Mind and Iron: What We’re All Getting Wrong About AI

Hi and welcome back to another spicy episode of Mind and Iron. I’m Steven Zeitchik, veteran of The Washington Post and Los Angeles Times, senior editor of tech and politics at The Hollywood Reporter and journalistic OG Anunoby of this tech-journalism NBA Finals (at least, that’s the aspiration).

Every week we try to make sense of this dizzying orb of technology and AI hurtling toward us. Join our rocket ship.

This week I wanted to take a pause from the hurly-burly of the news to pull back and look at a big-picture issue, namely: when exactly should we use AI? Or, better, are we lumping too many use cases together and not drawing sharp enough distinctions between AI that could be very very good and AI that could be very very bad?

The answer is yes (or I wouldn’t have phrased it that way). In doing so we’re not just making it easier for tech companies that want to impose a dicey agenda to neutralize opposition with the Luddite label but also, in the other direction, we’re missing out on opportunities to benefit from AI by dragging the good down with the bad.

It’s a rich topic, and I’d like to offer what I hope is a handy and reliable way to separate between the categories, with an assist from South Park (!)

Since we’re riffing in this issue, I’ll forgo the quote of the week, but I hope you’ll find some good lines in the piece below. We all desperately need some lampposts on this dark twisty road ahead of us. Here’s hoping this provides just such high wattage, or at least a cluster of fireflies.

Let’s get to the messy business of building the future.

The Matt Stone Test

1. EVER SINCE I INTERVIEWED SOUTH PARK CREATOR MATT STONE THIS SPRING, I’VE BEEN THINKING A LOT ABOUT ONE LINE HE SAID.

Stone and his creative partner Trey Parker have been running a boutique AI studio called Deep Voodoo for going on seven years now, and it’s produced some of the coolest, most interesting and, to use a technical term, less slop-y creations out there. When I asked his larger philosophy about when AI and AI should and shouldn’t be used, Stone said this:

“I find that a lot of discussions about AI become tiresome. You know, ‘Put your taxes in and it can do them. And it’s like, ‘Cool, but a human can do your taxes.’ What we’re trying to do is something no amount of humans can do.”

Even when Stone said the line it made an impression, and I’ve found myself unable to forget it since, returning to it again and again in recent weeks. The creator was referring primarily to video generations, of course, articulating a philosophy about what kind of work the company would devote its energies to. Having Kendrick shapeshift, Donald Trump deepfaked or Bill Clinton de-aged are all the kinds of entertainment twists you couldn’t pull off with all the makeup in the world, so they were Deep Voodoo projects; creating digital stand-ins for locations or reconstructing a historical period are time-honored conventions that will work just fine if we have humans do them, so they weren’t.

But Stone’s comments turn out to be a surprisingly durable rule across the board, sliding easily in to so many realms: into scientific research, politics, education, medicine, legal work (and, yes, taxes and finances). In some cases we are simply trying to make some efficiency gain or not even that, just save a little money by dropping a machine into where humans were doing just fine. In such cases, Stone is saying he’s bored or even turned off by AI.

But drop it in to places where no human could possibly compete — in modeling drug experiments that would take decades and can now be done in months, saving countless lives, or in climate prediction where AI can tell us where a hurricane might strike and thus how to spend critical conservation resources — and now you have him excited. I use grandiose examples, but this can come down to something as granular as a risk for an individual patient or whether a hurricane will strike your house. In all of these cases, the model isn’t replacing the human — it’s one-upping the human. And we’d be silly not to take advantage.

Call it the Matt Stone Test: are we just trying to automate what’s already there or iterate what isn’t? Is AI an efficiency tool or an innovation? Tech companies like to say they’re all about “innovation,” but much of this doesn’t feel that way at all. Coding and graphic design, for instance, are two tasks humans do pretty well — we’ve had a lot of practice. Yet plenty of AI effort is channeled in these directions.

I’ve sometimes trotted out a distinction between professional use cases (eg, scientific discoveries) and consumer utilization, and that can map onto the Stone Test, but not always, and I think this is more exact. When it comes to the dividing line on whether AI is worth the cost — and there is a cost, in jobs, in the environment, in the general outsourcing of human thought and toil — it seems like whether an application can pass the Stone Test might be the best metric we can muster.

At its most elemental AI amounts to the scary-sounding but surprisingly simple idea: We shouldn’t be scared of what AI can do that we can’t. We should be annoyed when it’s trying to do what we can.

Now, obviously there are cases where the efficiency has a qualitative effect on the bottom line. Could the ability to shoot dozens of historical reconstructions make for better TV shows simply because we’ll have so many more? Perhaps. Even AI drug discovery works on this principal — the more you can look and the more quickly, the better the odds of finding cures. I recognize that the line between the same task done quicker and a new task done different isn’t always clear. Still it strikes me that enough applications fall clearly on one side of the line that we can use it as a determining rule.

