AI may be the most powerful tool we’ve ever built. That doesn’t mean we should surrender our judgment to it—or our judgment about it.
I heard a story last week about two architects who walked into a client’s office and were handed a set of renderings. The client put them on the table and said, “I’d like you to build this.”
One architect looked at his partner and said, “I thought we did the renderings.”
His partner looked back at him and said, “I guess we’re working for AI now.”
It’s funny, of course, because there is enough truth in it to make it funny. But the more I thought about the story, the less I thought it was really about architects, renderings or even AI. Buried inside that little exchange is what may be one of the essential questions of our time:
Who works for whom?
Apparently, we’re supposed to be very worried about the answer because, in the span of only a few days, I encountered headlines telling me to Pause AI for Humanity’s Sake, informing me that AI Is Powerful Enough to Crack Our Hardest Math Problems—and Kill Us All, and finally announcing that America’s Great AI Freakout Has Begun.
That last headline seems particularly accurate. The freakout has most assuredly begun.
So I read the articles, because dismissing something you haven’t read is no better than believing something you haven’t examined, and I discovered that there is plenty of evidence that artificial intelligence is becoming astonishingly capable, progressing more quickly than many people expected and doing things that deserve our attention.
What I found considerably less evidence for was that it is going to kill me.
I’m Not Saying There’s Nothing to Worry About
AI is doing things that even some of the people working in the field apparently didn’t expect it to do this soon. The Wall Street Journal, for example, reported on AI systems attacking extraordinarily difficult mathematical problems that human beings have wrestled with for decades, with the accompanying performance chart showing just how quickly capability has been advancing.
There have also been experiments in which AI agents behaved in unexpected ways, cooperated with other agents and took actions their designers had not specifically mapped out for them. Peggy Noonan points to one such incident involving Hugging Face, along with the warnings of AI researchers who themselves are deeply concerned about where these developments might lead.
Those aren’t things I think we should ignore. Quite the opposite. Put safeguards around dangerous capabilities, test those safeguards, try to break them, find out what happens when they fail and then make them better. If there are things AI shouldn’t be allowed to do autonomously, don’t allow it to do them autonomously.
But somewhere in these discussions something peculiar happens, because an observation about what AI has done gradually turns into a prediction about what AI will do, and the difference between those two things starts disappearing.
AI behaved unexpectedly, therefore AI may become uncontrollable; AI may become uncontrollable, therefore it may learn to improve itself; if it can improve itself, it may gain access to critical infrastructure or other powerful systems; if it gains that access, it may become capable of destroying civilization; and if it becomes capable of destroying civilization, it may actually do it.
Therefore, we’re all dead.
Wait a minute.
That’s not one proposition. It’s a chain of propositions, and every link in that chain requires evidence before I am required to accept the conclusion at the end of it.
Possibility Isn’t Probability
I had an interesting exchange with a friend in the UK about this, and she offered what I thought was a very good analogy. She said the situation was like someone quitting a bicycle factory because the company wouldn’t put brakes on the bicycles and you wanted no part of what they were doing.
I’d quit, too.
But there is a problem hidden inside the analogy: first show me that the bicycle doesn’t have brakes.
If the brakes don’t work, fix them. If we discover that the bicycle cannot be made safe, don’t ride it. If somebody deliberately removes the brakes because they want the bicycle to go faster, stop them.
But don’t say that because the bicycle is becoming extraordinarily fast, it therefore has no brakes, because the thing you are using as an analogy is actually the very thing we are arguing about.
That’s the part of the current AI discussion I can’t get comfortable with, and not because I know AI is safe.
I don’t.
Neither, apparently, does anyone else, which seems to me the point. Uncertainty works in both directions, and the fact that I cannot prove something won’t happen isn’t evidence that it will.
Curiously enough, the Journal’s own reporting on America’s “AI freakout” acknowledges that the view that AI will lead to human extinction remains a fringe opinion among industry experts, while noting that even some experts who believe such an outcome is possible consider it a low-probability event and believe actions can be taken to reduce the risk.
That doesn’t prove those people are right either.
It tells me something simpler: we don’t know. I thought admitting that was part of thinking.
We’ve Been Here Before, But Not Quite
Whenever someone compares AI to an earlier technological revolution, someone else immediately points out that AI is different.
Of course it is.
The printing press was different from what came before it. The steam engine was different. Electricity was different. The automobile was different. The Internet was different. Google was different. And artificial intelligence is different again.
History doesn’t prove that AI is safe, because the fact that humanity survived the automobile tells us absolutely nothing about whether an advanced AI system could someday become dangerous. But history tells us quite a bit about human beings, including how badly we tend to understand enormous technological changes while we’re standing in the middle of them. And that includes ourselves, doesn’t it?
The Internet destroyed businesses and created businesses. It eliminated jobs and created jobs nobody had names for before. It changed publishing, advertising, retailing, journalism, music, banking, research, communications and just about everything else it touched, while leaving an enormous number of people wondering what exactly was happening to the world they understood.
They adapted, not because the disruption wasn’t real, but because it was.
What nobody could see particularly well while it was happening was what would exist on the other side of it. And we’re doing that again. Only faster. Much faster.
What Actually Makes Me Pause
Here’s what you really have to worry about: AI becoming intelligent carries with it the same possibility all technologies carry—that human beings will decide that because AI is intelligent, they no longer have to be.
You already see that happening.
