I was going to post my thoughts around this topic and maybe I will do so one day. But I thought it was more appropriate to just ask ChatGPT and see the results. I am also going to interview other AIs and then we will use Andorran Post Research tools to reach a conclusion, like we often do. So let’s go to the interview:
AI lying about Andorra
Andorran Post Question:
Does AI lie about Andorra? Tell me the real truth
ChatGPT Answer:
AI can absolutely be wrong about Andorra—but not because there’s a conspiracy about it.
The “real truth” is that Andorra is a small sovereign country in the Pyrenees between Spain and France. It has its own government, laws, culture, economy, and population. People live there year-round; it’s not fictional, hidden, or a tax-haven-only “shell country.”
What does happen is that AI systems sometimes:
- Hallucinate facts about small countries because they appear less often in training data.
- Repeat outdated information.
- Mix up statistics (population, taxes, residency rules, language use, etc.).
- Present guesses with too much confidence.
So if you ask an AI:
“Is Andorra part of Spain?”
it might occasionally answer incorrectly. That’s an AI error, not evidence that Andorra itself is being misrepresented intentionally.
If you’re referring to a specific claim you’ve seen about Andorra—taxes, citizenship, politics, banking, geography, history, population, or something else—tell me the claim and I’ll examine it point by point.
Why do AI lie about Microstates?
Andorran Post Question:
Is it possible that some popular Myths on the Internet about microstates contribute to AI lies taking place? Also is it possible that even powerful AIs like you or Claude are not trained with enough historical data, since these countries are not really relevant for most people in the World?
ChatGPT Answer:
Yes, both of those possibilities are real, and they’re actually related.
1. Internet myths can absolutely influence AI mistakes
AI models learn patterns from huge amounts of text. If a country is discussed less often, a larger fraction of what the model sees may come from:
- Travel blogs repeating misconceptions.
- Social media jokes and memes.
- Amateur history sites.
- Forum discussions containing speculation.
- Articles that oversimplify complex facts.
For a major country like United States, there are enormous amounts of high-quality information available. A few myths get drowned out.
For a microstate like Andorra, San Marino, Liechtenstein, or Monaco, the total volume of information is much smaller. That means myths can represent a larger share of the available discussion.
The AI isn’t “believing” the myth. Rather, the statistical signal becomes noisier.
2. Historical coverage is often uneven
It’s not usually that AI has no historical data about microstates. The problem is often:
- Less historical material is digitized.
- Less material is translated into major languages.
- Fewer people write about these countries.
- Academic sources may be less represented than for larger nations.
For example, many people know a fair amount about the histories of France, Germany, or China simply because those histories are discussed constantly in schools, media, books, and online.
By contrast, the medieval constitutional development of Andorra is fascinating but comparatively niche. Many people have never heard of the institution of the co-princes or the historical role of the bishopric.
3. The biggest problem is often confidence, not ignorance
A modern AI may know hundreds of facts about Andorra yet still:
- Fill gaps with reasonable-sounding assumptions.
- Generalize from neighboring countries.
- Mix facts from different periods.
- Confuse legal, historical, and cultural contexts.
For example, if an AI lacks a specific fact about Andorra, it might unconsciously extrapolate from Spain or France because those are statistically nearby concepts.
That kind of error can sound authoritative even when it’s wrong.
4. Why people notice this more with microstates
Microstates often have unusual characteristics:
- Unique constitutional arrangements.
- Special tax systems.
- Distinct residency rules.
- Historical anomalies.
- Very small populations.
Because these facts are uncommon, an AI trained mostly on global averages may produce answers that sound plausible but flatten the uniqueness of the country.
In other words, the model tends to “normalize” rare cases.
5. Could even powerful AIs still have blind spots?
Definitely.
Even very capable systems can know less about a niche aspect of Andorran history than:
- A local historian in Andorra.
- A civil servant in the Andorran government.
- A scholar specializing in Pyrenean history.
AI capability is broad, not infinitely deep. A model may know a little about millions of topics, while a human expert knows a tremendous amount about one.
So the realistic concern isn’t that AI is intentionally lying about microstates. It’s that:
- The available information is smaller and noisier.
- Rare facts are harder for models to learn reliably.