What’s funny to me is that AI companies were once themselves all about this principle, and now either seem to be forgetting their roots or even actively muddling them.

The original mission for OpenAI was exactly this kind of sweeping, change-the-world philosophy. (You can still see the residue of this in Sam Altman’s pronouncements about super-intelligence and all the positive change it will bring to the world.) But somewhere along the way — possibly around the time they decided to become profitable — the company pivoted and entered the accountant-replacement business, which makes those Altman pronouncements seem either silly or hypocritical. The company became about selling the quickest tool to make our lives marginally improved for the sake of improve their margins. Yet was claiming otherwise.

Of course it’s not just OpenAI. Google DeepMind was all about moonshots and big ideas; now it’s a lot more about Gemini improving your ability to search. Anthropic wanted a more ethical version of OpenAI’s sweeping-hood; now it’s reveling in how well Claude can help you code. These are all cases of humans-doing-taxes. We have a system and it works, and instead of AI applied to new systems we’re just turning it loose on the host organism — on what humans already do well.

The problem isn’t just mission but messaging, or perhaps the messaging necessity. For years AI companies felt so pushed against the wall by critics saying they were trying to replace humans that they began underselling — they began saying they were “just a tool,” no different from a hammer or nail. But this was dumb, and anyone with their eyes half- open could see it was dumb, because a hammer and a nail don’t autonomously walk over and fix your dresser. What was more accurate — what they should have done to reassure people — is go the OTHER way, and say this was not a hammer or a nail at all, this was something entirely less recognizable and more magical. Had they focused on that idea — and showed how we really couldn’t live without these technologies as opposed to how these technologies make us feel like life is not worth living — we would be both infinitely better off and a lot less resistant, which is something you’d think the tech companies themselves want.

In short, we’re having the wrong argument. AI companies are trying to convince us they’re going under the human bar and not replacing us, when what they should be arguing (and doing, they key part) is vaulting OVER that bar and doing things no human could possibly conceive of. As ambitious as that would be, it would have the benefit of not actually replacing humans because no human is trying to predict drug impact or climate change at that level. But it would also have the benefit of not making humans feel like they’re being replaced since the whole rhetoric now would be about doing something that might help us.

Once we understand AI this way it will, I think, bring a lot more people to the responsible AI movement — people who intuit something is off here but have yet to discern it enough to pick up a placard — and get us collectively pushing for and against the right goals.

AI companies like to talk a game about building a complement to humans — some entity that will be smart enough to will help us but not so smart it will edge us into irrelevancy. But anyone can see that if this is your goal it will automatically lead to exactly what you fear, because the drive to make something more effective will naturally inevitably push us out; otherwise why would we use it in the first place? What they should be talking about (and, again, doing) is building an entire parallel infrastructure, not so much a complement as a new conduit, and focusing on the benefits of that.

This means fewer pieces of video and image slop, fewer inferior search engines, fewer AI Agents booking travel, fewer programs doing legal reviews, fewer use cases of AI as a shortcut to what humans were already doing fine taking the regular cut.* Fewer of those quick-profit plays, and more of those rewarding (and, I’d argue, ultimately financially enriching) humanity-advancing ones.

[*Not to mention deployments where it’s not just being used as a short cut for reputable endeavors but as a short cut for disreputable ones, like employee and citizen surveillance.]

This is why Eric Schmidt gets booed at college graduations. Not because anything technological is automatically bad. But because he has completely flubbed the Matt Stone Test while trying to pretend he hasn’t. He sees AI as something that is replacing your accountant but is trying to pretend it’s something that’s curing cancer. It’s sad when even a college senior who’s not yet in the work world can see through that elision.

I don’t know what the Matt Stone Test is realistic for publicly traded companies who see the “efficiency gains” (read: “replacing human-labor savings”) dancing in front of their eyeballs. When faced with two choices, the healthy and the heart-stopping, we know what tech companies reach for; that cardio-smart almond gets passed over for the dopamine hit of junk food every time. But man, this test is a good way to think about the issue. And clarify what AI is about — not at all bad itself, just unleashing some pretty awful effects when it’s cynically weaponized to make a buck.

Every week we bring you the TSAS — the TOTALLY SCIENTIFIC APOCALYPSE SCORE (tm). It’s a barometer of the biggest future-world news of the week, from a sink-to-our-doom -5 or -6 to a life-is-great +5 or +6 the other way. 2025 ended on an up note, but the score for the year was dismal — a horrendous -42.5. Can we turn things around in 2026? Last week, we pushed into positive territory for the year for the first time — oh yeah. This week? More good news.

FINALLY, A SIMPLE WAY TO KNOW WHETHER AI IS WORTH PURSUING: +2.0