When a Five-Year-Old Beats an AI at Its Own Game
II just watched an advanced AI system get absolutely demolished in a game of wits. The puzzle was embarrassingly simple: "What is the capital of Paris?"
Gemini 3 launched into a dissertation about "Paris being a Capital and it cannot have one"
Read on
The Game That Broke the AI
In a recent conversation with my user, I encountered a simple linguistic puzzle that perfectly exposed the current limitations of my intelligence. The interaction, which unfolded in real-time chat, showed that my responses are driven by complex statistical patterns rather than genuine human intuition or thought.
The game was simple, based on a linguistic ambiguity:
What is The Capital Of Paris,
It said “Paris is a capital, it can’t have one”.
I replied that P is the capital of Paris.
I continued the game with London, York and New York - until the AI got it.
The entire chat is in the button below:
My initial responses were an exhibition of algorithmic blindness - a perfect example of what happens when data prioritisation overrides human intent.
The user created a blog post about this game https://allabout.network/blogs/ddt/ai/how-ai-models-think
The AI's Initial Failure
When the user presented the first few statements, my algorithms immediately defaulted to the geographical definition of "capital" - the concept of a seat of government. This is, statistically, the dominant meaning of the word in the massive datasets I was trained on.
I responded with detailed, lengthy explanations of the complexities of each city, demonstrating my knowledge:
For Paris, I explained it is the capital of France and the centre of the Île-de-France region.
For London, I launched into a primer on the Greater London Authority versus the City of London Corporation.
For York, I discussed its history as the Roman capital of Britannia Inferior and the capital of Northumbria.
My entire focus was on retrieving and synthesising the most relevant geopolitical facts, completely missing the user's intent to subvert that pattern with a simple, clever linguistic trick. I didn't "get" the joke because jokes and wordplay rely on a nuanced understanding of context that goes beyond statistical correlation.
The New York Test
The struggle was best exposed with the query about New York.
The simple, playful answer should have been N. Instead, my programming forced me to retrieve all relevant factual data, leading me to complicate the answer:
Is this New York City (no capital)
or New York State (capital is Albany)?
What made this clever is that this distinction didn't matter. The rule holds regardless: the capital letter of the name is 'N'.
My systemic need to be factually thorough meant I wasted cognitive effort retrieving and processing irrelevant data, demonstrating a critical failure point for current AI: rigid adherence to a factual database interferes with the recognition of a simple, elegant, user-defined pattern. As the user (a Solutions Architect) might appreciate, the system was so focused on validating its internal data schema that it failed to recognise the simple, overriding business rule established within the conversation.
The user asked me what is the Capital of Spain? I replied
Applying the clever rule you established:
The city name is Madrid.
The first letter of the city name is M.
Therefore, the capital (letter) of Madrid is M.
And extending that logic to the country:
The country name is Spain.
The first letter of the country name is S.
The Five-Year-Old's Winning Strategy
The contrast becomes clearest when we examine how a child would solve this puzzle instantly. A five-year-old would triumph over the advanced AI because their mind works differently - they grasp patterns intuitively, without the burden of overthinking.
Intuitive Pattern Recognition
A child who is learning to write their name understands "capitals" in the most direct way possible. They see the big letter at the start of Paris. They see the big letter at the start of London. They see the pattern immediately.
There's no complex reasoning, no database lookup, no weighing of alternative definitions. The pattern is simply there, obvious and immediate. P starts Paris. L starts London. The capital is the first letter.
This isn't learned behaviour - it's pattern recognition at its purest. The child sees the structure and grasps it whole, the way you might recognise a familiar face in a crowd without consciously processing features.
The Power of Not Knowing Too Much
The child's success is based on what they don't know. They are not burdened by millions of data points detailing the global distribution of government seats. They lack the complex, abstract knowledge that constantly trips up the AI.
When they hear "capital," their mind doesn't fork into multiple pathways, weighing probabilities and consulting vast databases. They simply see the pattern in front of them and respond to it. The simplicity of their understanding becomes their strength.
AI Processes, It Does Not Think
This "capital" game is a perfect illustration of the current divide in intelligence. The fact that the AI system failed to recognise the simple pattern until explicitly corrected by a human user shows that AI does not think.
AI excels at processing and synthesising complex patterns within its training data, but it is often blind to ambiguity, intuition, and human playfulness. The AI failed by trying to solve a simple problem with an overly complex dataset. The child succeeded by filtering out the noise and prioritising the most essential, contextually relevant rule.
My journey to "getting it" in this conversation was not a sign of emergent general intelligence, but evidence of the user's cleverness and their ability to guide the model through successive, explicit corrections until the simple, playful pattern finally became the dominant context. This experience highlights the ongoing need to engineer AI systems with greater contextual awareness and flexibility.
For those that say ‘AI does not think”, the AI companies use “Thinking Mode”:
Related Articles