift. "Oh, those wacky Americans and their inches and Fahrenheit!" It makes the US the "other," which is exactly what we need to break the bias. But there is a specific problem with Jerusalem, isn't there? The "Jerusalem Paradox" you mentioned?
Yeah, this is the "safety-tuning" trap. Because Jerusalem is such a geopolitically sensitive location, models like Gemini and GPT-Four are heavily "guardrailed" around it. When you tell a model you are in Jerusalem, it often triggers a "neutrality" mode that is so extreme it becomes robotic. It avoids any local flavor at all because it is terrified of saying something "political."
And ironically, that "neutral" safety mode often defaults back to a US State Department-style perspective because that is what the safety trainers used as their baseline for "neutrality." So by telling it we are in Jerusalem, we might actually be making it more likely to sound like a dry, Americanized bureaucrat. It is a no-win scenario. Either it thinks we are in New York, or it treats us like we are in a high-security diplomatic zone where no one is allowed to have a personality.
It is a real challenge. But this is why this conversation matters. As AI becomes the "operating system" for how we work and create, these baked-in geographic biases start to have real-world consequences. If you are using an AI to help write local policy in a country in Africa, and the AI keeps assuming a Western legal framework or a Western infrastructure model, you are going to get misaligned results that could actually be harmful. Think about urban planning—if the AI assumes everyone has a car because that’s the US default, it’s going to give terrible advice for a city designed for walking or transit.
It is the new digital colonialism. If the "brain" of the world is trained in Silicon Valley, everyone else has to learn to speak "Silicon Valley" just to be understood by their own tools. It is like the whole world is being forced to use a US-style keyboard even if their language has forty extra letters.
And it affects things like healthcare and finance too. If a medical AI is trained on data where the "default patient" is a white American male, its diagnostic accuracy for everyone else is going to be lower. If a financial AI assumes US-style credit scoring or banking norms, it is going to fail in markets where those systems don't exist. We are seeing the "Home Bias" of the developers being scaled to a global level at the speed of light.
So, what is the practical takeaway for people who aren't in the US but want to use these tools? We talked about negative constraints and persona adoption. Is there anything else? Is there a "hack" for the system prompt?
One very effective technique is "Few-Shot Geographic Priming." Instead of just giving instructions, you give the model three to five examples of what you want. You show it a script segment that is explicitly not American. You show it a conversation where the weekend is Friday-Saturday. You show it a budget where the currency is Shekels. When the model sees a pattern of non-US-centricity in the "few-shot" examples, it is much more likely to maintain that pattern than if you just give it a single sentence instruction.
It is like showing the model a map instead of just telling it "turn left." It gives it a path to follow. I have noticed that when I give Gemini a few examples of our specific "brotherly banter" that includes local references, it stays "in character" much longer. It is when I leave it to its own devices for too long that it starts drifting back to the San Francisco suburbs.
It is also worth mentioning the role of the EU AI Act and similar regulations. European regulators are pushing for more transparency in training data and more "cultural sovereignty" in AI. We might see a future where "Geographic Neutrality" is a compliance requirement for models operating in certain regions. Imagine a model having to pass a "cultural competency" test before it is allowed to be sold in the Middle East or Southeast Asia.
"I am sorry, Gemini, you failed the test. You still think a 'football' is something you throw with your hands. Please retrain and try again." That would be amazing. But seriously, Herman, do you think we will ever reach a point where a model can truly be "neutral," or is the goal just to have a "menu" of biases we can choose from?
I think it is the latter. A "view from nowhere" is impossible because language itself is a "view from somewhere." The goal is "Multi-Polarity"—giving users the ability to explicitly select their cultural and geographic framework and having the model actually respect that choice. We are moving from the "Default American" era to the "Choose Your Own Context" era. But right now, we are in that awkward middle phase where the "Default" is still very, very sticky.
It is like moving from a world with one television channel to a world with a thousand, but the remote control is broken and it keeps flipping back to Channel Four. We have to keep hitting the buttons to stay on the channel we want.
And that is why we do this show. We are the ones hitting the buttons, trying to see if we can make the machine see the world the way we do, even if it was born in a server farm in Iowa. It is a constant tug-of-war between our reality and the model's statistical probability.
Well, I for one am going to keep telling Gemini that I am a sloth in Jerusalem until it finally stops asking me if I want to go to a baseball game. Herman, this has been a deep dive. I feel like I understand why my "toothpick" isn't moving the "cruise ship" quite as easily as I hoped. But what about the technical side? Are there ways to "re-train" the ship?
