The Cost of Being an AI Doomer

Let's put numbers on the doom-versus-progress conversation. As of September 27, 2026, CNN's September 9 interview with former Anthropic researcher Jacob Coxon, titled Ex-Anthropic insider tells CNN how AI could kill all humans by 2030, had 9,708,414 YouTube views. Neuralink's September 18 video showing Terry, a man with ALS, using a brain implant to speak had 8,741,639.

The warning led by 966,775 views, with a nine-day head start. Two videos from different channels can't tell us what the public prefers. But how does a man with ALS speaking through an implant not have ten times the views? It's baffling to me that a prediction about the end of the world can draw even more attention than that.

This is not a cure for ALS. Neuralink's system is still experimental, not a device every person with ALS can get tomorrow. But look at what it shows. Terry, whose family struggled to understand his speech, can form words through a system that translates brain activity into a version of his own voice. What could this look like in six months, a year, two years? I don't know, but what's happening now feels almost miraculous.

The same goes for cancer. Merck and Moderna reported positive Phase 3 results for a personalized mRNA therapy, used with Keytruda, in patients whose melanoma had been surgically removed. The combination improved measures of how long patients went without recurrence or distant spread compared with Keytruda alone. The full data still need to be presented, overall survival is still being studied, and calling it a "cure for cancer" would be wrong. The actual result is exciting enough. An individualized treatment can be designed around the mutations in a patient's own tumor, with AI algorithms helping select its targets.

I don't know whether AI will be overwhelmingly net positive. Nobody does. But that's my bet, and the case for it gets stronger when you look at what the technology is already doing rather than imagining only the worst thing it might do.

Even Anthropic CEO Dario Amodei's recent warning argues for preserving AI's medical and economic upside while addressing serious risks. We can disagree about the right pace while taking both sides of that equation seriously.

We're teaching for a world that already left

I listened to an episode of The Journal about Estonia's AI experiment in schools. Estonia is putting AI tools in students' hands and trying to shape how they learn with them. The episode includes a student worried that data centers use drinking water. It also shows students frustrated with a special chatbot that asks guiding questions instead of giving direct answers.

I understand the instinct behind that design. If a kid uses a chatbot to skip every hard thought, the assignment didn't teach them much. But forcing them through a clumsy imitation of a teacher is no answer either. Give someone a bad tool, then ask why they don't see its potential, and you'll get exactly the result you engineered.

Jordi Visser deserves credit here. In Brainstorming With the Future, he argues that people learn AI by using it, and that the starting point can be as simple as a conversation. In The AI Mindset, he makes the related case for experimenting, updating your view, and judging the output instead of quitting after one bad answer. That thinking helped shape my view of what kids should be learning.

Using AI well is a skill. The interface is also getting easier. You can open an app, tap voice, and talk through a question. A kid curious about how a rocket lands can ask why it doesn't tip over, ask for a drawing, ask for a simpler explanation, then wander into engines, gravity, coding, or the history of flight. They can interrupt. They can try again. They can follow the part that actually grabs them. They get as many rounds as they want, at their own pace.

That is an extraordinary learning opportunity. Keeping kids away from it may leave them less prepared to use it, question it, and compete in a world where everyone else has access. I'd rather teach them to challenge an answer than pretend the answer machine doesn't exist.

Of course the machine can be wrong. OpenAI says hallucinations still happen. Pretending otherwise would make the argument for AI weaker. The same is true of math. Frontier systems have reached gold medal level on International Mathematical Olympiad problems and produced research-level proof attempts, but those attempts still need expert checking, and one initially promising attempt was later judged wrong. "AI is bad at math" is outdated as a blanket statement. "Check its work" is still excellent advice.

The deeper problem is school itself. If AI makes the homework trivial, what was the homework measuring? Calculators changed which calculations we do by hand. AI reaches into writing, research, coding, and problem solving, so the curriculum needs a much bigger rewrite. Students still need foundations, because you can't critique an answer you don't understand. But they also need practice asking good questions, thinking in systems, solving unfamiliar problems, and building things they care about. Curiosity can pull a student deeper than another worksheet ever will.

