Why We Outsource Judgment Despite Our 20-Watt Brain

With a 20-Watt Brain, Why Are We Still Outsourcing Judgment?

Back when digital cameras were just emerging, there was a saying that the best value-for-money camera in the world was the human eye. It was a way of highlighting how the eye recognizes complex scenes using very little energy. Still, it’s hard to compare a camera and an eye on a single, unified scale of performance.

Lately I think we can add another sentence to that observation. The human brain does a huge range of things on very little energy. But you can’t directly compare cost-efficiency between a brain and AI without first defining what task is being measured, and how well it’s being done. The brain runs on roughly 20 watts — about as much as a dim incandescent bulb.

And that raises a question. If we already have brains, why are we building massive AI data centers? One key reason is that once a model finishes training, you can copy it and run it in many places at once, simultaneously. It’s hard to see this difference if you simply compare the power one person’s brain uses against the power a data center burns through to handle countless requests at once. And that’s exactly where the real problem starts — right after that, when people go on to hand off even their judgment to the machine.

Where the camera pulled ahead of the eye: copying and transmission

The eye and the camera work differently. The human eye has roughly 6 million cone cells, and some studies estimate the resulting input at about 1.6 billion bits per second. But you can’t directly compare that estimate against a camera’s image quality or its price-to-performance ratio.

What I want to compare here is the ability to copy and transmit a record. A scene the eye has seen can’t be pulled out of your head. It can’t be duplicated, transmitted, or shared identically with someone else. A digital photo, on the other hand, is saved as a file. It’s copied perfectly and opens identically on any device. The camera turns what I saw into a record that other people can also check.

What’s interesting is that even this last wall is now being breached, bit by bit. Research that scans the brain and uses AI to reconstruct the image a person is looking at grows more sophisticated every year. It still requires a scanner the size of a room, and it’s far better at reading what someone’s eyes are actually seeing than what they’re merely imagining in their head. Still, the principle has already been proven.

Replication Is What Separates Brains from AI, Too

The exact same thing happens between brains and AI.

Let’s start with a fact-check. It’s true that the brain runs on 20 watts. But there’s a misconception worth clearing up first: this isn’t electricity the brain generates, but metabolic energy that brain tissue consumes to stay active and alive. And to be precise, it’s a rough estimate somewhere between 14 and 20 watts — not a spec you could write into a purchase order. No matter how hard you think, consumption only rises by about 8% at most. Still, the basic fact that it’s roughly on the order of a single light bulb doesn’t change.

One advantage of AI models is that their trained parameters can be copied and run across multiple systems. You can scale up processing capacity and keep logs of the work. Of course, every run still requires hardware and power, and the same model won’t always produce identical answers. Still, this scales in a fundamentally different way than educating people one by one.

The 20-watt figure represents the brain’s current metabolic energy use. It doesn’t include the cost of raising and educating a person, or the cost of maintaining the entire body. On the AI side, too, we need to separate training cost from inference cost. The real advantage of replication is that a one-time training cost can be shared across many downstream uses.

In practice, this field has already reached the commercial stage. Australia’s Cortical Labs became famous for training roughly 800,000 living neurons to play the game Pong¹, and in 2025 it launched CL1, a commercial biocomputer priced at $35,000. Switzerland’s FinalSpark unveiled a research platform connecting 16 human brain organoids² and offers a remote-access service. In 2024, researcher usage fees were reported at $500 per month — though that doesn’t mean all 16 organoids are leased to a single user.

That said, selling a research device is different from claiming it can replace a general-purpose computer. Cultured tissue has a limited lifespan, and it’s difficult to mass-produce tissue with consistent properties. The line “1,000,000 times more efficient than silicon” circulates widely, but that’s a company-stated possibility “in principle.” To actually compare real-world efficiency, you’d need to account for the energy spent on culture equipment, nutrient supply, and tissue replacement. And even after adding up those costs, whether the same efficiency gap holds up still needs to be independently verified.

This division of labor only works if humans keep making the judgment calls

Looking at these differences, I find myself thinking about how humans and machines might split the work based on what each does best. Computers handle computation and repetitive tasks; humans evaluate the results and make the decisions they’re accountable for. Other computing paradigms, like neuromorphic³ chips or quantum computing, are being researched too, but they haven’t yet secured a practical edge across every optimization problem.

But there’s a quiet premise built into this division of labor: it only works if humans keep holding onto their own share of the work — judgment. And lately, I’ve been watching that very premise start to wobble.

The dangerous part isn’t machines outpacing the human brain. It’s that because machines can cheaply substitute for the kind of thinking that’s always replicable, people end up handing over even their judgment — the one thing that can’t be replicated — to the machine. Even when we have the ability to review the outcome, we skip the step entirely.

