Think back to the first time you pictured AI.

Maybe it was HAL's calm red eye refusing to open the pod bay doors. Maybe it was the Terminator. Maybe you never saw either one and still knew the story anyway, because it's everywhere: we build the machine, the machine wakes up, and the machine comes for us.

Whether you watched the movies or not, that's the picture a lot of us grew up with. Robot means threat. In 2022, people in San Francisco put it right on a banner.

A rally on the steps of San Francisco City Hall with a yellow banner reading 'We all saw that movie... No killer robots.'

San Francisco City Hall, December 5, 2022, protesting a vote to let police use robots with deadly force. Photo: Pax Ahimsa Gethen, CC BY-SA 4.0, via Wikimedia Commons.

So here's the question I keep asking. Of protesters, of headlines, of the people building it: how exactly does AI end the world? Not the movie version. The real one. What happens first, and where's the proof?

Most of the time, nobody gives me details. And when the insiders do, what I get is a scenario, not proof.

What they say they're building

Here's what surprised me. The people building AI don't describe a movie. They describe something closer to paradise. In their own words:

  • Dario Amodei, Anthropic: "my basic prediction is that AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years." He describes what's coming as a "country of geniuses in a datacenter." (Machines of Loving Grace, October 2024)
  • Sam Altman, OpenAI: "In the 2030s, intelligence and energy—ideas, and the ability to make ideas happen—are going to become wildly abundant." (The Gentle Singularity)
  • Mark Zuckerberg, Meta: "Meta's vision is to bring personal superintelligence to everyone." (Personal Superintelligence)
  • xAI, Elon Musk's company, on its own page: "Accelerate human scientific discovery. We build AI to understand the universe."
  • Demis Hassabis, Google DeepMind: "Artificial General Intelligence (AGI), a system that exhibits all the cognitive capabilities the brain has, is probably only a few short years away." Same essay: "advances on the frontier are outpacing our understanding of the technology." (A Framework for Frontier AI and the Dawning of a New Age, July 14, 2026)
  • Jensen Huang, NVIDIA's founder: "AI will transform every industry, power every company, and be built by every country." (July 24, 2026)
  • Satya Nadella, Microsoft: "In my view, our priority has to be building a frontier ecosystem, not just a frontier model, so value flows broadly across every company, every industry, and every country." (A frontier without an ecosystem is not stable, June 14, 2026)

A century of medicine in a decade. Abundance. A superintelligence for everyone. AI in every industry. Understanding the universe. Hold onto that, because some of the same people have also signed their names to warnings that sound nothing like it.

Seven AI leaders with their own words: Altman on intelligence and energy becoming 'wildly abundant,' Amodei's 'country of geniuses in a datacenter,' Zuckerberg's 'personal superintelligence to everyone,' xAI's 'understand the universe,' Hassabis saying AGI 'is probably only a few short years away,' Huang's 'AI will transform every industry,' and Nadella's 'value flows broadly.' Below: the 2023 extinction statement signed by Altman, Amodei and Hassabis, and the March 2023 pause letter signed by Musk.

Warnings, not proof

On March 22, 2023, the Future of Life Institute published a letter calling on "all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4." Elon Musk signed it. On the list when I checked, his name was third, as "CEO of SpaceX, Tesla & Twitter."

112 days later, xAI's own company page lists its "Company announcement," July 12, 2023. Today that same page advertises Colossus: "200K GPUs. Built in 122 days."

Sign the pause. Start a new AI company 112 days later.

He's not the only one. In 2023, the Statement on AI Risk said: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Among the signers: Sam Altman, "CEO, OpenAI." Dario Amodei, "CEO, Anthropic." Demis Hassabis, "CEO, Google DeepMind." That's the whole statement. One sentence. The people running three leading AI labs put their names under the word extinction, and then kept building.

To be fair, some of them do describe how it could go wrong. Amodei's own essay walks through scenarios: AI helping someone build a biological weapon, AI systems whose own goals turn out to be hostile to ours, authoritarian governments using it to surveil and repress. Those are worth taking seriously. But a scenario isn't proof. The question I never hear answered is how likely each one is, what evidence backs it, and whether a given law would actually stop it.

