For 100 years, neuroscience has been trying to figure out why we dream.
Why do we spend a third of our lives unconscious — and during a substantial chunk of that, generating elaborate, false, often nonsensical scenarios? Freud said it was wish fulfillment. Crick and Mitchison said it was junk-flushing. Hobson and McCarley said it was the brain stem firing randomly and the cortex inventing stories to make sense of the noise. Activation-synthesis. Memory consolidation. Emotional regulation. Threat simulation. Dozens of theories, none fully satisfying.
In May 2021, a research assistant professor at Tufts University named Erik Hoel proposed something different. In a paper titled “The Overfitted Brain: Dreams evolved to assist generalization,” published in the Cell Press journal Patterns, Hoel argued that the strangeness of our dreams isn’t a bug — it’s the entire point. Dreams exist to keep the brain from becoming too narrowly tuned to everyday life, in exactly the same way that techniques like dropout prevent overfitting in artificial neural networks.
The hypothesis has a beautiful symmetry to it. We spent a century trying to figure out dreams using neuroscience and got nowhere conclusive. Then we built machines that learn — and the machines started having the exact same problem the brain seems to have evolved an elegant solution for. The solution showed us the question.
I want to write about this today because the implications go far beyond sleep science. This is the cleanest biological explanation I’ve ever encountered for why senior leaders gradually lose the ability to navigate change — and what the most adaptive leaders do, consciously or not, to prevent it.
What overfitting actually is
For readers who haven’t worked with machine learning, here’s what overfitting means in plain language.
You’re training an AI model to recognize, say, cats. You show it a dataset of cat photos. It starts to learn the underlying features — fur, whiskers, pointy ears, four legs, certain face proportions. Good model.
But you keep training it on the same dataset. And eventually something subtle starts happening. The model stops learning the concept of “cat” and starts memorizing the specific photos in your training set. It recognizes those exact 10,000 cats perfectly. Show it a 10,001st cat — slightly different lighting, slightly different angle — and it fails.
The machine learning field discovered this is ubiquitous. All deep neural networks face it. The standard solution is what researchers call “noise injection” — deliberately corrupting the inputs to force the network to learn the underlying pattern rather than memorize the specific examples.
Dropout, the most common technique, randomly turns off neurons during training so the network can’t rely on any single pathway. Data augmentation — flipping images, rotating them, distorting them — produces variations the original dataset didn’t contain. All of these techniques have one purpose: prevent the network from learning the specific data set so well that it becomes blind to anything outside it.
This is the mathematical and computational problem.
Now look at your brain.
The overfitted brain hypothesis
Hoel’s hypothesis is that the brain faces exactly the same challenge — that nightly dreams evolved to combat the brain’s overfitting during its daily learning. Dreams, in his formulation, are a biological mechanism for increasing generalizability via the creation of corrupted sensory inputs from stochastic activity across the hierarchy of neural structures.
In plainer language: every day you live a relatively narrow slice of life. Same commute. Same faces. Same routines. If your brain just consolidated that experience on loop every night, you’d become exquisitely fitted to the exact life you’ve been living — and progressively worse at handling anything different.
So your brain does something strange. It generates experiences you didn’t have. Distorted memories. Impossible rooms. People you’ve never met. Conversations that didn’t occur. Hoel argues this isn’t random noise — it’s the brain’s biological version of data augmentation, deliberately corrupting its own inputs to force generalization.
Hoel’s claim: sleep loss, specifically dream loss, leads to an overfitted brain that can still memorize and learn but fails to generalize appropriately.
Every night, your brain generates training data for a life you haven’t lived yet — so you can continue to process new things.
We’ve always thought we dream because we sleep.
Maybe we sleep because we need to dream.
Why this matters for leadership
Here’s where this stops being a meditation on neuroscience and starts being a working diagnosis of why most senior executives stop being able to lead transformations.
Consider the typical inputs an experienced senior leader receives.
Same industry. Same conferences. Same peers. Same board members. Same advisory networks. Same McKinsey reports. Same management books. Same vendor pitches. Same internal town halls. Same quarterly cycles.
The CEO is reading what the other CEOs are reading. The CIO is hearing from the vendors the other CIOs are hearing from. The CFO is benchmarking against the companies that benchmark against them. It’s an ecosystem of converged inputs, sealed off from out-of-distribution data, in which the same patterns are reinforced from every direction.
This is, biologically and computationally, the perfect recipe for overfitting.
The brain solves this problem by generating its own out-of-distribution inputs every single night. The mind doesn’t get a choice — the architecture of sleep enforces it.
