Large language models have learned not only to write texts but also to solve complex mathematical problems. However, that does not yet mean they truly understand what they are doing. The next major challenge for AI is to learn logical reasoning and an understanding of cause‑and‑effect relationships, to transfer knowledge to entirely new situations, and to grasp the consequences of its own decisions. "Did You Know?" discussed this with Perouz Taslakian, a machine learning researcher at ServiceNow Research in Montreal and lead of the Multimodal Learning programme.

When the first large language models appeared, they performed well at what they seemed designed to do: continuing texts, answering questions, writing poetry and code. But even basic mathematics could stump them.

The reason was that initially, the model largely learned to reproduce patterns it had already encountered in the data. If it had seen a similar example, it could give the correct answer. But that did not mean it understood the underlying principle.

Today, the situation has changed.

From memorisation to reasoning

Modern models have learned to break complex tasks into sequences of steps and to check their own reasoning along the way. For mathematical problems, this is particularly important: instead of trying to recall a ready‑made answer, the model can construct a chain of reasoning and arrive at the result.

According to Taslakian, it is precisely the development of reasoning that has recently enabled AI to tackle a number of mathematical problems that had remained open for researchers for decades.

But an important caveat remains: a correct answer does not prove that the model actually arrived at it independently.

If the solution already existed in published materials, it is not always possible to determine whether the model derived it logically or simply reproduced something it had seen during training.

And mathematical problems are only part of the issue.

The hardest part is understanding what caused what

Humans constantly try to establish cause‑and‑effect relationships. If a person takes a pill and a headache goes away, we cannot automatically conclude that the pill helped—it might be that they simply slept well.

For AI, such situations are particularly difficult. A model may find that two events often occur together, but that does not imply an understanding of causality.

This is why, Taslakian believes, the next major step in AI development is to teach models to work not only with statistical correlations but also with cause‑and‑effect relationships. This is especially important in situations where the system encounters something that has not appeared in its training data.

Imagine a completely new type of economic crisis. Or the emergence of a disease with no precedent in historical records. A model that relies primarily on past examples finds itself in a difficult position: it literally has nothing to fall back on.

In such a situation, a human tries to reason about mechanisms: what affects what, what consequences a change in one factor might cause, which elements of the system are truly important.

It is precisely this ability that allows knowledge to be transferred from one situation to another.

Why AI needs this

Causal reasoning is especially important if we want to use AI not just as a conversational partner but as a tool for scientific research and decision‑making.

For example, a model could help discover new treatments. But for that, it is not enough to know that certain events are statistically correlated. One must understand which intervention actually leads to the desired result. The same applies to economics, climate, medicine, and other complex systems where dozens or thousands of factors operate simultaneously.

The problem is that even humans are not always good at correctly identifying the causes of complex events.

Therefore, teaching AI causal reasoning remains a separate research challenge. It may require changes not only to training methods but also to the architectures themselves.

Will better reasoning make AI safer?

Not necessarily.

More advanced reasoning may reduce the number of errors, but AI safety is a much broader issue.

A model may give a logically correct answer that turns out to be dangerous in a particular context. At that point, it is no longer enough to know how to arrive at a result. One must understand whether it makes sense to obtain it and what the consequences might be.

This is especially clear in situations where the cost of error is very high.

An incorrect weather forecast is unpleasant, but survivable. A mistake by a system making decisions in a military context or affecting a human life carries consequences of an entirely different magnitude.

Moreover, models face the problem of values. Humans understand that human life matters and that some actions are unacceptable not only because they lead to an incorrect computational outcome.

Yet humans themselves are not always able to give definitive answers to questions about what is right.

And here a paradox emerges: AI inherits not only humanity's knowledge but also its mistakes, logical fallacies, and contradictions.

When answers become cheap, the key skill is asking questions

According to Taslakian, the rapid development of AI will change not only technology and science but also education.

If systems are becoming increasingly capable of answering almost any question, simply possessing a certain amount of information becomes less important for a human. Far more important is knowing what exactly to ask.

"Knowing how to ask the right question is already becoming a skill that we need to develop much further."

This could be one of the biggest shifts in education.

The problem is that no one yet knows exactly what an education system should look like in a world where every student has a powerful AI model at their side.

One thing is clear: teaching in the way that was done ten or twenty years ago is no longer enough.

Why do we still need humans, if AI knows almost everything?

This very question lies at the heart of the summer school on artificial intelligence that Taslakian and her colleagues have been organising in Armenia for the third year.

The school invites specialists from other countries to tell students about the latest AI research and developments. Participants do not just listen to lectures—they also work on practical tasks.

At first glance, an obvious question arises: why spend time on a school when the same information could be asked of ChatGPT?

Taslakian acknowledges that technically, ChatGPT can indeed tell you about the latest developments.

But learning is not just about receiving information.

Every researcher has experience that has not yet been turned into a paper or a textbook. There is intuition, professional habits, personal mistakes, and observations. There are questions that arise only during live discussion.

And there are other students.

One person may ask a question that would never occur to another. Discussion forces you to clarify your own thoughts, argue, seek explanations, and notice things that are easy to miss on your own.

That is why the organisers deliberately make the school in‑person.

As long as AI does not have a human body or its own life story, humans retain a source of information that is not present in text data: direct experience of interacting with the real world.

And that, perhaps, is one of the most important things that AI cannot yet derive from the internet.