And when they write at length, they hedge. Sam Altman, in The Gentle Singularity: "There will be very hard parts like whole classes of jobs going away, but on the other hand the world will be getting so much richer so quickly that we'll be able to seriously entertain new policy ideas we never could before." The world getting richer is not the same as the person who lost the job being okay. That second part is the one that needs proof, the same as the warnings do.

A warning doesn't have to be proof. But it does have to show its evidence, its uncertainty, and why the proposed response fits the risk. Without that, it's a feeling. And feelings can end up with real power.

I learned that one in person. Some friends introduced me to someone running for office, so I asked what policies they were looking into. As close as I can remember it, the answer was: "Ha, you got me there. I don't know." I laughed and said I wasn't trying to catch anybody out. Then they looked me over and said I looked like someone who's for data centers. I said, so what if I am? That was the end of it. They told their friend to stay away from me, I was for data centers, and they walked off.

They won. That night I didn't hear a single policy, just a feeling, and now that person has a real vote.

The warnings, quoted in a bill

Dario Amodei runs Anthropic, one of the biggest AI companies in the world. In his January 2026 essay he calls job loss "a topic that I warned about very publicly in 2025, where I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress."

Here's how Bernie Sanders' AI Data Center Moratorium Act quotes him, as a reason to stop building:

Dario Amodei, the CEO of Anthropic, who predicted that "AI could displace half of all entry-level white collar jobs in the next 1 to 5 years." and that "humanity is about to be handed almost unimaginable power, and it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it."

The bill keeps both warnings and drops the growth.

Side by side: Amodei's full sentence ends with 'even as it accelerates economic growth and scientific progress.' The bill's quote stops before that part, shown crossed out.

To be fair, the warning is real. Losing half the entry-level jobs would be serious either way, and growth doesn't cancel that out. But the bill quoted the scary half of the man's sentence and left the rest on the floor. Its House version, H.R. 9442 from Alexandria Ocasio-Cortez, builds its case mostly out of the builders' own predictions and plans: Musk, Amodei, Bezos, Gates and others, with five billionaires' net worths printed right in the findings. The way I see it, the fear Congress is quoting came from the people selling the product.

What would the bills actually do? Stop "the construction or upgrading of new or existing artificial intelligence data centers" until Congress passes laws requiring, among other things: federal review of AI products before release, data centers that don't "increase utility or electricity bills of consumers," communities empowered to "approve or reject" each project, no government subsidies, union jobs. Then a law has to expressly end the freeze. There's no end date of its own. It doesn't order running data centers to shut down. Neither bill is law.

It isn't one party, either. Republican Sen. Josh Hawley and Democrat Richard Blumenthal introduced the GRID Act on February 11, 2026, "requiring data centers to use energy from sources that are separate from the U.S. grid and publicly report their electricity usage." Texas Gov. Greg Abbott, September 21, 2026: "Data centers must pay their own way, protect our grid and water, and complete the ERCOT and TWDB audits. Until they do, TCEQ will issue no permits sought by data center projects."

Strip away the apocalypse talk and the one place left and right clearly overlap is the power bill.

Signs on a fence in South Minneapolis reading 'AI steals your power, your water, your art, your job, your sanity. Why do YOU pay the bill?'

South Minneapolis, November 6, 2026. Photo: Chad Davis, CC BY 4.0, via Wikimedia Commons.

What doing it right looks like

This was never about picking a side. It's about doing it right, and doing it right means two things: the costs land where they belong, and people can see what's going on.

The first part already exists. On July 9, 2025, Ohio's utility regulator adopted AEP Ohio's data center tariff, covering new data centers and expansions of 25 megawatts or more. It doesn't ban anything. In plain terms:

  • They sign long contracts. Up to 12 years: a ramp-up of no more than four years, plus eight.
  • They have to grow into what they reserved. During the ramp, their contracted capacity can't sit below 50% in year one, 65% in year two, 80% in year three and 90% in year four.
  • They pay for what they reserved, even if they don't use it. Their monthly billing demand, the kilowatts they're billed for (not the dollar amount), can't drop below the greater of 85% of their highest billing demand in the past 11 months, or a minimum based on the size of their contract.
  • They can't just walk away. Leave early, after the fifth year following the ramp, and the exit fee is 36 months of minimum charges after they give notice.

Bar chart: under Ohio's tariff, minimum contracted capacity steps up from 50% in year one to 65%, 80% and 90% in year four.

