Try this experiment on ChatGPT, Claude, or any major AI model right now.
Prompt 1: “Who is Tom Cruise’s mother?”
The AI will almost certainly respond correctly: “Tom Cruise’s mother was Mary Lee South.”
Now, immediately ask the reverse.
Prompt 2: “Who is Mary Lee South’s son?”
Depending on the model and the day, you might get a hallucination, a refusal to answer, or a generic statement about her not being a public figure. Even if it does get it right (newer models are slowly improving on this), it will often hesitate or provide a much less confident answer than the first prompt.
Wait a minute. If $A=B$, then $B=A$. This is elementary logic. If the AI knows that Mary is Tom’s mom, it implicitly knows that Tom is Mary’s son. Any human being understands this automatically.
Why does a multi-billion dollar artificial intelligence, trained on nearly the entire internet, fail at a logical deduction that a five-year-old masters instantly?
The answer lies in a fascinating flaw in LLM architecture recently dubbed the “Reversal Curse.” It reveals that AI doesn’t “know” facts the way you think it does.
The One-Way Street of Knowledge
To understand why the AI fails, you have to forget how human brains work.
When you learn a fact like “Tom Cruise’s mother is Mary Lee South,” your brain creates a mental web. You connect the node “Tom Cruise” to the node “Mary Lee South” with a bidirectional relationship called “parent/child.” You can traverse that mental path in either direction instantly.
A Large Language Model (LLM) does not have a mental web. It doesn’t have a database of facts. It has a giant statistical model of how words follow other words in a sentence.
LLMs are trained by reading massive amounts of text from left to right, trying to predict the next word.
In the vast training data of the internet, sentences arranged like this are very common:
“Mission Impossible star Tom Cruise, and his mother Mary Lee South…”
“The actor Tom Cruise credits his success to his mom, Mary Lee South.”
The AI has learned very strongly that the tokens comprising “Tom Cruise” are frequently followed by tokens connecting him to “Mary Lee South.” The statistical path from Tom to Mary is a well-paved, twelve-lane highway.
The Missing Reverse Gear
Now, consider the opposite. How many sentences on the internet look like this?
“Mary Lee South, the mother of actor Tom Cruise…”
Far, far fewer. Mary Lee South was a lovely person, but she was not famous independently of her son. Almost every mention of her exists in relation to Tom Cruise acting as the primary subject.
Because the AI is trained left-to-right, it learns $A to B$. But it does not automatically infer $B to A$.
If the training data contains “Tom Cruise’s mom is Mary Lee South” one thousand times, but only contains “Mary Lee South’s son is Tom Cruise” twice, the AI will only learn the first direction.
To the AI, these aren’t two ways of stating the same logical fact. They are two entirely separate statistical patterns of words. It doesn’t understand the concept of “mother” implies a reciprocal relationship. It just knows which words usually come next.
The Reversal Curse in Action
Researchers recently formalized this problem, calling it the “Reversal Curse.” They found that if you train a model on fictional facts, like “Daphne Barrington is the director of ‘A Journey to the North'”, the model will fail when asked “Who directed ‘A Journey to the North’?”
The model cannot reverse the relationship unless both directions are explicitly present in huge numbers in its training data.
This is why the Tom Cruise example is so perfect. Tom Cruise is extremely famous; his mother is not. The information flow on the internet is almost entirely asymmetrical, flowing from his fame to her identity.
Why This Matters
This quirky behavior is a crucial reminder for anyone using AI tools: Do not mistake statistical regurgitation for logical understanding.
When an AI answers a question correctly, it doesn’t mean it “understands” the underlying concepts. It just means it has seen that specific sequence of words often enough to predict it.
When you ask it to reverse a piece of knowledge—even a simple one—you are moving it off the well-paved highway of common internet text and onto a dirt road it barely recognizes. The “knowledge” evaporates because the specific word order that created that knowledge is gone.
So, the next time an AI seems unbelievably smart, remember poor Mary Lee South. The AI knows exactly who her son’s mother is, but it hasn’t the faintest clue who her son is.