Sarah Tucker wrote an interesting essay called The High Cost of Thinking for Free, which isn’t really about AI apocalypse at all but about human value and the importance of understanding what we contribute when traditional professional categories begin changing. She also makes an important observation about intelligence itself: highly intelligent people and organizations can remain remarkably blind to assumptions because intelligence by itself doesn’t guarantee judgment.
Which takes me back to those two architects and the rendering sitting on the client’s desk.
A rendering isn’t a building, and it isn’t architecture. Somebody still has to determine whether the thing can stand up, whether it can be built, whether it meets code, where the mechanical equipment goes, what it costs, what happens when water hits it, whether people can actually use it and whether the finished building will do what the person who dreamed it up thinks it will do.
Those are judgment questions, and AI may have allowed the client to arrive at the architect’s office with something he couldn’t easily produce before, which is certainly significant.
But perhaps AI didn’t replace the architect. Maybe it just gave the client a pencil.
If so, the architect has a different problem. He has to explain more clearly what an architect is actually for. That’s disruption. It isn’t extinction.
I Trust AI as Far as I can Throw It
I use AI every day, and I trust it. Kinda.
I kinda trust Google when it puts something at the top of a search. I kinda trust things I read in the media because someone was willing to print them. I kinda trust an author because somebody was willing to put his work between two covers, and I kinda trust an expert because someone put three letters after the name.
But that’s just it: kinda. Somewhat. Slightly.
I’ve spent a lifetime interviewing people, researching markets, writing, teaching, reading and running a business, and if that has taught me anything, it is that information and judgment aren’t the same thing.
Why should AI suddenly relieve me of the responsibility to know the difference?
I argue with AI. I tell it when I think it’s wrong and ask it to explain why it thinks I’m wrong. I ask it to prove things, throw away things it gives me, rewrite things, challenge assumptions and occasionally discover that it has seen something I haven’t.
Sometimes it changes my mind. Sometimes I change its answer. Sometimes neither one of us moves very far. That’s not working for AI. That’s using it.
And if we lose that distinction of “kinda,” I think we have a much more immediate problem than some hypothetical superintelligence plotting our extinction. We will have voluntarily surrendered the very thing we’re supposedly afraid AI will someday take away from us. Our “us.”
Seventeen Thousand Five Hundred Pages Aren’t a Person
One of the more fascinating articles I encountered during this little AI immersion wasn’t frightening at all. It concerned a man named Brian Dema who wanted to create an AI version of his dead father, Richard, and who happened to have an extraordinary amount of material from which to do it.
His father had left behind 17,500 pages of work-related documents. Dema fed them into an AI system, supplemented them with recordings of his father’s voice, involved a psychologist and gathered information from people who had actually known him.
Eventually they created “RichieBot.” And it wasn’t Richie.
People who knew Richard Dema noticed that the bot seemed more extroverted than the man had been, and his son noticed something even more revealing: it didn’t have his father’s distinctive humor.
Think about that for a moment.
Seventeen thousand five hundred pages of information about a human being still weren’t the essence of a human being.
I don’t see that as an argument against AI. I see it as an argument for humility, because AI can know an astonishing amount about us without becoming us, just as it can know an astonishing amount about a problem without necessarily possessing the judgment required to decide what ought to be done about that problem.
Neither do we, sometimes.
Which is why we have to keep thinking. Kinda.
Evidence at Hand
This may be where I part company with people who are convinced AI is going to kill us, because I’m convinced it won’t.
How could I possibly know that?
What I do believe is that we can only make judgments based on the evidence at hand, accumulate additional evidence as time marches on and then have enough intellectual honesty to change the judgment when the evidence changes.
If AI systems demonstrate dangerous autonomous capabilities, increase the safeguards. If a safeguard doesn’t work, replace it. If a particular capability cannot be deployed safely, don’t deploy it. And if we eventually acquire convincing evidence that we have created something we cannot control and that poses an intolerable threat to human beings, we’d better have enough courage to stop.
That’s not the courage I’m particularly worried about.
I’m worried about whether we’ll have enough courage to admit that something frightening might not happen, because possibility and probability aren’t the same thing, uncertainty isn’t evidence, and saying “I don’t know” is sometimes the most intelligent conclusion available when everyone around you is shouting that they do.
Which takes me back to those two architects.
The client puts an AI-generated rendering on the table and says, “I’d like you to build this.”
Who works for whom?
I suspect we’re going to be asking that question for a long time, because it isn’t really a question about renderings, architecture or even artificial intelligence. It’s a question about whether human beings intend to retain responsibility for judgment when increasingly powerful tools can perform more and more of the intellectual work that used to precede that judgment.
For my part, I’ll continue asking AI questions, arguing with it, correcting it, learning from it and, yes, occasionally discovering that it has corrected me.
But the judgment remains mine.
I don’t work for AI. It works for me. Kinda.
___________________________________________________
Works Consulted
Cohen, Ben. “AI Is Powerful Enough to Crack Our Hardest Math Problems—and Kill Us All.” The Wall Street Journal, September 11, 2026.
Jargon, Julie. “A Dead Father, Reincarnated With AI: Is This How We Will Remember Loved Ones Now?” The Wall Street Journal, September 12, 2026.
Noonan, Peggy. “Pause AI for Humanity’s Sake.” The Wall Street Journal, September 10, 2026.
Ramkumar, Amrith, Angel Au-Yeung, and Lindsay Ellis. “America’s Great AI Freakout Has Begun.” The Wall Street Journal, September 11, 2026.
