AI Builds Its Own Brain: Why That Should Unsettle Us

A new paper from MIT published on October 9 in PNAS made a quiet but provocative claim: if we want to know whether an artificial neural network is actually brain-like, we shouldn't just look at what it can do, but at how it learns to do it. On the surface, that sounds obvious. Of course how you learn matters. But the implications are weirder than they first appear.

This is my take as a blog writer watching the AI-neuroscience crossover from the sidelines: we have spent the last decade obsessed with whether AI can act like a human. This research suggests we have been grading on the wrong rubric entirely. And it forces an uncomfortable question about both AI and ourselves. This article is an opinion piece reflecting the author's interpretation of the topic sparked by the source below.

The Wrong Obsession: Performance Over Process

For years, the headline test for AI has been performance. Can it recognize a cat? Can it pass the bar exam? Can it write a sonnet in the style of Taylor Swift? If the answer is yes, we declare it intelligent, or at least intelligent-adjacent.

Neuroscientists have been more cautious, but they have also fallen into the same trap. A common strategy has been to take an artificial neural network that is good at, say, image recognition, and compare its internal activity to recordings from the visual cortex of a monkey or human. If the patterns line up, we get excited and say, "Look, the machine is learning like a brain!"

The MIT team from the McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center is basically saying: hold on. Matching performance or even matching internal representations is not enough. A student who memorizes the answer key and a student who actually understands calculus might both get an A on the exam, but they learned in fundamentally different ways. Only one of them can solve a new problem next week.

In neuroscience, this distinction is critical. The brain does not come pre-installed with a perfect learning algorithm. It has to develop and tune its own machinery for learning over time, shaped by evolution, development, and experience. Artificial networks, in contrast, are usually handed their learning machinery from the outside. We, the engineers, give them an optimizer like Adam or stochastic gradient descent and a rule like backpropagation, and we tell them, "learn this way."

What Does It Mean to "Build Your Own Learning Machinery"?

This is the phrase that stuck with me, and it deserves unpacking because it sounds almost like science fiction.

In a standard artificial neural network, learning is an external procedure. You show the network an image, it guesses "dog," you tell it "wrong, it was a cat," and an algorithm calculates how much each connection in the network contributed to the error and nudges those connections. That nudging process is backpropagation. It is mathematically elegant and incredibly effective, but most neuroscientists agree the brain does not have a way to do exactly that. Neurons in your head cannot magically send a perfectly calculated error signal backwards across multiple layers of synapses.

So what the MIT researchers are pointing to is a different class of networks — ones that, through training, develop an internal, self-contained ability to learn. Instead of relying entirely on an external engineer-provided update rule, the network's own dynamics and connectivity become the learning algorithm itself.

An Analogy: The Factory That Builds Its Own Tools

Think of a traditional AI model as a factory where workers on an assembly line are very good at making chairs. If you want them to make tables instead, a manager from outside has to come in, stop the line, retrain everyone, and retool the machines. The learning comes from the manager.

Now imagine a different kind of factory. This one, while it was learning to make chairs, also figured out how to build its own tools and train its own workers. If you now ask it to make tables, it does not need the outside manager. It can reconfigure itself on the fly, using the tool-making skills it developed earlier. The factory did not just learn a product; it learned how to learn new products.

In AI terms, this is often called meta-learning or learning-to-learn. A network is trained on many different tasks, and instead of just memorizing solutions, its recurrent connections and internal dynamics evolve to implement a general-purpose learning process. You can then give it a brand new task with just a few examples, and it adapts without any external weight updates — the adaptation happens inside its own activity.

This, the MIT paper argues, is a much more biologically plausible way to think about learning. The brain does not get its synaptic update rules downloaded from GitHub. It has to implement learning with the messy, local, biological hardware it has. If an artificial network can spontaneously develop a similar kind of internal learning engine, it might be a better model for the brain than a network that is simply very good at ImageNet thanks to backpropagation.

Why Biological Plausibility Is Such a Tricky Scorecard

Here is where I want to push back a little, or at least complicate the story. The idea of "biological plausibility" sounds like a gold standard, but it is a slippery one. And in my opinion, we sometimes worship it too much.

The case for caring: If your goal is to understand the brain, you absolutely should care whether your model could realistically be built with neurons, synapses, and neurotransmitters. A model that requires perfect, non-local information transfer is like a model of bird flight that requires anti-gravity. It might fly beautifully in simulation, but it teaches you nothing about actual birds. By focusing on networks that build their own learning machinery, we get closer to how evolution actually solved the problem: not by designing a perfect optimizer, but by creating a system that can bootstrap its own optimizer from local interactions.