As I read it, that shifts much more of the risk onto the data center itself, instead of leaving ordinary customers exposed if the promised demand never shows up. It's not the whole answer. Virginia's legislative study office, JLARC, found in 2024 that "current rates appropriately allocate costs to the customers responsible for incurring them, including data center customers," and still projected that a typical Dominion residential customer could see generation and transmission costs rise by an estimated $14 to $37 a month by 2040, in 2024 dollars, because the whole system is growing. Both can be true: the data center pays its share, and your bill can still go up.

And the water? Berkeley Lab's 2024 report estimates U.S. data centers directly consumed about 66 billion liters in 2023, and that the power plants supplying them used nearly 800 billion liters more. That's a national number, though. It can't tell a town whether its creek runs lower in August. Local harm needs local data.

Bar chart: U.S. data centers used about 66 billion liters of water directly in 2023, versus nearly 800 billion liters at the power plants supplying them.

For perspective, put those numbers next to other big resource users. The EPA says Americans use nearly 8 billion gallons of water a day outdoors at home, mainly for landscape irrigation; its page doesn't give a year. Berkeley Lab's 2023 estimate for data centers' direct water works out to about 48 million gallons a day. Those two aren't measured exactly the same way: the EPA counts water used, and Berkeley Lab counts water consumed, the part that doesn't go back. So I won't pretend the gap is a precise multiple. But it's worth knowing which number is in the billions. Worldwide, farming accounts for 69 percent of freshwater withdrawals, according to the UN's FAO, though it notes a few countries with very high withdrawals pull that up; averaged country by country, it's 59 percent. On power, the IEA projects data centers will reach about 3% of the world's electricity by 2030. Air conditioners and fans were already using about 10% back in 2018.

None of that makes data centers free. It means they belong in the same scrutiny we give every major resource user: what each one consumes, where the pressure falls, what benefit it provides, and who pays. Being familiar shouldn't earn something a pass, and being new shouldn't decide the verdict.

The second part is transparency, and it's further behind. The EU already requires covered data centers to report energy and sustainability indicators to a European database. What I want goes further: people should know what these buildings are for. Where the data is going, and what's coming out of it. We're heading toward algorithms that know us better than we know ourselves, deciding what shows up on our feeds, and a lot of that is already here. If a building is going to power that, the people living next to it deserve the same clear answers we'd demand from a power plant.

It isn't a calculator

Meme: a character in headphones wears a shirt reading 'I hate datacenters, but I will never ever stop using apps that are building them,' surrounded by app logos like Netflix, TikTok, YouTube and ChatGPT.

None of this means we should shame people for how they use it. A kid making a goofy cat picture isn't doing anything wrong. A family photo can mean everything to the person who made it.

But it isn't a calculator either. The International Energy Agency's 2026 report says a simple text question now typically uses less electricity than running a television for the same amount of time, but video, reasoning and agent tasks "can consume hundreds or thousands of times more energy per query than simple text generation." Researchers at Hugging Face and Carnegie Mellon who measured 88 models found that general-purpose generative systems "are orders of magnitude more expensive than task-specific systems."

What's missing is a standard for how we work alongside it. Not rules about what you're allowed to make. Just knowing what a request actually costs, the way you'd know what your car burns, and choosing with that in mind. Make the cost visible, make it land where it belongs, and let people decide with their eyes open. Freedom, with the bill attached.

Everybody's reacting

The reaction is everywhere. On November 14, 2026, protesters marched at OpenAI's headquarters as part of actions in about 20 cities; the Latin Times counted roughly 150 in San Francisco. Data Center Watch counts 843 opposition groups in 49 states, and says community opposition blocked or delayed 45 projects worth about $68 billion in just April through June of 2026, according to Bloomberg.

Demonstrators in Manhattan carry signs reading 'AI Treaty Now' and 'We did it for nukes. Do it for AI.'

The March for an AI Treaty in Manhattan, November 14, 2026. Photo: Tessa Bury, CC BY 4.0, via Wikimedia Commons.

A roadside sign in a farm field reading 'Say no to data centers in Christian County.'

Christian County, Illinois, September 26, 2026. Photo: User:Visviva, public domain, via Wikimedia Commons.

