Artificial intelligence (AI) is opening up many breakthroughs, but there is still a core limitation that cannot be self-taught after being deployed.
Unlike humans, especially young children who constantly adapt to the environment, current AI models are almost "frozen", only operating based on data that has been trained before.
According to research published on March 17, 2026 by leading AI research scientists - Emmanuel Dupoux (field of cognitive science - FAIR at Meta), Yann LeCun (field of artificial intelligence, deep learning - Professor at New York University) and Jitendra Malik (field of computer vision, Professor at California University Berkeley), the reason lies in how AI is built.
Most modern systems depend on the MLOps process, where people collect data, train and update models in batches. When the environment changes, AI cannot self-adjust but needs to be retrained from the beginning.
This makes AI prone to failure in real-world situations different from training data. Language or visual models can recognize patterns very well, but lack adaptability and do not learn from their own mistakes.
Research points out two core learning mechanisms that need to be combined. The first is System A (learning from observation). This is how people build their understanding of the world through seeing, hearing and predicting.
Current AI models mainly belong to this group with the advantage of being able to expand and detect laws from big data. However, the weakness is that they are not linked to actual action and it is difficult to distinguish cause and effect relationships.
The second is System B (learning from action), based on trial and error. This is how people learn to walk, learn to speak or solve problems. The advantage of this system is the ability to discover new solutions, but it consumes a lot of data and time.
In nature, these two systems always operate simultaneously. Humans both observe and act, constantly adjusting to optimize behavior. Conversely, AI today separates these two mechanisms, limiting learning ability.
To overcome this, researchers propose adding the M System (super-control), which acts as a "controlling brain".
This system tracks errors, levels of uncertainty and performance, thereby deciding when to learn from observation, when to experiment. In other words, AI will know when to ask itself what to learn and how to learn.
This approach is inspired by people like children who will explore when uncertain, practice when they understand, and even consolidate knowledge while sleeping.
If successfully applied, AI can self-adjust learning strategies without constant human intervention.
The research group also proposed a development model in two time scales, including: "life cycle" - where AI learns during operation and "evolution" - where the super-control system is optimized through millions of simulations. This is considered a closer step towards AI capable of autonomous learning.
However, the challenge is not small, because building sufficiently fast and practical simulation environments requires huge computing resources. At the same time, self-learning AI also raises safety concerns when it can act unexpectedly.
However, scientists believe that this is a necessary direction. Not only does it help AI operate more efficiently in the real world, this research also contributes to explaining how humans learn and adapt, which is one of the biggest mysteries of intelligence.