Look at what the upside actually means

The upside is not just a faster email. In a randomized trial of young people with diabetes, every participant offered an AI eye exam during a clinic visit completed screening, compared with 22% of those given a standard referral. That's not a chatbot replacing a doctor. It's a tool helping close a gap that can end in preventable vision loss.

That matters in a country where millions of adults carry medical debt. I want a future where someone can understand a diagnosis, prepare better questions, find a missed possibility, or get an affordable screening without starting with a giant bill. AI will not replace a scan, a laboratory test, a surgeon, or a clinician who knows what they're doing. It can make expertise easier to reach and help people use the care they do get more effectively. That's worth building toward.

And here's a medical story I bet most people reading this missed. On June 17, 2026, Midjourney, yes, the AI image company, unveiled a prototype of a full-body ultrasound scanner. It uses 40 of Butterfly Network's ultrasound-on-chip modules. You step into water, a ring of sensors sends sound waves through your body, and powerful computers reconstruct a 3D map from the measurements. The images come from real scans, not Midjourney's image generator.

Midjourney's goal is to make a scan take about a minute and put the experience in a spa. Today's prototype takes more than 20 minutes, and the first product is intended to show body composition, not diagnose disease. It has not been shown to replace MRI or screen people for cancer. MRI itself doesn't use ionizing radiation. And routine whole-body screening has not yet been shown to prolong life in people without symptoms or known risks.

Still, think about the possibility if fast, affordable scans eventually prove useful for early detection. Going to check on your health could feel more like visiting a spa than navigating a hospital. That future needs clinical evidence, but the attempt is real. Midjourney announced it three months ago. How many people heard about that compared with the latest prediction that AI will end the world?

Food is another example. Researchers are using machine learning to understand how weather, soil, and terrain affect crop yields. Their findings could inform better crop breeding and farming decisions. That is a path toward more resilient food systems, not a promise that an app will end hunger by itself. Hunger also involves poverty, distribution, conflict, and policy. But it would be absurd to dismiss a tool that could improve one part of the problem because it cannot solve every part.

And yes, driving belongs in this conversation. AI systems could prevent crashes by detecting hazards and compensating for human mistakes. In Tesla's 2025 North American data, major collision rates were 40% to 69% lower per mile with FSD (Supervised) engaged than with manually driven Teslas using active safety features, depending on road type. That's encouraging, but Tesla says its comparisons don't fully control for weather, traffic, time of day, or the situations in which drivers choose to turn FSD on. They can't prove FSD caused the entire difference.

Here's a signal outside Tesla: Lemonade charges about 50% less per mile when an eligible customer drives with FSD engaged. That's a real pricing decision from an insurer, not a 50% discount on the entire policy. Lemonade says its pricing draws on Tesla's safety data, so it is directional evidence, not an independent verdict on crashes or deaths.

How many lives has FSD actually saved? I can't give you a defensible number from those collision rates. Federal officials estimate 36,640 people died on U.S. roads in 2025. If systems like FSD eventually cut that toll by even 10%, that would be roughly 3,700 lives a year. That's a scenario to show the stakes, not a claim about lives FSD has already saved. Regulators are also investigating specific failures. We need stronger independent safety evidence while we keep improving the technology.

The data center question deserves an adult answer

Every one of these possibilities needs computing infrastructure. That doesn't mean every proposed data center is a good deal for the town next door.

Submer's breakdown of data center water myths gets something right: "How much water does AI use?" is too vague a question to answer with one viral number. Google estimated that a median Gemini text prompt used 0.26 milliliters of water for on-site cooling in May 2025. That's one company's estimate for one kind of prompt, not a measurement of every AI conversation or the water used to generate its electricity.