Let me walk through a concrete scene. Handing a report draft to AI is fine — polishing sentences and organizing data is close to repetitive work. But if you take the AI’s conclusion and ship it without asking “does this actually fit our situation?”, what you just handed over wasn’t repetition — it was judgment. Code works the same way. The moment you paste in AI-generated code without understanding it, you’re left with a system where, later on, nobody can explain why it behaves the way it does.

The trouble is that none of this is obvious in the moment. Nothing happens on the day you hand off the judgment. If anything, it feels faster and more convenient. The bill comes due much later, after the organization has arrived at a state where nobody knows why a decision was made in the first place — and by then it’s the whole organization that pays. There’s also a concern that individuals lose chances to practice judgment for themselves.

Taken to an extreme, imagine a city where AI runs the power and water systems, but no human can explain how those decisions get made. This isn’t meant to describe reality as it is — I think of it as a scene that illustrates a real problem: the people using these systems need to understand the reasoning behind them.

Oswarld’s Lens

Honestly, I’ve told this story once already, back in September 2025. I wrote a book called People Who Outsource Their Thinking: Homo Brainless, subtitled “A Warning for Modern People Who Have Given Up Thinking in the AI Era.” Finalized in 2024 and released to the world in 2025, the book was, commercially speaking, a flop.

search.shopping.naver.comsearch.shopping.naver.com

Almost nobody read it back then. It was the peak of AI hype, so a warning to “not outsource your thinking to machines” landed as a buzzkill. Everyone was fixated on the upside of handing work off to AI. The stories that sold at the time were all about how amazing AI was, what it could now do, how you should be using it. But lately, the book has started getting read again. It’s not that my diagnosis suddenly became correct. It’s that the market is only now ready to ask that question.

There’s a pattern I’ve seen over and over while building GTM strategies: the better the diagnosis, the longer it takes for the market to become ready to accept it. Forecasts that technology will change something are usually right in direction — it’s the timing and the path that are almost always wrong. The “outsourcing of thought” was the same story. There was a lag between when I wrote about it and when people actually started caring about the problem.

So what I want to say isn’t “stay away from technology.” It’s the opposite. What’s a waste is skipping the judgment call when you could actually review the answer AI gives you.

Closing

The human brain does a remarkable variety of work on very little metabolic energy. AI models have their own advantage: what they learn can be copied and reused across countless systems. So the real question isn’t “how smart are machines getting?” It’s “am I still doing, myself, the judgment calls that can’t be copied?”

When you hand something off to AI, I’d suggest drawing one distinction: is what you’re passing along “repetition” or “judgment”? Hand off repetition all you like — that’s what machines are good at. But if you’re handing off judgment out of habit too, you’re giving up the chance to work it out for yourself.

Have you ever caught yourself mid-delegation lately, thinking, “wait, this was actually mine to judge, and I just passed it along”? Tell me in the comments where you crossed that line, and in what kind of task. Where exactly the boundary between repetition and judgment falls — let’s map it out together with reader examples in the next issue.

---

📨 If you have a colleague who seems to be outsourcing their thinking a little too easily these days, slip them this piece.

References & Further Reading

Primary sources

• Zheng, Jieyu & Meister, Markus, “The unbearable slowness of being: Why do we live at 10 bits/s?”, Neuron, 2025. Link ··· This is the paper that forms the backbone of today’s piece — the one arguing that while our senses take in a billion bits per second, consciousness only processes 10.
• Kagan, Brett J. et al., “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world”, Neuron, 2022. Link ··· This is the study where neurons learned to play “Pong.” That said, the word “sentience” has drawn pushback from other researchers, so read it with that caveat in mind.
• “This $35,000 Computer Is Powered by Trapped Human Brain Cells”, Gizmodo, 2025. Link ··· Gives you the pricing and specs on Cortical Labs’ CL1.
• “World’s first bioprocessor uses 16 human brain organoids”, Tom’s Hardware, 2024. Link ··· Details on FinalSpark’s organoid rental service.
• “The Human Brain Runs on Less Power than a Light Bulb”, Britannica. Link ··· Walks through where the “20 watts” figure comes from — and its limits.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. DishBrain: A system built by growing living neurons on an electrode array. Once researchers set up a signal loop with it, the neurons adapted to playing “Pong” within five minutes. Whether this deserves to be called “intelligence,” though, is still up for debate. ↩

2. Organoid: A tiny clump of tissue grown in a lab from stem cells, mimicking a miniature organ. Brain organoids replicate some of the structure and activity of a real brain, but they don’t “think” the way a human brain does. ↩

3. Neuromorphic computing: Chip design modeled on the brain’s structure. Unlike conventional chips that march in lockstep to a central clock, these chips exchange signals only when needed — using far less energy for optimization problems as a result. ↩