The executive calendar enforces the opposite. It is structurally designed to maximize in-distribution input. Every meeting, every report, every conference appearance, every advisory call is calibrated to deepen exposure to the same patterns the leader is already saturated in.
And then we wonder why senior executives — accomplished, intelligent, well-credentialed — gradually lose the ability to lead through change. Their judgment is calibrated against a world that has stopped existing, and they have no biological or organizational mechanism for generating the corrupted inputs that would force generalization.
The pattern recognition that made them valuable becomes the thing that makes them blind.
What “executive dropout” actually looks like
The good news, if you accept Hoel’s framing, is that the fix is structurally simple. You give yourself the same kind of out-of-distribution exposure your brain gives itself every night, but you do it consciously, during the day, in your professional life.
I think of this as executive dropout — borrowing the term from the machine learning technique that turns off random parts of a neural network during training to prevent overfitting.
What executive dropout looks like, practically:
One hour a week with someone outside your industry, generation, or function. Not for networking. Not for a project. For out-of-distribution input. A retired chef. A young creator on TikTok. A high school physics teacher. An architect. Anyone whose pattern library is genuinely different from yours.
One book a quarter from a domain you have no business reading about. Physics. Marine biology. Ancient military history. Bach. Sourdough bread. The point isn’t the topic — it’s that the patterns don’t reinforce the ones you already know.
Regular environments where you are clearly the least expert person in the room. Most executives haven’t been in that position for two decades. It atrophies the muscle of asking questions, sitting with uncertainty, and learning by absorption rather than performance.
At least one person on your senior team whose background is genuinely orthogonal to yours. Not for diversity optics. Because their pattern library is different from yours, and when the world changes underneath you, their library may match the new world better than yours does.
A regular practice of writing about topics outside your professional core. Writing forces the brain to organize information in new ways. When you write about something that isn’t your job, you’re doing in language what dreams do in sensory experience — generating training data that your normal work life would never provide.
None of these are productivity hacks. They are the structural mechanisms by which your judgment stays adaptive in a world that keeps changing.
The AI transformation angle
There’s a specific reason this matters in 2026, beyond the general principle.
AI transformation requires executives to make judgments about a domain that has no historical precedent. Not in their industry. Not in their career. Not in their education. The decisions they’re being asked to make — about workforce composition, about workflow redesign, about which functions to automate and which to preserve, about token-based budgeting models, about how to govern systems they don’t fully understand — fall outside any distribution of inputs their career has prepared them for.
This is the worst possible moment for an overfitted brain.
The leaders I see succeeding with AI right now share one trait, which I’ve now started thinking about through Hoel’s framework. They have a deliberate dream phase in their professional life. They expose themselves, on purpose and on schedule, to inputs their day job would never provide. They read outside their domain. They talk to people younger than them. They show up in rooms where their credentials don’t impress anyone. They build their generalization capacity actively, instead of letting their thirty years of accumulated expertise quietly overfit them to a world that’s disappearing.
The leaders who are struggling are the ones whose calendar contains zero out-of-distribution input. Same industry. Same peers. Same advisors. Same conferences. Same playbook. The pattern recognition that made them senior is the exact thing that’s making them blind, and there is no biological mechanism — no executive equivalent of sleep — that’s correcting for it.
The closing thought
Hoel proposes that dreams help prevent the brain from becoming too narrowly tuned to everyday life, much like dropout prevents overfitting in AI.
The brain figured out, sometime in the last several hundred million years of evolution, that adaptation requires generating inputs that reality doesn’t provide. Without that mechanism, intelligent systems become brilliant prisoners of their training data.
Modern executives are operating without the equivalent.
The conferences, the advisory boards, the executive networks, the curated reading lists — these are sophisticated mechanisms for deepening exposure to the patterns the leader is already saturated in. They are the exact opposite of dropout. They are over-training disguised as professional development.
The leaders who’ll thrive in the next decade aren’t going to be the ones with the deepest expertise in their specific industry. They’re going to be the ones who built their own architecture for executive dropout — a deliberate, scheduled, structural exposure to data their working life would otherwise never generate.
The brain has been doing this for millions of years. Machine learning has been doing it for twenty. Most executives have been doing the opposite for their entire careers.
That gap is the gap between the leaders who’ll lead the next decade and the leaders who’ll spend the next decade being puzzled by why their judgment, which used to be so reliable, suddenly isn’t anymore.
The world has changed. The leaders who notice will be the ones the next decade is built around.
