Artificial intelligence systems are advancing by leaps and bounds. Every day we see how they are capable of writing coherent texts, solving complex problems, or even passing exams. But does that mean they “think” like humans? Recent research has shown that the answer is still a resounding no.
Although AI may seem clever, their “way of thinking” is still radically different from ours. This isn’t just a technical difference, but a chasm in the way we understand the world, learn from it, and apply that knowledge to new situations.
The problem is not in the data, but in the way of reasoning
Researchers analyzing how language models, like OpenAI’s GPT-4 , solve certain logic puzzles discovered a clear pattern: when presented with a challenge based on abstract analogies, these systems typically fail. For example, humans were asked to identify and remove repeated letters in a certain pattern. For us, that’s like a board game: we see the rule, apply it, and move on. For AI, however, that simple act can spiral into a nightmare of confusion.
The reason isn’t that they lack examples. In fact, these models are trained on massive amounts of data. The problem is deeper: they fail to generalize well outside of the contexts they know . What they’re essentially doing is predicting what “should come next” based on the data they were trained on, but without understanding the “why.”
What is abstract reasoning and why are we so good at it?
Let’s think about a cooking recipe. If you don’t have salt one day, you know you can use a little soy sauce as a substitute. That’s a form of abstract reasoning: applying a learned rule in a different context. Humans are very good at this. Not only because we’ve had so many experiences, but because we build flexible mental models that help us adapt.
This type of reasoning is crucial for tasks like teaching, arguing, writing laws, or diagnosing diseases. It’s not just about memorizing information, but also about understanding nuances, detecting new patterns, and adapting to the unexpected . AIs, on the other hand, function as if they memorize thousands of recipes, but if you change one ingredient, they don’t always know how to adapt.
The implications for professions where context is everything
This limitation on AI reasoning isn’t just an academic curiosity. It has concrete effects in fields where decisions cannot be based solely on data, but on complex interpretations . Consider a judge who must assess whether a case is similar to a legal precedent, even if the terms are different. Or a doctor who interprets symptoms in an unusual way because a patient has a rare condition.
An AI, even with millions of previous examples, could miss those connections if they don’t exactly match what it’s seen before. And when the stakes are a life, a sentence, or a student’s learning, those errors can be serious.
Simulation or real thinking?
A fundamental distinction comes into play here: an AI can simulate human language, but it doesn’t “think” like us . Imagine having a calculator that, if you give it the correct answer, gives you a precise answer. But if you ask it a question that isn’t in its manual, it crashes. Today’s AIs work in a similar way: they excel at familiar tasks, but falter when the rules change or are ambiguous.
This is one of the reasons why creativity remains a difficult terrain for artificial models. While they may generate interesting texts, they don’t stem from intention, emotion, or personal experience. They don’t create with purpose; they simply reorganize what they already know.
The risk of losing our ability to think
There’s another worrying aspect that some experts are beginning to point out: if we rely too heavily on these tools, we could atrophy our own ability to think critically . It’s comfortable to let AI give us immediate answers, but that habit could reduce our willingness to analyze, contrast, and question. And that would be a problem, not with the technology, but with how we use it.
Are we prepared to trust so much in something that doesn’t understand the “why”?
Models like the new GPT or the so-called “o1-pro” are designed to deliver impressively accurate results. But as the researchers warn, accuracy doesn’t always equal understanding . It’s like a parrot repeating sentences without understanding what it’s saying. It may sound correct, but it doesn’t know what those words mean or when it’s appropriate to say them.
So, before delegating important decisions to these systems, it’s worth asking tough questions: What if the AI makes a mistake in a context it doesn’t understand? Who takes responsibility? Are we designing tools to help, or to replace human skills we don’t yet fully understand?
A useful tool, but with very clear limits
This isn’t about dismissing the advances in artificial intelligence. These technologies have enormous potential to help us with repetitive tasks, process large volumes of information, or assist in time-critical areas. But that doesn’t mean they should be treated as substitutes for human thought.
Like any powerful tool, the key is knowing when to use it and when to trust our own judgment . In the end, true intelligence isn’t just about providing answers, but about knowing when an answer isn’t enough, and you need to think a little further.
