The Metacognitive Bottleneck: Japanese Riddles Reveal Fundamental Limits of Machine Insight and Self-Evaluation in Reasoning AI

Imagine you are trying to teach a robot how to solve a mystery. You might think the hardest part is just knowing the facts, like memorizing a dictionary or a history book. But there is a trickier kind of thinking called "reasoning," where you have to look at a problem, realize your first guess is wrong, and suddenly see the answer in a completely new way. It's like looking at a pile of puzzle pieces, thinking they don't fit, and then suddenly realizing you've been holding one upside down. Scientists call this "insight."

Now, imagine you want to test if a robot can do this. You can't just give it a math test, because robots are great at math but might just be memorizing the answers from their training. You need a test that forces them to think creatively and check their own work. This is where the idea of "metacognition" comes in. Think of metacognition as the robot's internal supervisor. It's the part of the brain that says, "Wait, I just found the right answer, but I'm not sure if I should stop here or keep looking." If the robot can't trust its own supervisor, it might find the solution but then talk itself out of it, giving the wrong answer anyway.

This is exactly what a team of researchers from Japan wanted to investigate. They wondered: Are our smartest AI models actually getting better at this kind of creative thinking, or are they just getting better at guessing the right answer because they've seen similar questions before? To find out, they created a special test using Japanese riddles and discovered something surprising about how these machines think.

---

The Riddle Test: Can AI Solve the "Aha!" Moment?

The researchers, led by Masaharu Mizumoto and his team, decided to use a classic type of Japanese children's riddle called nazonazo. These aren't your average "What has four legs and barks?" riddles. These are tricky word puzzles that often rely on puns, breaking down characters, or seeing numbers in a new way. They are designed to make your brain hit a wall, or an "impasse," before you suddenly have an "Aha!" moment where the answer clicks into place.

The team built a new benchmark called the NazoNazo Benchmark. They gathered 201 of these riddles. To make sure the test was fair and fresh, they picked 120 of them to compare against real humans. They asked 126 people to solve these riddles, and the humans got about 52.9% of them right. This gave the scientists a "human bar" to measure the robots against.

Then, they put 38 different large language models (the fancy AI brains) released between 2023 and 2025 to the test. They made sure the AIs couldn't look up answers on the internet or use any outside help; they had to solve it using only what was inside their own "heads."

The Results: Robots Are Still Stuck in the Mud

The results were a bit of a shock. Even the most advanced "reasoning" models, which are supposed to be the smartest, only got about 17.6% of the riddles right. The simpler models did even worse, at just 7.6%.

Compare that to the humans, who were more than three times as accurate. While most models struggled significantly, one model stood out as a notable exception: GPT-5. It was the highest-performing and most recently released model in the evaluation, and its accuracy fell within the range of human performance. However, the researchers noted that this overlap doesn't prove the AI is statistically equivalent to humans, but it does show it was the only model to approach the human range. For almost everyone else, the AI was struggling to make that creative leap.

The researchers also noticed something interesting about the riddles themselves. For humans, the difficulty of the riddles was spread out evenly, like a smooth hill (a Gaussian distribution). But for the AI, the scores were heavily skewed toward very low values. The AI mostly got the easy ones wrong or the hard ones wrong, with very few models performing anywhere near the middle or high end. This suggests the AI isn't really "thinking" its way through; it's just guessing.

The Real Problem: The Robot That Talks Itself Out of the Answer

Here is the most fascinating part of the story. The researchers didn't just look at the final answers; they peeked inside the AI's "thought logs." These are the step-by-step notes the AI writes down while it's thinking.

They found a specific failure mode they call "verification failure."

Imagine you are looking for your keys. You look under the sofa, and there they are! You pick them up. But then, your inner voice says, "Wait, maybe they aren't really there. Let me check the kitchen just in case." You go to the kitchen, look around, don't find them, and then come back and say, "I guess I don't have my keys."

That is exactly what the AI was doing. In many cases, the AI would generate the correct answer in its thought process. It would write down the right solution. But then, instead of saying, "Okay, that's the answer," it would keep searching, doubt itself, and eventually pick a wrong answer.

This happened in a surprising number of cases:

• For one model, this happened in 39.34% of the times it got the answer wrong.
• For another, it was 30.57%.
• Even the best ones had this issue, with rates ranging from 5.23% to 39.34%.

The AI was like a detective who finds the criminal but then decides to arrest the wrong person because it didn't trust its own evidence. The researchers call this a "metacognitive bottleneck." The robot can find the solution, but it can't trust it enough to stop and commit to it.

What This Means for the Future

The paper suggests that the problem isn't just about making AI smarter or giving it more data. The problem is that these models are bad at checking their own work. They are great at generating ideas but terrible at knowing when to stop and say, "Yes, this is it."

The researchers also ruled out a few ideas. They showed that the AI wasn't just memorizing the riddles from the internet, because if it had, it would have gotten much higher scores. They also found that switching the language the AI thought in (Japanese vs. English) didn't really help much, proving the problem is deeper than just language skills.

So, what's the takeaway? We have built machines that can sometimes solve a puzzle, but they lack the confidence to admit when they've solved it. To make AI truly smart, we might need to teach them not just how to think, but how to trust their own thoughts. Until then, when it comes to tricky riddles, a curious human teenager is still likely to beat the robot.