Researchers propose three-system model to give AI human-like learning ability
A team of leading AI scientists has proposed a framework to address a fundamental limitation in deployed artificial intelligence: the inability to learn and adapt after training ends.
Emmanuel Dupoux of Meta's Fundamental AI Research, Yann LeCun of New York University, and Jitendra Malik of UC Berkeley published their findings on March 17, 2026. Their research identifies why current AI systems remain essentially static once deployed, unlike humans who continuously adjust to new environments.
The frozen model problem
Most modern AI relies on batch processing: humans collect data, train models, then deploy them. When conditions change, the system cannot adapt. It requires retraining from scratch.
This explains why language and vision models fail in real-world situations that differ from their training data. They recognize patterns well but cannot learn from mistakes or adjust their own behavior.
Two existing learning systems
The researchers identified two mechanisms humans use to learn. System A operates through observation-seeing, hearing, and predicting patterns. Current AI models primarily rely on this approach, which scales well with large datasets but lacks connection to actual action and struggles with cause-and-effect reasoning.
System B learns through trial and error. Humans use this when learning to walk, speak, or solve novel problems. It discovers new solutions but demands substantial data and time investment.
Humans deploy both systems simultaneously. They observe while acting, constantly refining behavior. AI systems today keep these mechanisms separate, limiting their learning capacity.
The control system solution
The researchers propose adding System M, a meta-control layer that monitors errors, uncertainty levels, and performance. This system decides when to learn from observation versus when to experiment-essentially allowing AI to determine what and how to learn.
The approach mirrors how children explore when uncertain, practice when confident, and consolidate knowledge during rest. If implemented successfully, AI could adjust learning strategies without constant human intervention.
The team also outlined a two-timescale development model. The "life cycle" phase involves learning during operation. The "evolution" phase optimizes the control system through millions of simulations, moving closer to truly autonomous learning.
Practical obstacles remain
Building sufficiently fast simulation environments requires enormous computational resources. Self-learning systems also raise safety concerns-AI that acts unexpectedly could create risks.
Despite these challenges, the researchers argue the direction is necessary. Beyond improving real-world AI performance, the work addresses a larger question: how humans themselves learn and adapt, one of intelligence's enduring mysteries.
Learn more about AI Research and Generative AI and LLM advances.
Your membership also unlocks:
