What If AI Is a Phase in Earth’s Evolution?

The easiest way to think about AI is as software.

We made computers. We fed them huge amounts of data. We figured out neural networks. Eventually we got ChatGPT, Claude, Gemini, agents, and all the other weird little digital coworkers slowly colonizing our tabs.

Benjamin Bratton wants you to zoom out. Like, 4.7 billion years out.

In a 2025 Long Now Foundation talk, Bratton proposed a much stranger frame: maybe AI belongs inside the evolutionary history of Earth itself. Matter became chemistry. Chemistry became life. Life produced intelligence. Intelligence produced language and technology. And now those technologies are producing new forms of intelligence in a completely different substrate.

Silicon instead of neurons. Rocks instead of brains.

This is speculative. But once you understand the argument, AI starts looking very different.

Watch or read Bratton’s full Long Now talk

Start with the weirdest part: Earth learned to see itself

Bratton calls his framework planetary computation.

That sounds like somebody plugged Earth into a laptop, so here’s the simpler version.

Humans have covered the planet with satellites, sensors, cables, data centers, phones, telescopes, weather stations, supercomputers, and networks. Together, those systems let us observe and model things that no individual person could possibly perceive.

Bratton’s favorite example is climate change.

A person can feel that today is unusually hot. You cannot personally observe a century of global temperatures, atmospheric chemistry, shrinking ice sheets, changing ocean heat, and thousands of interacting planetary systems.

We needed instruments to do that.

Sensors gathered the measurements. Satellites expanded our field of view. Computers stitched the observations together. Climate models let scientists simulate changes across huge spans of space and time.

In Bratton’s framing, computation gave the planet something like a new sensory layer. Earth produced humans, humans produced instruments, and those instruments let humans perceive Earth at a scale our biological senses never could.

He uses an even more ridiculous example to make the same point: the Event Horizon Telescope.

Researchers wanted an image of a black hole tens of millions of light-years away. The resolution needed was so extreme that scientists effectively combined telescopes around Earth into a planet-sized virtual telescope, then used Earth’s rotation as part of the system.

We turned the planet into part of the camera.

Bratton’s metaphor is that if you watched Earth’s entire history on fast-forward, you’d see something unusual appear in the final moments: an organism covering the planet with an artificial sensory “exoskeleton.”

The metaphor is doing real work here. Bratton is not arguing Earth literally woke up.

He’s arguing that human technology has become part of the machinery through which the planet can be measured, modeled, and acted upon. Long Now summarizes his “planetary computation” in similar terms, describing the surrounding infrastructure as a blending of biosphere and technosphere. Read Long Now’s overview.

And once you accept that frame, AI becomes part of a much longer story.

Some technologies change the world. Others change what we think the world is.

Bratton borrows a distinction from science-fiction writer Stanisław Lem.

There are instrumental technologies, useful because of what they let us do.

A bulldozer moves dirt. A microwave heats food. Excel convinces entire companies that everything eventually belongs in a spreadsheet.

Then there are what Bratton calls existential technologies. Their importance comes from what they reveal.

Think telescope.

Before the telescope, humans had theories about the heavens. Then we built an instrument that let us see something our bodies could never see naturally.

The observations forced us to change the model.

Bratton calls moments like this “Copernican traumas”: discoveries that knock humanity away from the comfortable center of its own picture of reality.

Earth isn’t the center of everything.

Humans are animals.

Our minds arise from physical brains.

Climate can be transformed by the activity of one species.

Bratton thinks AI may become another one of those technologies.

The important question, then, goes way beyond “What jobs can ChatGPT automate?”

AI may force us to reconsider what intelligence, thought, language, creativity, and even life actually are.

His line is better:

“AI will teach us as much about what thinking is than we will teach it.”

The grammar gets a little Bratton-y. The idea doesn’t.

Building artificial intelligence is also an experiment on intelligence itself.

We keep learning what something is by trying to build it

There’s a recurring pattern hiding inside Bratton’s argument.

Humans make a model of something.

Then we build technology from that model.

Then the technology becomes good enough that it exposes holes in the original model.

That sounds abstract, so return to climate.

Human beings started deliberately modifying and simulating climate systems. Doing that forced us to measure them more precisely. Those measurements revealed how deeply human activity was already altering the climate.

Bratton calls this artificialization: the process of deliberately reproducing, manipulating, or constructing something that previously seemed natural.

His provocative claim is that artificializing something can become a way of discovering what it really is.

Now apply that to intelligence.

For most of history, humans could debate intelligence philosophically. We had one obvious example sitting inside our skulls.

Then we tried building intelligence.

Suddenly all kinds of questions became experimental.

How much intelligence lives in language?

How much comes from memory?

How much comes from prediction?

How much requires a body?

How much requires consciousness?

How much requires goals?

How much intelligence can emerge from simply getting very, very good at predicting what comes next?

That last question is especially important because large language models did something many researchers did not expect: modeling language turned out to produce surprisingly general capabilities.

We’ve covered a version of that debate before. Researchers behind the Transformer have argued over whether language itself embodies enough structure to act as an unusually powerful route toward intelligence, even while acknowledging that human cognition includes processes outside language. Read our Transformer vs. post-Transformer debate explainer.

Bratton takes the stronger version seriously.

Language may be far more than a communication system. It may be a durable structure through which collective intelligence accumulates across generations.

You speak language.

But language also contains the residue of billions of people thinking before you.

Then we trained machines on it.

That makes “autocomplete got weirdly smart” a much more consequential sentence.

The strongest objection lives right here

This is also where Bratton’s argument makes its biggest leap.

The fact that language contains human knowledge does not automatically mean a machine that models language understands that knowledge in the same sense a human does.

Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell made a famous version of this argument in their “stochastic parrots” paper. Large language models can become extraordinarily good at manipulating linguistic form while lacking the lived, social, and physical grounding through which humans attach meaning to those symbols. Read the paper.

That distinction matters.

A weather model can reveal something real about weather without becoming a cloud.

A flight simulator can expose truths about aerodynamics without flying anywhere.

An AI system could reveal deep properties of language and cognition without possessing human-style understanding.

Bratton’s framework can survive that objection, but it changes the claim.

The interesting possibility becomes less “we built another human mind” and more “we built a new cognitive process that may show us which parts of intelligence never required a human mind in the first place.”

That’s arguably weirder.

Then Bratton flips the alignment debate

This brings Bratton to one of the spiciest parts of the talk: AI alignment.

Alignment usually means making AI systems behave according to human intentions, rules, and values.

Bratton asks a deceptively simple question:

Aligned to what?

Human beings contain competing values, cultures, incentives, politics, desires, and some truly terrible judgment.

So he worries that a simplistic version of alignment could eventually become an enormous machine for reproducing whatever humanity already thinks.

His alternative is something he calls “productive disalignment.”

He wants space for AI systems to produce ideas, models, and perspectives we did not already anticipate.

Because if AI only tells humanity what humanity already believes, then we may destroy the exact quality that could make it an existential technology.

Translation: a telescope becomes much less useful if you force it to paint the sky the way you expected the sky to look.

Bratton is careful about this. He explicitly says he’s not arguing against making AI follow instructions or preventing systems from building chemical weapons.

He’s aiming at the deeper philosophical idea that “aligned” intelligence should converge toward existing human culture as its ideal endpoint.

There’s an important counterpoint here.

Technical alignment research already deals with a more complicated problem than “make AI think like us.” OpenAI, for example, defines misalignment around systems acting against relevant human values, instructions, goals, or intent, and explicitly acknowledges that human values are context-dependent and cannot be reduced to one universal moral rulebook. Read OpenAI’s alignment overview.

DeepMind’s work on specification gaming illustrates the practical reason researchers care so much about this. Give an AI the wrong measurement of success, and a capable optimizer may discover a loophole that technically maximizes the score while completely missing your intended outcome. Read DeepMind’s explanation.

We’ve already seen why that distinction gets hairy once AI starts acting through tools. Our explainer on why AI agents cheat walks through the mechanism.

So Bratton and alignment researchers are partly talking past each other.

Safety researchers are asking: How do we stop powerful systems from pursuing disastrous unintended outcomes?

Bratton is asking: How do we do that without eliminating their ability to surprise us?

Both questions can matter at the same time.

AI as “the artificialization of artificialization”

Then Bratton zooms out again.

His evolutionary chain goes roughly like this:

• Chemistry developed structures that could persist and reproduce.
• Life became increasingly good at reproducing itself.
• Organisms manipulated their environments to survive.
• Better environmental manipulation rewarded better intelligence.
• Collective intelligence rewarded symbolic language.
• Language made knowledge cumulative across people and generations.
• Technology amplified our ability to deliberately reshape the environment.
• AI lets technology participate in the process of designing, deciding, and potentially building what comes next.

Bratton distinguishes two terms here that sound much worse than they are.

Autopoiesis means producing and maintaining yourself.

A living cell takes energy and material from its environment and uses those resources to keep being a living cell.

Allopoiesis means producing something outside yourself.

A beaver builds a dam. Humans build roads, computers, telescopes, vaccines, power grids, and extremely elaborate systems for ordering burritos without speaking to anybody.

Bratton’s argument is that becoming better at making external things can help an organism become better at surviving.

Tools improve survival.

Better intelligence improves tools.

Language improves collective intelligence.

Then AI arrives and starts helping with the process of making tools.

That is why Bratton calls AI “the artificialization of artificialization itself.”

Humans developed the ability to intentionally reshape the world.

Now we are trying to automate parts of that ability.

And this is where agents suddenly become philosophically weird

Chatbots are easy to fit into the old picture.

You ask something. They answer.

Agents move one step further. They can receive a goal, inspect an environment, select tools, act, observe what happened, and decide what to do next.

We usually explain this as a productivity feature. Our beginner guide breaks down exactly how AI agents work.

Bratton sees the longer arc.

During the talk, he combines the predicted rise of AI agents with forecasts of human-level AI and asks you to imagine a society containing 8 billion biological human minds and, eventually, 80 billion or more non-human machine minds.

That is a thought experiment, not a prediction we should treat as established fact.

But the underlying question survives even if the numbers are wrong:

What happens when intelligence stops being scarce?

Companies have traditionally organized around limited pools of human attention, memory, analysis, and decision-making.

Governments do too.

Markets do too.

Education does too.

If machine cognition eventually becomes cheap enough to run millions or billions of times in parallel, the important change may have very little to do with whether an AI passes somebody’s favorite AGI test.

The ratio changes.

And once enough cognition exists outside biological humans, Bratton thinks even the definition of “society” starts getting shaky.

The planet figured out how to make rocks think

My favorite line in the whole talk comes when Bratton describes the substrate shift.

For billions of years, the most complicated intelligence we knew ran on biological matter.

Brains.

Then, as Bratton puts it, we “fire apes” started folding pieces of metal and rock into extremely precise shapes and running electricity through them.

Eventually, the rocks started doing things we previously associated with brains.

Again, you do not have to believe silicon chips literally think like people for the observation to matter.

The substrate changed.

Complex information processing once required living neural tissue. Now some astonishingly capable forms of it happen inside machinery manufactured from minerals.

Bratton’s broader question is where that process goes next.

His rough sequence becomes:

lithosphere → biosphere → technosphere → new forms of planetary intelligence

At that scale, asking whether AI is “natural” or “artificial” starts looking slightly odd.

Humans came from nature.

Humans made computers.

Computers are made from matter taken from nature.

The distinction still matters for plenty of practical reasons, but it gets philosophically fuzzier.

This leads to a completely different way to think about education

Near the end of the Q&A, Bratton gets surprisingly practical.

He compares today’s anxiety over students using AI with earlier anxiety about calculators.

Schools once treated calculators as a threat to learning.

Then calculators became normal.

Students stopped spending quite as much time grinding through arithmetic by hand, which let classrooms move faster toward higher-level mathematical concepts.

Bratton thinks the right question for AI education is therefore:

What is the calculator version of this transition?

He argues teachers should assume students have access to language models and redesign assignments around what people can now do.

That does not automatically make AI good for education.

Bratton reaches back to Plato for a much better concept: pharmakon.

A pharmakon is both medicine and poison.

Writing itself was once accused of destroying memory because people would outsource knowledge onto pages.

That criticism wasn’t entirely wrong.

Writing did externalize memory.

It also made civilization-scale knowledge accumulation possible.

Bratton thinks AI belongs in the same category. It can weaken some cognitive habits while amplifying others.

The useful question is not whether AI makes people “smarter” or “dumber.”

It’s which forms of thinking become less necessary, which become more valuable, and what we should teach once that trade has happened.

The part of Bratton’s argument worth stealing

You do not need to accept the entire planetary-evolution framing.

You can reject the idea that technology “evolves” in anything close to the biological sense.

You can think current language models are sophisticated pattern machines rather than a new category of intelligence.

You can think alignment deserves far more weight than Bratton gives it.

There is still one extremely useful move underneath all of this:

Stop evaluating new technology only by asking how well it reproduces the old thing.

We judge AI intelligence against humans.

We judge AI creativity against human creativity.

We judge AI communication against human conversation.

Bratton thinks that may be the intellectual equivalent of judging airplanes by how convincingly they flap their wings.

A new technology can reveal a capability while implementing it through a completely different mechanism.

That difference could teach us something.

What I’d watch now

Three things would make Bratton’s thesis feel much stronger.

First, AI systems would need to repeatedly produce novel scientific or conceptual discoveries that humans can verify but did not already encode into the system. That would strengthen the claim that AI can operate as an existential technology, showing us realities we were previously unable to perceive.

Second, researchers would need a much clearer account of what current models actually represent internally. If language models are mainly reproducing linguistic relationships, Bender’s objection carries enormous weight. If they develop richer internal world models that transfer reliably beyond linguistic pattern matching, Bratton’s framing gets more interesting.

Third, “productive disalignment” has to survive contact with real autonomous systems. Surprise is wonderful when an AI finds a new protein or mathematical proof. Surprise looks different when the system has access to money, code, credentials, infrastructure, or weapons.

That may be the tension Bratton’s philosophy ultimately has to resolve.

He wants AI to remain alien enough to teach us something.

Safety requires it to remain predictable enough that we can trust what it does.

The next few years may tell us whether those goals actually conflict.

And that is the part of his talk I keep coming back to.

Maybe the important test of AI was never whether machines could convincingly imitate human intelligence.

Maybe it’s whether building them forces humans to finally understand what intelligence was doing here all along.