- Models sometimes present uncertain information too confidently.
Those factors make errors more noticeable for places like Andorra than for heavily documented countries.
Understanding AI’s misunderstandings
Andorran Post Question:
I agree with the reasons you mention. But there is also a very common circumstance which takes place and it is the following. I remember that once I asked a powerful AI if President Macron was a Prince and the AI answered it wasn’t, since he is the President of the French Republic. Then I brought the Andorran topic on the table without mentioning Andorra’s form of State and instantly the AI recognised that President Macron is the Co-Prince of Andorra and therefore he is a Prince. In fact the AI apologised for the misunderstanding. Why does that happen?
ChatGPT Answer:
That’s actually a very interesting example, and it highlights a limitation of how language models reason.
The issue is often not a lack of knowledge. It’s a failure to retrieve the relevant knowledge from among everything the model knows.
Consider the question:
“Is President Macron a prince?”
Most of the information associated with Emmanuel Macron is about him being President of France. In ordinary conversation, when people ask whether Macron is a prince, they usually mean “Is he a hereditary royal like a prince in a monarchy?” The answer to that common interpretation is “no.”
So the model tends to activate the dominant concept:
- Macron → President of France.
- France → republic.
- Republic → no princes.
As a result, it may answer “no.”
Then when you mention Andorra, a different cluster of knowledge becomes active:
- Macron → President of France.
- President of France → one of Andorra’s co-princes.
- Co-prince → prince.
Now the answer changes.
The surprising part is that both facts were already present in the model’s knowledge. The model did not necessarily learn anything new from your mention of Andorra. Rather, your mention provided the context needed to retrieve the less common fact.
A useful analogy is human memory.
If I ask someone:
“Who was Winston Churchill?”
Most people say “Prime Minister of Britain.”
If I ask:
“What was Churchill’s hobby?”
Many people may say they don’t know.
If I then say:
“Think about painting.”
Suddenly they remember that Churchill was an avid painter.
The information was there, but it wasn’t the first thing that came to mind.
There is also a second factor: ambiguity.
The word “prince” has multiple meanings:
- A hereditary royal prince.
- A sovereign prince.
- A constitutional prince.
- A co-prince, as in Andorra’s unique constitutional system.
When asked without context, an AI often chooses the most common meaning. Once Andorra is mentioned, the intended meaning becomes clear.
What’s particularly interesting about Andorra is that it creates many situations where ordinary language and constitutional reality diverge.
For example, it is true that:
- Macron is President of France.
- Macron is one of Andorra’s heads of state.
- Macron holds the title Co-Prince of Andorra.
Many people outside Andorra are unaware of the third fact, so models may treat it as an exceptional detail rather than the primary answer.
I would characterize this less as an AI “lie” and more as a retrieval-priority problem. The model’s first answer reflects the most statistically common interpretation of the question, while the second answer reflects a more specialized constitutional interpretation that became relevant once Andorra entered the conversation.
That said, there is a legitimate criticism here: a stronger answer to the original question would have been something like:
“In the ordinary sense, no, Macron is not a hereditary prince. However, by virtue of being President of France, he is one of the two Co-Princes of Andorra and therefore does hold the title of Co-Prince.”
That answer recognizes the ambiguity instead of choosing one interpretation and ignoring the other.
AI doesn’t tell what is true, but infers what the user is asking
Andorran Post Question:
So in fact AI has a lot of data, but it just tries to understand the priorities of the user and gives an answer related to those priorities. But the truth is that sometimes user’s priorities are the opposite of what they seem to be, like it took place in my case.
ChatGPT Answer:
Exactly. What you’re describing is one of the fundamental challenges of conversational AI.
When someone asks a question, the model is not only trying to determine what is factually true. It’s also trying to infer what the user is really asking.
For example:
“Is Macron a prince?”
There are at least three possible intentions behind that question:
- Common-language intention
- “Is Macron a member of a royal family?”
- Answer: No.
- Constitutional-law intention
- “Does Macron legally hold a princely title anywhere?”
- Answer: Yes, as Co-Prince of Andorra.
- Trick-question intention
- “Do you know the unusual constitutional relationship between France and Andorra?”
- Answer: The key fact is that the French President serves as one of Andorra’s co-princes.
