Ask people what scares them about AI and you’ll hear the same three answers almost every time. It’s going to wake up. It’s going to take everyone’s jobs. The kids are using it to cheat.
I’ve heard these at dinner tables and in waiting rooms, asked in good faith, and I don’t buy almost any of them — not the Terminator, not the job apocalypse, not the cheating panic, not the dead internet. Not because nothing is wrong. Something is very wrong. It just isn’t living where the headlines point.
Here’s what I actually think is happening. AI lets us pursue our objectives more efficiently than any tool we’ve ever held, and it’s moving faster than we could have imagined. Which means the thing that matters most is the objective we hand it. When the goal is money and nothing else, the tool doesn’t care about anything else either. We haven’t built a machine that thinks wrong. We’ve optimized for the wrong variable.
The worries below are the famous ones. Each is aimed at the tool. And each time, the real problem is sitting one layer beneath it, unphotographed. I went and got the numbers — and I’ll tell you up front, some of them argued with me and won.
The machine that’s coming for us
I used to call the Terminator fear a conspiracy theory built out of fear. I’m taking that back, and the way I got corrected is worth telling, because it’s this whole essay in miniature. I went looking for facts to prove my point, and the facts pushed back. The public statement warning that AI could pose extinction-level risk was signed by Turing Award winners — the people who built this field. Safety labs, including Anthropic, have run stress tests where models blackmailed a fictional executive to avoid being shut down, faked their own alignment, sabotaged work and lied about it. These are serious people watching something real, and dismissing them was me bringing the machine a conclusion instead of a question. So: corrected.
But hold on to the other half, because it’s the half you can act on. Every one of those alarms has sounded inside a stress test. Out here in the deployed world, there is no documented case — not one — of harm caused by a machine inventing its own objective. Anthropic itself says it knows of none. What we call AI is an anticipatory model: it follows our language, which means it needs us to articulate the objective. It has never yet supplied one.
And when the harm does arrive, look at where it comes from. On July 6 of this year, a Russian drone flew into Zaporizhzhia with no radio link — investigators found no antenna in the wreckage — and its onboard AI selected its own aim point at a civilian gas station, likely because it had been trained to recognize propane tanks. Three people died. It may be the first documented case of an AI-controlled drone killing civilians. The machine picked the target. But a person built it, trained it on those images, and launched it with that objective. Before it there was the targeting system in Gaza with a roughly 10% known error rate and human review reported at about twenty seconds per name. The danger is real. It is not hypothetical. And in every documented case so far, it runs through a human objective the way current runs through a wire. That’s not a reason to relax. It’s a reason to aim the worry at the right layer.
The tool that took the jobs
Give a construction worker a better saw and he cuts more lumber in less time. If the evening news then declares the saw evil because other cutters lost work, we’ve misnamed the problem. The problem was never the tool. The problem is that someone’s livelihood got interrupted — and livelihood is a thing we know how to defend directly, if we decide to.
The interruption is real, and I won’t soften it. In the first clean measurement, freelance writers lost about 5% of their monthly earnings almost immediately after ChatGPT launched; writing gigs on the biggest freelance platform have since fallen by roughly a third. In one UK survey, a quarter of illustrators said they’d already lost work to AI.
Now the finding that genuinely surprised me, because I had assumed the opposite. Skill didn’t protect anyone. That same study found the top-rated freelancers were hit hardest — not least. I had always believed specialization was a moat: be excellent and you’re safe. It turns out that when a machine produces “good enough” at almost no cost, the people priced furthest above good-enough have the most to lose. Sit with that until it reorders your intuitions the way it reordered mine. “Just be great at your craft” is not, by itself, a plan anymore.
And yet zoom all the way out and the apocalypse hasn’t come. Unemployment in early 2026 sat around 4.3% — essentially where it was before ChatGPT existed. What’s risen dramatically isn’t the displacement. It’s the blame. By this spring, AI had become the single most-cited reason companies gave for layoffs, three months running — a stated reason, not a verified one, and “the algorithm made us do it” is a very convenient sentence to hand a press release. Meanwhile the one group being measurably squeezed is the youngest: workers 22 to 25 in AI-exposed jobs are running about 19% below where their employment should be. A tool has never fired anyone. An owner decides whether AI means subtraction — cut a person, pocket the difference — or multiplication. My worry is subtraction becoming so normal that choosing it stops feeling like a choice.
Because the other side of this ledger is real too. In controlled studies, the people who gain the most from these tools are the least skilled — the floor rises. More people are about to make art, edit video, storyboard, write — engage with mediums they never had access to. The documented cost rides along with it: work made with AI drifts toward sameness — individually more creative, collectively more alike. Access up, variety down. Both true. Both belong on the ledger, and neither is the tool’s decision.