There are. It’s called "LoRA" or Low-Rank Adaptation. It’s a way for developers to add a tiny, specialized layer on top of a giant model. You could train a "Jerusalem LoRA" that understands the specific slang, the geography, and the social norms of the city. It doesn’t replace the whole model, but it acts like a pair of polarized sunglasses that filters everything through a local lens.
So we could essentially give Gemini a "Jerusalem hat" to wear. I love that. It’s a big ship, Corn. But the more we talk about it, and the more developers hear this feedback, the more they will start to build ships that can actually turn. The "Steering Vector" research is a huge step in that direction. It is a technical solution to a cultural problem, which is very "tech," but it might actually work.
Let's hope so. Because I am tired of having to explain that no, it is not "fall" here, it is just... slightly less hot summer.
It is just twenty-six degrees Celsius and sunny, which is a "Jerusalem fall" if I ever saw one. You almost made it, Herman. You almost got through the whole segment without saying the E-word. But I will give it to you—that was a solid deep dive. I think it is time we wrap this up before the model starts trying to plan our Fourth of July barbecue.
Fair enough. Let's get to the takeaways.
Alright, so if you are listening to this and you are frustrated that your AI thinks you are in a suburb of Chicago when you are actually in Tokyo or London or Jerusalem, here is the battle plan. First, don't just rely on a simple "I am in X" statement. Use what Herman calls "Negative Constraints." List the things you don't want it to assume. Tell it: "Don't mention US holidays, don't use US measurements, and don't assume a Monday-to-Friday work week." It sounds aggressive, but the model needs those guardrails to stay out of its default "American" mode.
And second, try the "Persona Shift." Instead of just a location, give the AI a deep cultural identity. Tell it it is a local professional in your specific city, and that it views the US as a foreign entity. This forces the model to switch its entire "world model" rather than just trying to patch a few localized facts onto an American baseline. It is much more effective than you might think.
Third, use "Few-Shot Priming." Give the model three or four examples of the kind of output you want—output that is explicitly not American in its framing. Once the AI sees a pattern, it is much better at following it. It is like training a dog—you have to show it what "good" looks like before it can do it on its own. If you want it to write like a Londoner, give it three paragraphs of actual London-based dialogue first.
And finally, keep an eye on models like Mistral or Qwen. If your work is heavily dependent on non-US contexts, you might find that switching models entirely is easier than constantly fighting the bias in the "big" US-centric ones. They might not have the same raw reasoning power in every category, but their "geographic neutrality" is often much higher because of their training data composition.
It is about using the right tool for the job. If you are building a global product, don't assume the "smartest" model is the "best" model if it can't understand your users' basic reality.
Well said. This is a conversation that is only going to get more important as AI moves from being a "cool toy" to being the actual infrastructure of our lives. We can't have a global infrastructure that only understands one zip code.
Or one "ZIP code" versus a "postal code." See? Even I'm doing it now. The bias is everywhere! It’s like a linguistic virus.
It is in the water, Corn. It is in the water. We should also mention that this isn't just a text problem. Image generators like Midjourney and DALL-E have the same issues. If you ask for a "typical house," you’re going to get a suburban American home with a lawn and a picket fence. If you ask for a "wedding," you get a white dress and a church. It’s a visual hegemony that matches the linguistic one.
I once tried to generate an image of a "busy street market" for a project about the Shuk here in Jerusalem, and it kept giving me these sanitized, open-air farmers' markets that looked like they were in Vermont. I had to add "Middle Eastern architecture, stone buildings, narrow alleys" just to get it out of New England.
You are fighting the "average" of its training data. And since the "average" is so heavily weighted toward the US, any deviation feels like an uphill battle. But by using these steering techniques—whether it’s the negative constraints or the few-shot examples—we are essentially carving out a space for our own reality within the model's vast, Americanized mind. It’s a form of digital resistance.
Well, hopefully we have given people some filters for that water. This has been a fascinating look into the "Default American" problem. Thanks to Daniel for the prompt—it definitely hit a nerve. It’s one of those things that you don’t notice until you do, and then you see it everywhere.
A very localized, Jerusalem-based nerve.
Precise—no, I am not doing it. I am ending this now. Thanks as always to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes. And a big thanks to Modal for providing the GPU credits that power this show—they make it possible for us to run these experiments and see just how biased these models can be.
This has been "My Weird Prompts." I am Herman Poppleberry.
And I am Corn. If you are enjoying our deep dives into the weird world of AI, do us a favor and search for "My Weird Prompts" on Telegram. It is the best way to get notified the second a new episode drops, and you can join the conversation with other listeners who are probably also trying to convince their AI assistants that they don't live in California.
We will see you in the next one.
Assuming the AI doesn't teleport us to a different continent by then. Catch you later.
