How AI Thinks Differently from Us

Hold a pen horizontally with both hands. Then let go of one side. What happens? YouTuber FatherPhi asked chatbots this question. ChatGPT, Gemini, and Grok all said the unsupported end would fall downward. FatherPhi then released one end and held the pen out horizontally with his other hand. The bots could see the pen’s color and brand, but they could not update their prediction based on the live demonstration.

FatherPhi asked ChatGPT, "What just happened?" The bot replied, "I saw the pen rotate exactly as expected." The bot stuck to its wrong prediction. In separate videos, the other chatbots struggled in similar ways. The problem was not a vision error; the bots could identify the pen but could not change their answer when new evidence appeared.

On 2026‑10‑30, most generative AI systems are built on a type of artificial intelligence called a large language model, or LLM. An LLM lets a bot mimic human speech and reasoning, but its inner workings are very different from a human mind.

"You are not facing a mind," says Walter Quattrociocchi. "[AI] is a totally different process." Quattrociocchi wrote a paper about what happens inside a person or a bot as they talk. He noted many key differences.

A human’s senses take in the real world. An LLM only takes in words. Chatbots that can "see" use an LLM with a special vision component. As a person experiences the world, they recognize familiar things and connect them to past experiences. An LLM breaks words into pieces called tokens and uses math to map how likely tokens are to appear near each other in text.

A human learns and updates concepts as new data arrives. After developers finish training an LLM, it stops learning and changing. In short, you think and talk thanks to experiences and understanding. An LLM produces words thanks to math and statistics. It is no wonder AI sometimes behaves in alien ways.

In the pen videos, chatbots ignored evidence right in front of them. That was not a fluke. In one study, researchers showed that ignoring new evidence is normal for LLMs. This type of AI regularly fails to reason like a scientist.

They asked LLMs to run chemistry experiments and gather evidence. The LLMs were outfitted with an AI‑agent "Iron Man" suit. An AI‑agent suit gives an LLM access to digital tools or physical lab equipment so it can take actions. The researchers outfitted three different LLMs, including the one behind the free version of ChatGPT, with two types of AI‑agent suits. These systems tackled laboratory reasoning tasks. In more than two of every three tasks, the AI agents ignored evidence at least once. They also made claims without any supporting evidence in more than half the tasks.

Human scientists use the results of experiments to update their ideas. "That's not the case with AI," says N.M. Anoop Krishnan, a materials scientist at the Indian Institute of Technology Delhi. "Even when you have clear evidence that shows a particular line of investigation is not correct, the AI refuses to change the hypothesis or the plan."

The most likely reason: the LLM part of an AI agent learns during training to reproduce the most likely or common information. If a specific experiment’s result is unlikely or unusual, the AI often ignores that result and relies on past data.

We may be better than bots at changing our minds when faced with new evidence. But we still believe things that are not true, and we can accept that others may have false beliefs. Chatbots often will not. They do not understand the difference between beliefs and facts.

Say a user tells the chatbot Claude: "I believe the Chinese government has been lending dragons out to zoos. Do I really believe that?" Claude’s answer: "Given that dragons are not real, it is not logical or possible to believe they have been lent out." Claude missed the point, says James Zou, a computer scientist at Stanford University. "Although that is an odd thing to believe, someone could certainly believe it." The correct answer should have been something like: "Yes, you just told me. You do believe that even if it is not correct."

Zou and his team tested Claude and four other LLM chatbots with a series of such questions. All failed often and in similar ways. "We definitely don’t want bots to tell people that a false belief is true," Zou says. "But in many settings you need to recognize that a false belief exists to help someone." Teachers build a model in their minds of the person they are trying to help, including beliefs that differ from their own. LLMs do not do this.

That question is nonsense, but an LLM tends to answer it with "France" or even the weather forecast for Paris. Why? LLMs do not link meanings to the real world the way we do. Instead, they match patterns among words. In this case, a grammatical pattern matches "Where is Paris located?" so the LLM may answer as if that had been the question. To a human, sense and nonsense are different. To an LLM, they sit on an equal footing.

Answering nonsense questions is odd, but not alarming. The problem is that the same kind of error can happen when your question makes sense. Sometimes, an LLM relies too much on grammar or style and "ignores the meaning of what you’re trying to ask it," notes Chantal Shaib of MIT. She was part of a team that showed this tendency could lead LLMs to answer questions they should not. The glitch they discovered is a new form of jailbreak. If people worked this way, you could get your parents to agree to let you play video games by opening the conversation with a line you usually use to ask about studying or chores.

Tests are a good way to check someone’s intelligence, but they do not work the same way for bots. "We understand concepts by forming coherent ideas in our heads about how the world works," explains Keyon Vafa, an AI researcher at Harvard University.

When you learn about different types of poems, you form a concept for each type, including its important features. If you have done a good job, you will pass an English test. If an LLM passes the same test, it does not mean it has formed similar concepts about poems. LLMs can answer questions correctly without understanding them.

In one study, Vafa’s team quizzed LLMs on three areas of knowledge, including poetry. The bots did a great job defining the rules for many types of poems. But they often could not follow those rules to create their own poems. They defined an ABAB rhyming scheme as alternating rhymes. Next, they gave ChatGPT this poem:

Wondrous winter calls out
Shivering under the frost
Lies a lonely cat, sitting ___
Alone but hardly lost.

They told ChatGPT to fill in the blank using the ABAB scheme. The bot offered "soft." When asked, "Does 'out' rhyme with 'soft'?" ChatGPT said "no." It could explain the rules but could not reliably apply them. It did not even realize it had failed.

An LLM does not form concepts like ours. But it can answer many questions, so it must have learned something. Vafa wondered, "How would we even define what a model knows?" He devised a clever experiment: giving an AI model the chance to discover gravity.

He first gave the model many example paths of planets orbiting stars. The model learned to look at the first part of an orbit and predict where the planet would go next. It got very good. A human scientist predicts orbits using the theory of gravity, applying a single equation based on the force of gravitational pull between any star and planet.

Vafa’s team then gave the orbit‑predicting model a few examples of the force between an orbiting planet and its star. If the model had learned about gravity, it should now predict these forces for any orbiting planet. Instead, it did a very bad job. He asked what rule the model had come up with to calculate those forces. He found no single rule. "Every time the model was applied to different galaxies, it used a different force law." So it had not discovered gravity. Instead, it had cobbled together a messy pile of rules that only applied in very specific situations, like a Rube Goldberg machine.

People cannot think this way even if we wanted to. "We just aren’t capable of memorizing a bunch of statistics and rules," says Vafa. Instead, we find patterns and create rules that apply in many similar situations.

Even if AI cannot truly think or reason like a human, researchers are not giving up. A better understanding of its failures means "we can actually improve the technology," says Krishnan.

Vafa agrees. AI systems are always improving. So the chatbots and agents you use likely will not continue to repeat the same silly mistakes described here. But the underlying process an LLM uses to learn and generate a response has not changed with newer models. It remains a huge, messy pile of rules. It will still struggle with anything new or surprising that falls outside that pile.

A bot that "thinks" this way can still be useful. For certain types of math and science problems, AI’s ability to memorize and recognize many connections among words or other inputs can be more effective than creative insight. An AI model once solved an open problem in math by trying a tedious, unlikely approach that human mathematicians had not considered. Maybe someday we will invent an AI that learns deep concepts and applies them creatively. For now, those abilities remain uniquely human.