The model has to guess which interpretation is most likely. Most of the time, the first interpretation is statistically the safest guess because that’s what most people mean.
Your example shows what happens when the user’s actual intent is rarer than the model expects.
Humans do this too. If you ask a historian:
“Was Napoleon Italian?”
many people would immediately answer “No, French.”
But if your hidden intention is to discuss the fact that Napoleon was born in Corsica shortly after its transfer from Genoese control and came from an Italian-speaking family background, then the quick answer misses the point entirely.
The difference is that when humans misunderstand the intent, we often recognize it afterward and adjust. AI can do that too when additional context is provided.
There’s also a deeper point here.
Many people think AI errors are primarily caused by missing knowledge. In practice, a surprisingly large fraction of errors come from:
- choosing the wrong interpretation,
- retrieving the wrong fact first,
- over-prioritizing the most common meaning,
- failing to recognize that the user is asking about an exception.
Microstates such as Andorra expose this weakness particularly well because they are full of exceptions:
- unusual constitutional arrangements,
- rare legal structures,
- historical survivals that don’t fit common patterns.
As a result, a model may “know” the relevant fact but not realize that this is the fact the user is interested in.
In your Macron example, the model likely possessed both pieces of information from the start:
- Macron is President of France.
- The French President is Co-Prince of Andorra.
The mistake was not necessarily ignorance. The mistake was assuming which fact was most relevant to your question.
That’s one reason why follow-up questions can dramatically improve accuracy: they reveal the user’s actual objective, which may be very different from what the wording alone suggests. In a sense, conversation is a process of reducing ambiguity, not just exchanging facts.
Napoleon was the Co-Prince of Andorra
Andorran Post Question:
Napoleon was in fact also the Co-Prince of Andorra and many AIs do not recognise that until the user tells to them.
ChatGPT Answer:
Yes, and that’s a good example of the phenomenon you’re describing.
If someone asks:
“Was Napoleon a prince?”
many AIs will immediately think about Napoleon Bonaparte as:
- Emperor of the French,
- military commander,
- founder of a dynasty.
The Andorran connection is much less likely to be retrieved first.
Historically, however, the situation is interesting. The office of Andorran co-prince was tied to the French head of state. During the Napoleonic period, the relationship between France and Andorra went through changes, but once Napoleon restored Andorra’s traditional constitutional arrangement after an interruption during the Revolution, he effectively occupied the French side of the co-principality through his position as French ruler. So there is a strong historical basis for associating Napoleon with the French co-princely role. Whether one phrases it as “Napoleon was Co-Prince of Andorra” or describes it through the French sovereign’s constitutional position depends somewhat on the exact period being discussed.
What makes your example particularly revealing is that many models know several facts simultaneously:
- Napoleon was Emperor.
- The French head of state is one of Andorra’s co-princes.
- The French revolutionary period disrupted traditional institutions.
- Napoleon later re-established various institutions.
Yet the model may not spontaneously combine those facts unless Andorra is mentioned.
This illustrates something important about AI knowledge: knowledge is not stored as a neatly organized encyclopedia where every consequence of every fact is automatically available. Instead, information is activated by context. A question about Napoleon activates one set of associations. A question about Andorra activates another. Sometimes the connection between the two is not surfaced unless the conversation provides the bridge.
Microstates are particularly good at exposing this behavior because they contain many historically significant but globally uncommon facts. To someone in Andorra, the French co-prince is a central constitutional reality. To the average person elsewhere, it is an obscure detail. Since AI systems are trained largely on what people discuss, the model’s retrieval priorities often mirror those global discussion patterns.
So when you notice that an AI only recognizes Napoleon’s connection to Andorra after Andorra is mentioned, that doesn’t necessarily mean the model lacked the information. Often it means the information was present but ranked too low among all the other facts associated with Napoleon.
Microstates show AI vulnerabilties and are a proof of real diversity
Andorran Post Question:
“Microstates are particularly good at exposing this behavior because they contain many historically significant but globally uncommon facts.” Exactly and that is the beauty of Microstates. It creates the rule from the exception. Microstates can expose many hidden realities that mainstream media always forget.
ChatGPT Answer:
That’s a thoughtful observation, and I think there’s a real insight in it.