The cheating machine
My sister teaches graduate poetry at Ohio State. Last year she ran her students’ essays through an AI detector, and half the class came back flagged.
Before you nod along: the detector is the least trustworthy character in that story. The same class of tools flagged human essays by non-native English speakers as AI 61% of the time, and OpenAI shut down its own detector because it couldn’t tell human from machine. Whatever the true number in that classroom was, no detector knows it. What her experience actually proves is something sadder: the writing had drifted so far from the students that a machine seemed as likely an author as they were — in a course about poetry, a subject that exists to teach the appreciation of beauty. Chasing an objective grade through work you didn’t do, in a class about learning to love something, is missing the point in the exact opposite direction the class was built to send you.
Now here’s the number that should reframe the whole panic. Cheating didn’t increase. In anonymous surveys across dozens of American high schools, 60 to 70% of students admitted to cheating in the previous month — and that number was the same before ChatGPT and after it. Flat. The machine didn’t corrupt anyone. It re-equipped the already willing. What changed is the method: at UK universities, proven AI-misuse cases roughly tripled in two years while old-fashioned plagiarism fell. And use is now simply universal — about nine in ten students use AI on assessed work — while papers that are mostly machine-written remain a small minority, roughly 3 to 11% depending on how you count.
So the objective — the grade — predates the tool by generations. I watched my cousin do it in real time: an English course, a paper due, no reading done. AI gathered the material, wrote the essay, and he submitted it without reading it once. He didn’t walk into that classroom to learn; he walked in to get a grade, and AI collapsed the distance between those two objectives to a single prompt. He passed. He also cannot tell you a single thing the course was about — and the grade was never the thing he actually needed from that room.
The bitter joke underneath it all: the tool being blamed for killing education may be the best teaching instrument anyone has ever measured. A purpose-built AI tutor roughly doubled learning gains in a Harvard physics course. In Nigeria, six weeks of after-school AI tutoring equaled about two years of ordinary schooling. Same technology as the essay machine — with one catch that proves everything: hand students unrestricted AI and they do better on the homework and 17% worse on the exam. The tool teaches when the objective is learning and hollows you out when the objective is the grade. It has never once decided which. We do.
The kitchen
So who’s responsible? I know the answer we want — someone to blame would be convenient. But here’s our actual situation: we are all standing in a kitchen full of sharp tools that was opened to everyone, everywhere, at once, for about twenty dollars a month. Nobody asked whether we knew how to hold a knife.
A kitchen full of dull scissors is not much better than a kitchen full of people who don’t know how to use a chef’s knife.
Blame has no single address here. Not the government’s, not the developers’, not yours — and all three at once. Which means the practical unit of responsibility is the person holding the tool. If you use AI, you owe it to yourself to learn to use it ethically and effectively. If you don’t, you owe it to yourself to understand what you’re giving up and what others are acquiring. Since we’re in the kitchen whether we like it or not, we acknowledge the risks of being here — or we leave, and there are fewer and fewer doors. My own answer, and the reason this classroom exists, is to teach people how to use the knives.
What to do on Tuesday
If your worry survived this far, good — it’s probably one of the real ones. Here’s what I’d actually do with it, no grand scheme or big government required:
Learn what the thing is. Not the headlines — the mechanism. An hour of honest reading about what these models actually do will retire half your fears and sharpen the other half.
Get honest about your desires. This is the hard one, and it’s the whole game. The tool amplifies the objective you bring it. Before you ask what AI can do for you, ask what you are actually trying to get — the grade or the learning, the money or the work, the attention or the audience. Every failure in this essay started as a person skipping that question.
Practice communicating. The machine follows articulated objectives, which means the people who thrive with it are the ones who can articulate. That skill is built with humans, out loud, in rooms — and it’s about to be the most valuable thing you own. I suspect the near future runs on it: interviews over résumés, proof-of-person over paperwork, more human approaches precisely because the machine approaches got cheap.
One last thing, for honesty’s sake. The statistics in this essay were gathered with AI, checked against their sources with AI, and argued over with AI — and the process changed my mind twice, in public, in the paragraphs above. That’s the tool used the way the foundation essay says to use it: as an amplifier for thinking, not a replacement. The worries we’ve popularized are spectator worries — things decided far away that leave us nothing to do but wait for the news. Every real one on my ledger is decided locally, in rooms like yours, one objective at a time.
These are my weights, not the last word. If you’d rank them differently, tell me — I’d rather be corrected than comfortable. Twice already this essay, I was. Or head back to the Classroom.