Look at the larger scale too. Berkeley Lab estimated that all U.S. data centers consumed about 17 billion gallons on site in 2023. U.S. golf facilities applied 1.63 million acre-feet for irrigation in 2024, about 531 billion gallons. The California Water Impact Network estimates that almond orchards use 4.7 to 5.5 million acre-feet annually, about 1.5 to 1.8 trillion gallons. Familiar uses of water rarely get framed as proof that an entire industry should not exist. Why does AI?

Those figures need their labels. The golf number is irrigation water applied, the almond number is an estimate built from crop acreage and assumed water needs, and the data center number is water consumed on site. They also cover different years. Berkeley Lab estimated another 211 billion gallons consumed indirectly to generate data centers' electricity. Water use varies by cooling design, climate, and power supply. The national comparison puts the scale in perspective, but it can't tell a town whether one proposed facility will strain its water supply. Ask for the local numbers, the water source, and a plan that holds up in a drought.

Power demand is real too. The International Energy Agency expects global data center electricity use to roughly double between 2025 and 2030. And if a utility builds expensive infrastructure for one huge customer, residents should not be stuck with the bill. The U.S. Department of Energy explicitly identifies cost shifting as a problem that rate design needs to prevent.

There is real evidence that data centers can help balance the grid when they're designed and contracted to respond to it. Google says it has reduced data center electricity demand during grid stress in Oregon, Nebraska, and the Southeast by shifting non-urgent computing work. It has signed utility agreements for a total of 1 gigawatt of demand-response capacity, although contracted capacity is not the same thing as power actually shed. In a separate Phoenix field demonstration, a 256-GPU AI cluster cut its power draw 25% for three hours while maintaining its workloads. That's a cluster-level test, not proof that an entire facility stabilized a grid. But it shows how a big new electricity user could also become a flexible one. The IEA says on-site batteries offer another way to make some facilities useful grid assets. None of it happens automatically. The operator and utility have to build it into the deal.

There are benefits, but let's not inflate those either. Construction brings work and a facility can generate tax revenue. Brookings researchers estimate roughly 100 to 200 local jobs in a typical county that receives its first large data center, depending on the facility. That is meaningful. It is not an automatic economic miracle. If a developer wants land, power, and water, the community should get transparent numbers, fair rates, enforceable commitments, and a clear account of the tradeoffs.

That brings me to the Kentucky family that turned down a $26 million offer for part of its farm. The land is theirs. They have every right to say no, and their concern about losing farmland isn't stupid. I still can't stop thinking about the choice. A data center could bring new jobs to the county, though it doesn't follow that the farm supports only the family or that every promised job would materialize.

Then there's the money. The family's taxes, costs, and plans are private, so I won't pretend to know their take-home amount. But AI could help anyone explore the decision. Suppose a sale left $10 million to invest after taxes and after setting aside enough to live well. What if they put it in a fund tracking the S&P 500? Historical data from 1928 through 2025, including reinvested dividends, work out to about a 10% annual compounded return. Repeat that rate for 30 years and $10 million grows to about $175 million. Give it 40 years and it's about $453 million. Those are future dollars before inflation, investment taxes, fees, or withdrawals. Ten percent is a historical average, not a conservative promise about the next 40 years. Change the assumptions and the answer changes. That's the point of running the numbers, not a prediction of what this family should do.

What do we build for across generations? Is the point to make sure our great-grandkids inherit the same work, or to give them more security and more options than we had? They might value keeping this particular land above any financial return. I don't know their answer. I do think fear of AI can keep the rest of us from asking the question honestly.

The reflexive doom take counts every cost of building and almost none of the cost of standing still. That's a bad way to make decisions. We should scrutinize the costs, negotiate hard, and fix what doesn't work. A world with better treatments, more accessible learning, earlier diagnosis, stronger food systems, and safer roads is worth pursuing. The risks deserve a serious response. So does the possibility that refusing to build will carry its own enormous cost.