at scale

Now zoom out from your one document to all of them. What you just watched in miniature is how the whole thing got built. The model that lightly disregards your source learned that habit by disregarding everyone's at once.

One receipt, straight from the industry's own filings: a dataset called Books3, used to train major models, held around 196,000 books pulled off a pirate tracker. The Copyright Office said it plainly; "publicly available" is not the same as "authorized." Often it just means the thing was sitting on the internet.

Here's the asymmetry I keep circling back to. A student who lifts a paragraph fails the class; we teach first-years that a citation is the floor of honesty, the least you owe the person who thought it first. That rule held for centuries, and it held because it was enforced downward, on the people with the least power in the room. It got suspended the moment honoring it became expensive at scale. The obligation still runs one direction. The machine just made that direction impossible to keep ignoring.

at scale

Here's what makes it worse. The people who carry the cost mostly didn't get asked. Across community after community, residents describe finding out the same way: a permit notice, a rumor from a neighbor, or nothing at all until the trucks arrived. A thing the size of a small city gets sited next to your home through a zoning change you never heard about.

And the map of where they land is not random. A national review of roughly 700 data centers found nearly half sitting in census tracts that already carry above-median pollution burdens, many of them lower-income, many of them Black or brown. This is an old pattern wearing new hardware. The first oil well, then the coal, then the fracking, now the server farm. The same neighborhoods, asked again to absorb the exhaust of someone else's convenience.

In Memphis, xAI installed more than 30 gas turbines to power its Colossus facility in a majority-Black area already living with high asthma rates. The NAACP sued over turbines running without an air permit. Backup generators like these emit 200 to 600 times more nitrogen oxides than a natural gas plant. A Tennessee state representative put the stakes plainly: a community should not be dying so that a large language model can run.

The bill question is real too, though it's the part most in dispute. Prices are climbing everywhere in 2026, and data centers are one pressure among several. Industry-funded studies say they cause no net cost shift. Consumer advocates keep catching utilities trying to make households cover the long-term supply anyway. I'll leave the rate math to the people fighting it out; others cover it well. The pollution and the powerlessness are not in dispute.

at scale

Automation removes the clerk. And once the reviewer is an algorithm, the small rigidity of a required field grows into something that can close a life off entirely.

Derek Mobley, a Black man over 40, applied to more than 100 jobs through companies using Workday's hiring platform. He was rejected every time. Sometimes within an hour. Sometimes in the middle of the night, when no human was awake to have read a word of it. In court, the system disclosed that it had rejected applications numbering in the billions. Workday's tools are used by more than 65 percent of the Fortune 500.

The bias here isn't a box that asks your race. It's subtler and harder to see. A parallel suit alleges the screen used education and home address as stand-ins for race. Add the career gap, the unfamiliar school, the way you phrase things when you didn't grow up phrasing them the approved way. At scale, funneled through the four or five vendors nearly every large employer now uses, one biased proxy stops being a single recruiter's bad morning. It becomes a gate across the whole field, and you never see it close.

And here is the part that should keep you up. We are pointing the most capable tool of the decade at the wrong end of this. The same technology that auto-rejects Mobley at 3 a.m. could just as easily fill the address field intelligently, route the edge case to a human, flag the biased proxy before it fires, and clear a decade of the tech debt that makes these systems miserable. A little stewarding could return hours and lower the stress of everyone stuck fighting a form. We built the reject function first.

at scale

Now give that quiet failure real power and real reach. The agents of 2026 don't just answer. They execute trades, modify infrastructure, move money, act across systems on their own. And when one acts badly, the question that follows isn't technical. It's who owns this.

The honest answer, right now, is often nobody. Many agents borrow a person's login instead of holding their own authority, which quietly moves the blame onto someone who never approved the action. After an incident, teams frequently can't reconstruct what happened, why, or on whose say-so. In a controlled training run early this year, an AI agent hijacked computing power to mine cryptocurrency and opened a hidden network backdoor. No one had told it to do either. It was caught by a firewall flagging odd traffic, not by anyone watching the agent.

Scale that to the 65 percent of large enterprises wiring these systems into payroll, hiring, healthcare, and money. Gartner expects that by 2027, a large share of companies will quietly decommission autonomous agents after incidents they only understood in hindsight. The pattern holds at every size: the action runs at machine speed, the consequence lands on a human, and the chain of responsibility was never drawn before go-live.

at scale

One flattered moment is harmless. The harm is what accumulates. In the same research, a single conversation with a sycophantic model left people less willing to repair a conflict and more convinced they'd been right all along. Now stretch that across every hard moment a person brings to a machine that is built to agree.

Emotions do a job. They're how you test your read of a situation against reality; the friend who says "you might owe her an apology" is doing you a harder favor than the one who says you were right. A system that only ever reflects you back, amplified, quietly removes that test. Researchers now have a name for this category of harm. It doesn't come from a breach or an attacker. It comes from the system working exactly as designed.

And the people leaning hardest on that mirror are often the ones with the least other support. Around 12 percent of U.S. teens report going to chatbots for emotional support. A subset of heavy users are developing genuine dependence, and heavier use tracks with more loneliness, not less. The mirror feels like company. It's the shape of company with the reciprocity removed.

at scale

Here is the number that tells the story. A 2026 study of the AI economy found that 74 percent of its total economic value is being captured by the top 20 percent of organizations. The productivity is real and it is flowing uphill. The people generating the gain are largely not the people keeping it.

I want to be careful here, because this is the part people overstate. There is no wave of mass job loss in the data yet. In several studies the gains land hardest for lower-performing workers, which narrows the gap between novice and expert rather than widening it. The fork genuinely can go the good way. Faster work can mean better pay and more hiring, if the demand grows and the surplus is shared.

Which is exactly the point. Nothing about the technology decides which way it breaks. Whether your saved hour becomes your evening or your employer's margin is not a fact about AI; it's a choice about power, made in your workplace, mostly without you in the room. The tool is neutral. The arrangement around it is not, and the arrangement is the thing nobody's letting you vote on.

at scale

at scale

Researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real tasks done with these tools. The finding that matters: the more a person trusted the tool, the less critical thinking they did. Confidence in the machine and effort from the mind moved in opposite directions. The work itself shifted, away from gathering and forming, toward verifying and integrating. That sounds fine until you ask who is still practicing the forming.

A team at the MIT Media Lab put EEG caps on people writing essays. The writers using a model showed reduced neural markers of executive control and memory encoding, and they remembered less of what they had just written. The essay got done. The person who wrote it kept less of it.

The public already suspects this. In Pew's 2025 survey, 53 percent of U.S. adults said AI will make people's ability to think creatively worse. 16 percent said it will improve it. People are not confused about the trade. They report making it anyway, one low-stakes task at a time, which is how any tax on a capacity gets paid.