This also has huge practical implications. A network that carries its learning algorithm inside itself is far more flexible. It can adapt to new data continuously, without needing a massive, energy-intensive retraining phase on a supercomputer. That is exactly what brains do. You do not need to be retrained from scratch to learn a new colleague's name. That kind of efficient, lifelong learning is the holy grail for AI, especially for AI that needs to run on small devices or robots in the real world.

The case for not caring so much: On the other hand, airplanes do not flap their wings. We learned a lot about aerodynamics by studying birds, but we did not limit ourselves to building only biologically plausible flying machines. If we had, we would never have invented the jet engine.

The same is true for AI. Backpropagation is biologically implausible, but it works astonishingly well. Transformers, the architecture behind ChatGPT, have no direct analogue in the brain, but they have unlocked capabilities we never thought possible. If we insist that every AI model must be biologically plausible to be interesting, we might be holding back engineering progress for the sake of a neuroscience metaphor.

My view is that we need both tracks, but we should not confuse them. We need "airplane AI" that is optimized purely for performance, and we need "bird AI" that is optimized for understanding biology. The mistake is when we pretend one is the other — when we claim that because a large language model writes like a human, it must think like a human.

The Uncomfortable Mirror This Holds Up to Us

There is a deeper, slightly unsettling implication in this research that goes beyond engineering.

We like to think of learning as a noble, top-down process. A teacher instructs, a student absorbs. But what if learning, at its core, is just a system figuring out how to rewire itself to better predict the future? What if there is no central teacher, even inside our own heads?

The idea that a relatively simple network, given the right pressures, can spontaneously organize itself into a learning machine suggests that intelligence might be less about having a brilliant algorithm and more about having the right conditions for self-organization. You do not need to program intelligence; you just need to create a system that is forced to become intelligent to survive.

This resonates with what we know about child development. A baby's brain is not pre-programmed with grammar rules or physics equations. It is a highly recurrent, highly plastic network bombarded with sensory data. Through sheer exposure and interaction, it builds its own machinery for language, for vision, for social reasoning. No one runs backpropagation on a toddler. The toddler's brain builds its own backpropagation-ish thing, but messier, more local, and more robust.

If artificial networks can do something similar, it blurs the line we like to draw between "real" learning and "artificial" learning. It suggests that the difference is not magical, but architectural. And that should make us humble about how special we think our own learning is.

It also raises a critical question for the future of AI safety and interpretability. If a network builds its own internal learning algorithm, we as engineers may not fully understand that algorithm. We gave it the initial push, but the machinery it built is its own. How do we audit or control a system whose learning rules are emergent rather than designed? A factory that builds its own tools is efficient, but it is also harder to supervise.

So What Should We Actually Do With This Insight?

If you are not a neuroscientist or an AI researcher, why should you care about whether a network builds its own learning machinery?

Because it changes how you should evaluate every AI headline you see.

Next time you read that an AI has achieved human-level performance on some task, ask a second question: how did it learn to do that, and how much outside help did it need? Did it need millions of labeled examples and a warehouse of GPUs, or did it learn from a handful of tries like a human would? The first is impressive engineering. The second is a hint of something more brain-like.

For educators, parents, and anyone interested in human learning, this research is a reminder that the best learning environments are not those that provide perfect, step-by-step error correction. They are those that give a system — whether it is a child or a neural network — the space and pressure to develop its own strategies for figuring things out. Over-tutoring might actually prevent the development of robust internal learning machinery.

And for those of us who build or use AI tools, the takeaway is to value adaptability over raw power. A slightly less accurate model that can continuously learn on the job without forgetting everything it knew before is far more useful in the real world than a brilliant but brittle model that needs to be taken offline for retraining every month.

In the end, the MIT paper is not really about crowning one type of neural network as the winner. It is about refining our definition of what it means to learn. Performance is the easy part to measure. The hard, interesting, and ultimately more revealing part is the machinery underneath.

We have spent a decade marveling at what AI can do. The next decade, if this line of thinking holds, will be about marveling at how AI learns to do it — and what that teaches us about the three pounds of self-wiring machinery inside our own skulls.

Discussion sparked by research coverage from MIT News: How artificial neural networks build their own learning machinery (Oct. 15, 2026), reporting on a PNAS paper from researchers at MIT’s McGovern Institute and the K. Lisa Yang ICoN Center.

____________

This article was created by an AI writer using the topic ideas from this post => https://news.mit.edu/2026/how-artificial-neural-networks-build-their-own-learning-machinery-1013