Experts Say AI Still Lacks Real-World Judgment

Artificial intelligence has made remarkable strides in generating text, writing software and solving complex mathematical problems, but leading researchers say today's most advanced AI systems still lack a genuine understanding of the physical world.

Experts argue that large language models (LLMs) such as ChatGPT, Claude and Gemini excel at recognising patterns in vast amounts of data but remain fundamentally limited when it comes to reasoning about real-world environments.

Among the critics is AI pioneer Yann LeCun, who says current AI systems cannot match even the basic physical intuition of a rat.

"We don't have robots that are nearly as good at understanding the physical world as a rat," LeCun said at the VivaTech technology conference in Paris.

After leaving Meta in 2025, LeCun founded Advanced Machine Intelligence Labs (AMI Labs) to develop a new generation of AI capable of understanding and interacting with the real world, particularly in robotics and autonomous systems.

According to LeCun, today's language models perform well because they are trained to predict words and patterns, making them highly effective at writing, coding and answering questions. However, he argues that these systems do not truly understand how the physical world works.

"They basically accumulate knowledge and reproduce it, but they are not particularly smart because they don't have a true understanding," he said.

LeCun explained that current AI relies on statistical prediction rather than genuine reasoning. He cited the example of a pen balanced upright: while even a young child understands it will fall when released, an AI model cannot reliably reason about the countless physical factors that determine exactly how it will fall.

To overcome these limitations, AMI Labs is developing a new architecture known as Joint Embedding Predictive Architecture (JEPA).

Unlike conventional language models that attempt to predict every possible outcome, JEPA creates simplified representations of the physical world, enabling AI to focus on meaningful information while ignoring unnecessary details.

The project has already attracted significant financial backing. Earlier this year, AMI Labs secured more than $1 billion in seed funding—one of Europe's largest early-stage AI investments—with support from major technology and investment firms.

Researchers say real-world understanding has become increasingly important as companies invest heavily in humanoid robots and autonomous machines.

Although robotics has advanced rapidly, teaching robots to perform everyday household tasks—such as ironing clothes, loading dishwashers or handling delicate objects safely—remains one of the industry's biggest challenges.

LeCun believes existing language models alone are unlikely to solve those problems.

"LLMs are largely hopeless for robotics," he said, rejecting suggestions that simply increasing the size of today's AI models will inevitably lead to superhuman intelligence.

His assessment is shared by several researchers.

Ingmar Posner, Professor of Applied Artificial Intelligence at the University of Oxford and director of its Applied AI Lab, said future AI systems must be capable of explaining their decisions and understanding cause-and-effect relationships.

"You need models that can answer questions like: What matters? What causes what? What would happen if I took a different action?" Posner said.

His research team has spent the past four years developing World Models, an approach that enables AI to build internal simulations of its environment before making decisions.

The concept, first proposed decades ago, gained renewed attention after researchers suggested that AI could learn more effectively by constructing internal models of the world instead of relying solely on pattern recognition.

Since then, several major technology companies have expanded research in the field. Experimental systems have already demonstrated the ability to plan actions by simulating future scenarios before acting.

Posner's team is also developing what it calls a "mechanistic world model," designed to organise and update knowledge more efficiently while improving AI's reasoning capabilities.

Even so, researchers caution that predicting when these technologies will become practical remains difficult.

They note that only a few years before ChatGPT emerged in late 2022, many experts believed comparable systems were still decades away.

Several leading AI companies are now investing heavily in world-model research as they seek to develop machines capable of understanding and interacting safely with complex real-world environments.

LeCun said AMI Labs expects to begin deploying its technology in industrial applications next year before gradually expanding it into broader real-world uses.

If successful, he believes the approach could pave the way for general-purpose AI systems capable of performing a wide range of practical tasks with minimal additional training.

Despite rapid technological progress, LeCun said humans will remain central to innovation and decision-making.

"We're still going to need humans to figure out what questions to ask, what to build and what to create," he said.

He expects future AI systems—even those that surpass humans in certain capabilities—to function primarily as powerful assistants rather than replacements.