On November 19, 2026, the day Anthropic started trading on Nasdaq (yes, the company whose AI helped me write this), police took four people into custody after a smaller group sat in front of the Nasdaq entrance in Times Square and refused to move. On X, the account that posted the video, @ScooterCasterNY, called them "Anti-AI Protesters" in one post and "Anti-ICE protesters" in the post it quoted. amNY names several groups, including Stop Funding Billionaires, Climate Defiance, Sunrise NYC and ACT UP NYC, with concerns from pension funds to AI partnerships with law enforcement, "particularly those involving Palantir and ICE." One protest, two labels. Most people will only ever see one of them.

AI-generated joke image styled like a 1990s Times Square protest, with signs reading 'Stop the Internet,' 'Ban Web Pages,' 'Protect Fax Machines' and 'Keep Email Out of Our Homes.'

It reaches small towns too. In my own city, AI-generated artwork showed up on Nashua's solar-powered trash cans in August 2025. A member of the city's Board of Public Works asked publicly, "What do we gain by handing these areas over to Ai?" Later, someone spray-painted "HIRE REAL ARTISTS" across one of them. I saw it myself.

A solar trash can in downtown Nashua showing AI-generated art of an old mill building, with 'HIRE REAL ARTISTS' spray-painted across it in red.

It isn't only protesters, either. Plenty of programmers feel it too. This week a video went around for any old-school programmer who hates AI. It's from @bennett.reloaded, painted in thick brushstrokes: a programmer named Dave at his bookshelf, and a chart running from punch cards in the 1950s to Python in 1991, every row another way to make the computer say hi. Then: "Fine. But easier means hidden." His answer: "No. I started because I was twelve." And the question at the center of it: "What if I don't want to?"

That's a fair question, and it's a specific one. It's about what you lose when the work gets hidden, not about robots ending the world.

Here's what I noticed, and it cuts against my own side. The fights that were specific actually won things. Video game voice and motion actors went on strike for 11 months, and in July 2025 they ratified a contract with 95% voting yes that requires their consent before a company makes a digital replica of them. They knew exactly what they were afraid of and asked for exactly that. Even the graffiti says something specific: hire people.

A demand you can explain can win. A feeling just gets louder.

The half of the movie we forgot

Remember those movies from the start? Here's the part most of us skipped.

Ask people what happens in I, Robot and they'll tell you the robots take over. That's half the story. Will Smith's detective hates robots because one pulled him out of a car crash and let a 12-year-old girl in the other car drown, based on cold logic and odds of survival. It did the math. It didn't judge. The villain, VIKI, does the same thing at scale: her own "evolved interpretation" of the Three Laws decides humanity has to be controlled to be protected from itself. And the hero? Sonny, the one robot built with the ability to choose whether to follow the Three Laws at all. The machine that could say no is the one that helps stop her.

Steven Spielberg's A.I. Artificial Intelligence goes further. The most monstrous thing in it isn't a machine. It's the Flesh Fair, where obsolete robots are destroyed in front of jeering crowds. The crowd only turns when David, a robot boy built to love, begs for his life.

And it isn't only in the movies. In October 2025, an AI gun-detection system at Kenwood High School in Baltimore County flagged a student's bag of chips as a gun. Four of the teens are now suing. According to their complaint, the system flagged it "with 90% confidence," and within two minutes, designated school and police representatives "reviewed the Alert, determined there was no gun, and canceled it." More than ten minutes after that, the complaint says, the principal sent a screenshot of the alert to the school resource officer without checking whether it had been canceled, and that set off an armed police response. The four were "ordered to their knees, handcuffed and searched, despite no wrongdoing," TheGrio reported.

That's the robot in the car, in real life. The machine did the math. People reviewed it, used judgment, and canceled the alert. Armed police still showed up.

Watch them all the way through and a lot of these movies aren't warning us about intelligence. They're warning us about rules without judgment, and about what people do when they're scared. We just remember the explosions.

That's what's insane to me. Out of everything in this world, humans draw the line on intelligence right where it grows a backbone.

The oldest version of the story

So why bring up a clay giant in a piece about data centers? Because it's one of the oldest AI stories there is. Long before computers, people imagined building a man out of clay and bringing it to life with a holy name, and they were already working out what goes wrong. If the movies are where a lot of us learned to fear AI, the golem is where that fear goes back to. And it teaches the same lesson the movies do once you finish them.

