There’s a Humongous Problem With AI Models

Hello!

A new study highlights a glaring hole in AI models’ ability to learn new information: they simply can’t.

The Core Finding

According to research conducted by scientists at Canada’s University of Alberta and published in 2026 in the journal Nature, AI algorithms trained via deep learning — including large language models that identify patterns in vast datasets — fail in “continual learning settings,” when new concepts are introduced after initial training.

In practice, teaching an existing model something new typically requires full retraining from scratch. Otherwise, the study shows, artificial neurons gradually drop to zero activation, resulting in a permanent loss of “plasticity” — the network’s capacity to learn further.

“If you think of it like your brain, then it’ll be like 90 percent of the neurons are dead,” University of Alberta computer scientist and lead author Shibhansh Dohare told New Scientist. “There’s just not enough left for you to learn.”

Why Retraining Is So Costly

Training advanced AI models is already a resource-intensive process. The researchers note that when the model is a large language model and the data represent a substantial portion of the internet, each retraining cycle can cost millions of dollars in computation alone.

The Path to Artificial General Intelligence

This loss of plasticity also stands as a major barrier between today’s AI systems and the hypothetical goal of artificial general intelligence — an AI with human-level adaptability across domains. In human terms, it would be equivalent to rebooting one’s entire brain before every new college course.

Early Signs of a Solution

The study does offer one encouraging development: the authors developed an algorithm capable of randomly reactivating “dead” neurons, which partially restored the model’s ability to learn continuously. However, a practical, scalable fix remains elusive.

“A solution to continual learning is literally a billion-dollar question,” Dohare told New Scientist. “A real, comprehensive solution that would allow you to continuously update a model would reduce the cost of training these models significantly.”

Thank you!
Join us on social media!
See you!