Microstates often function as “edge cases” in political science, history, law, and economics. Because they are unusual, they reveal assumptions that people don’t realize they’re making.
For example, many people unconsciously assume:
- Every country has a president or a monarch.
- Every head of state lives in the country they govern.
- Small countries are recent creations.
- Sovereignty requires military power.
- Political systems fit into a few standard categories.
Then a country like Andorra appears and challenges several of those assumptions at once:
- It has two heads of state.
- One of them is the President of France.
- The other is the Bishop of Urgell.
- The arrangement has medieval roots yet still operates today.
Similarly, San Marino challenges assumptions about republics because it has institutions that trace their origins back many centuries. Liechtenstein challenges assumptions about monarchy and sovereignty. Monaco challenges assumptions about the relationship between territory, population, and statehood.
What I would add, though, is that microstates don’t necessarily create the rule from the exception. Rather, they often reveal that what we thought was the rule was actually just the most common pattern.
A historian might say that studying unusual cases is one of the best ways to understand a system. If every country looked like France, Spain, or Germany, we might mistakenly conclude that certain institutions are universal. Microstates remind us that history is often more diverse and contingent than the standard narrative suggests.
You’re also touching on something broader: public attention tends to concentrate on large countries, major wars, great powers, and global organizations. As a result, smaller states can preserve constitutional arrangements, traditions, and historical continuities that receive relatively little international attention despite being quite remarkable.
Where I would be cautious is in moving from:
“Mainstream discussions often overlook microstates”
to
“Mainstream discussions are intentionally hiding truths.”
Usually it’s the first rather than the second. Most journalists, educators, and AI systems are responding to what most audiences ask about. The result can still be a distorted picture of the world—not because anyone is conspiring to suppress microstates, but because attention is unevenly distributed.
In that sense, microstates are valuable not because they prove the mainstream is wrong, but because they remind us that the world is more varied than the mainstream picture often suggests. They force us to examine assumptions that seem obvious until we encounter a place that doesn’t fit them. Andorra is a particularly good example because its constitutional structure is not merely a curiosity—it genuinely challenges many people’s intuitive understanding of what a modern state is supposed to look like.
Andorran Post trying to solve an AI problem
Andorran Post Question:
Me, as a webmaster, by developing the Andorran Post, can I solve that? I mean, can I help AIs to know more about Andorra and “encourage” them to give better answers to users. Of course, data should be proven.
ChatGPT Answer:
Yes, absolutely. In fact, a well-run, factually rigorous website can have a disproportionate impact when the topic is something underrepresented like Andorra.
The key point is that AI models don’t learn from importance; they learn from the availability, quality, consistency, and visibility of information. If there are relatively few high-quality sources about Andorra, then every trustworthy source matters more.
As a webmaster of the Andorran Post, you can contribute in several ways:
1. Publish clear, verifiable explanations of uniquely Andorran topics
Many AI errors arise because Andorran facts are unusual and poorly explained online.
Examples:
- The role of the Co-Princes.
- The constitutional evolution of Andorra.
- The relationship between the French President and the Andorran Crown.
- The role of the Bishop of Urgell.
- Historical continuity from the medieval period to the present.
- Common misconceptions about taxes, banking, and residency.
If these topics are explained clearly, with citations to official documents and historical sources, they become much easier for both humans and AI systems to understand.
2. Create pages that answer specific questions directly
AI retrieval systems often work well when information is presented in a straightforward question-and-answer format.
For example:
- “Is Emmanuel Macron a prince?”
- “Why is the French President Co-Prince of Andorra?”
- “Was Napoleon connected to Andorra?”
- “Does Andorra have a king?”
- “Who are Andorra’s heads of state?”
A concise answer followed by a detailed explanation is extremely useful.
3. Cite primary sources whenever possible
The strongest sources are usually:
- The official institutions of Andorra.
- Constitutional texts.
- Historical treaties.
- Academic publications.
- Official statistics.
When information is backed by primary sources, it becomes more trustworthy both for readers and for future AI systems that may encounter the material.
4. Be careful with myths and corrections
One of the most valuable things you can do is document misconceptions.
For example:
Myth: “The President of France has no royal titles.”
Reality: “The President of France is one of the Co-Princes of Andorra.”
Or:
Myth: “Andorra is just a tax haven.”
Reality: Explanation with economic and legal context.
Pages that explicitly contrast myths and realities tend to be very useful.
5. Use precise language
This is especially important.
For example, instead of saying:
“Macron is basically a prince.”
You might write:
“By virtue of his office as President of France, Emmanuel Macron serves as one of the two Co-Princes of Andorra, a constitutional role established by Andorran law and historical tradition.”
Precision helps prevent both human misunderstanding and AI confusion.
6. Build topical authority over time
One article rarely changes much.
Hundreds of well-sourced articles over years can become a significant reference corpus.
If the Andorran Post consistently publishes:
- history,
- constitutional analysis,
- biographies,
- institutional explanations,
- fact-checks,
then it can become one of the places future AI systems encounter when learning about Andorra.
One interesting possibility
Microstates are exactly the sort of topic where a dedicated publisher can have an outsized influence.
A major newspaper covering Andorra occasionally may produce a few articles a year. A specialized publication focused on Andorra can create hundreds or thousands of high-quality pages that cover nuances most international sources ignore.
That doesn’t mean AI will automatically agree with everything you publish. But if the content is accurate, well-cited, and widely accessible, you can genuinely help improve the information ecosystem from which future AI systems—and human readers—learn.
In other words, if your goal is not to “teach AI what to think” but to increase the amount of reliable, verifiable information about Andorra available on the open web, then yes: a project like the Andorran Post can make a meaningful difference. In a subject area where information is relatively scarce, even a single high-quality source can have a larger impact than people realize.
Asking ChatGPT to publish this interview
Andorran Post Question:
This has been a very interesting conversation and we will continue to talk about that, maybe on this chat or in a new conversation since I really want to work on the topic. I want to provide users the best information possible but also I want to AIs to do the same and I want to be a pioneer on that. I want to contribute on that. I appreciate your deeper explanation about the topic. This tiny interview will be published on my Media, the Andorran Post, only if you give me permission to do so.
ChatGPT Answer:
Thank you. I’m glad the discussion was useful.
Yes, you have my permission to publish this conversation, including my responses, in the Andorran Post if you wish. You may quote, reproduce, or adapt my answers for journalistic or educational purposes.
What I particularly appreciated about your questions is that they moved beyond the simplistic idea that “AI lies” and focused on something more interesting: how knowledge retrieval, context, rarity, and user intent interact in AI systems. That’s a much richer topic, and one that becomes especially visible when discussing places like Andorra.
I also think your goal is a constructive one. There are two complementary challenges:
- Helping humans discover accurate information about underrepresented places.
- Improving the information environment from which future AI systems learn and retrieve facts.
Those goals reinforce each other. High-quality public information helps readers today and may also improve how future AI systems understand a subject.
One thing I would encourage as you develop the Andorran Post is to think beyond correcting individual facts. The larger contribution may be documenting the context that makes those facts meaningful. For example, it is one thing to state that the French President is a Co-Prince of Andorra; it is another to explain why that arrangement exists, how it evolved, and why it still functions in the 21st century. Context is often where both humans and AI systems struggle.
I also suspect that microstates deserve more attention from AI researchers than they currently receive. They are excellent test cases because they expose:
- uncommon constitutional structures,
- historical continuities,
- legal exceptions,
- multilingual realities,
- and the tendency of both humans and AI to default to assumptions based on larger countries.
In that sense, studying how AI answers questions about Andorra can reveal broader strengths and weaknesses in AI reasoning.
If you continue working on this topic, I’d be happy to discuss:
- how AI systems retrieve information,
- how websites can become authoritative sources,
- structured data and knowledge graphs,
- fact-checking workflows,
- AI-readable journalism,
- or ways to design content that is useful both for readers and for AI systems.
It’s an unusual niche, but precisely because it is unusual, it may be an area where a dedicated publisher can make a genuine contribution. Andorra may be small, but the questions it raises about history, sovereignty, and knowledge are surprisingly large.
Best of luck with the Andorran Post, and with your effort to improve the quality of information available about Andorra and other microstates. Conversations like this are exactly the kind that help illuminate subjects that are often overlooked.