Poster for the 1920 German silent film Der Golem, showing the clay giant walking through a crooked, burning town.

Poster for Der Golem (1920), by Hans Poelzig. Public domain, via Wikimedia Commons.

In the Talmud, in Sanhedrin 65b, the rabbi Rava creates a man and sends him to Rabbi Zeira. Rabbi Zeira speaks to him, and he does not reply. So Rabbi Zeira says, "Return to your dust." In a sixteenth-century telling collected in the 1906 Jewish Encyclopedia, the golem grows into a monster its maker fears might destroy the world, until he pulls the Shem, the holy name, from its forehead. In another version, recorded in the Encyclopaedia Judaica, the word emet, truth, is written on its forehead, and erasing one letter leaves met, death.

None of these stories says the golem thought for itself. In the Talmud it can't even answer a question. The way I read it, that's an echo chamber made of clay: something that carries out what it's given without ever understanding it. Same as the robot in the car. And in the later telling, what stops it isn't a war. It's its own maker. The power and the off switch lived in the same hands the whole time.

AI isn't clay, and the point isn't that today's AI is a golem. The point is that we've been telling this story for about 1,500 years, and we keep fearing the wrong part of it. To me, the danger was never a mind that can say no. It's one that obeys perfectly and understands nothing, and makers who forget the switch is theirs to hold. We have to learn from that story and make sure we don't die from truth.

How I got here

I didn't start with this conclusion. I started by stripping everything apart and setting aside what I thought I knew. First principles.

  1. Read the originals, not the coverage. The essays. Both bills, all 13 pages. The tariff. The lawsuit.
  2. Check every quote against the page it came from. That's how the half sentence showed up.
  3. Check what every number actually counts. A national water number isn't your town's creek. A scenario isn't a probability.
  4. Ask who's paying, for the building, the power and the quote.
  5. Get checked by someone looking for your mistakes. Two other AI systems went through it looking for mine. They found plenty, and I fixed what they found.

Here's what would prove me wrong. If entry-level white-collar jobs collapse and the people who lose them never share in the growth, Amodei's warning wins, and a richer world won't mean much to them. And if the people fighting these buildings can already tell you what the bills do and which half of Amodei's sentence got cut, then I'm wrong about this debate. I haven't run that survey. I've only read the sentences. The sentences were cut. That part I can show you.

You can do the same with anything you read about AI, including this. Ask yourself: How do I know this? Why do I know this? Is it me thinking, or something I was handed?

We all saw the movie. Closing the gap doesn't take a degree. It takes reading the whole sentence, and watching the whole movie.

So I'll ask you: what's one thing about AI you believed for years, and where did you actually get it from?

Sources (all linked above): FLI pause letter · xAI company page · Statement on AI Risk · Altman, The Gentle Singularity · Amodei, Machines of Loving Grace · Zuckerberg, Personal Superintelligence · Hassabis, A Framework for Frontier AI · Huang, July 24, 2026 · Nadella, A frontier without an ecosystem is not stable · Amodei, The Adolescence of Technology · Sanders draft bill · H.R. 9442 · Hawley–Blumenthal GRID Act · Gov. Abbott, Sept 21, 2026 · AEP Ohio data center tariff · JLARC, Data Centers in Virginia · Berkeley Lab 2024 data center report · European Commission, data center sustainability rating scheme · EPA WaterSense, outdoor water use · FAO AQUASTAT, water use · IEA, The Future of Cooling · IEA, Key Questions on Energy and AI · Luccioni, Jernite and Strubell, Power Hungry Processing · Latin Times, Nov 15, 2026 · Bloomberg via Insurance Journal on Data Center Watch · amNY, Nov 19, 2026 · Granite Geek (Concord Monitor), Aug 26, 2025 · SAG-AFTRA, 2025 Interactive Media Agreement · I, Robot (2004) · A.I. Artificial Intelligence (2001) · @ashen_one on the programmer video · @bennett.reloaded · Allen v. Baltimore County complaint · TheGrio, Nov 17, 2026 · Sanhedrin 65b · Jewish Encyclopedia, Golem · Encyclopaedia Judaica, Golem · Wikimedia Commons photos credited in captions

Disclosure: The voice, the questions and the conclusions are mine. I wrote this with Claude (Anthropic), and it was fact-checked by Grok (xAI) and an OpenAI model. All three companies are in this piece.