A conversation with Josh Tyrangiel, the author of “AI for Good,” on the right uses of the transformative technology
Most human intelligence about artificial intelligence falls into one of two camps. The accelerationists promise that AI will cure cancer, fix the climate and usher in a golden age. The doomers warn that it will eliminate our jobs, hollow out our institutions and possibly even end the species.
Josh Tyrangiel — a longtime journalist who is now a writer at The Atlantic — went looking for the people in between. His new book, “AI for Good: How Real People Are Using Artificial Intelligence to Fix Things That Matter,” reports from inside the Cleveland Clinic, the IRS, a school district in Indiana and a Pentagon war room. The people are doctors, teachers, engineers and civil servants trying to use AI to reduce the friction of real institutions — messy data, misaligned incentives and bureaucracies built decades before the technology existed.
We asked Tyrangiel to walk us through what he found, and what it means for New York. Among other things, the conversation touched on classroom rules and the temptation of cheating; lessons that the IRS, of all places, has to offer city government; and whether AI can finally unstick the permitting process.
Tyrangiel is an optimist, but a chastened one. He has been through enough technology cycles to know that passivity in the face of a powerful new tool tends to produce the worst version of them. The work, he argues, is to be neither dazzled nor afraid — but to insist that human judgment stay in the loop.
The following transcript has been edited for length and clarity.
Vital City: The conversation around AI is polarized, like so many other things. Your book is an attempt to weigh both sides of the debate, but you ultimately come out with a fairly positive read of what the technology can accomplish. Can you talk about why you came down there and whether you struggled with that conclusion?
Josh Tyrangiel: I approached the whole thing from a place of detachment, because that's what a good journalist is supposed to do.
On the accelerationist utopian side, inevitably I heard that we were on the verge of an incredible breakthrough. Our lives were going to be so much richer and so much better, cancer was going to get cured, climate change was going to be mitigated. At a certain point, it was clear this had become a kind of dogma. So I would say, "Tell me more about how the cancer curing is going to happen." You'd find they actually hadn't thought about that. It was just their plucked example of how much better things would be.
Then I would talk to people who were like, "You have no idea. You better pick a side. You have to come with me if you want to live." — basically quoting The Terminator. I thought, oh, boy. Both sides might be right, but my hunch, having been through one or two of these moments before, is that the answer's somewhere in the middle, and that we may not be dealing with a deity, but we may be dealing with something a lot better than the other software we've seen.
So my goal was to get as far away as possible from the extremes and figure out how people are using the technology right now, to make the things that I care about better — the lives of my family and my children, the health of the republic, the quality of life of all the people around me. Naturally, that meant going into healthcare and hospitals, education, government agencies and human connection.
Vital City: The two extremes seem to agree on one thing: This is really powerful technology, and it's going to mean massive change.
Josh Tyrangiel: I do think the technology itself is incredibly powerful. I also think it's important to parse what that power is. The name "artificial intelligence" is very deceptive. What this actually is, is human intelligence on steroids. It’s taking tons of data created by human beings and processing it at speeds and at a volume that human beings simply can't. It's going to take us some time to adjust to how that works.
There are downsides. People are very impressed with speed, but speed isn't always quality. Even now, you'll get hallucinations, and those can be meaningful if they turn up in the wrong documents — health studies, legal documents. You're also not getting real-time feedback from human beings into the state of our systems and the states of our needs and desires. It's almost like this abstracted power right now, and it's very seductive. But that doesn't mean it arrives without complications, and it doesn't mean there's no place for humans in managing it and bringing it into the systems and institutions we care about.
Vital City: It's not just speed — it's the authoritativeness of presentation. It can produce things that just look so good, so polished.
Josh Tyrangiel: Yes, It produces results that convey an aura of certainty. Part of that is the programming, and part is just the speed. It produces so fast that you're like, "Wow." But we need this kind of secondary response, which is, okay, how good is this really? The trick is dazzling, but you have to sit there for a minute and process it with your own brain to understand where it's maybe a little less dazzling than it may have seemed. That's going to take years of adjustment.
Vital City: You spend a lot of time in the book on the use and misuse of AI in education. Public education has so many tech failures littered through its recent history. Already we've seen a number of AI failures, but we've also seen some qualified successes. As the New York City Department of Education or any other department of education thinks about how to actually use the technology, what level of concern should they have? Should they look at this like a threat? A tool? Something else?
Josh Tyrangiel: They should look at it as both. It's a threat in that if you've structured your curriculum solely to assess credentials, you have to know that every student out there has the ability to fake their way to a credential. So there needs to be a wholesale examination inside of education: What are we doing here? You've already seen a rise in blue books and oral exams — those changes will probably continue.
I also think there needs to be some questioning about what the function of the classroom is. Is it merely to get a student to a credential, or do we want to focus on identity formation, on teaching kids who they are in relation to the world? Should we be changing the criteria for what the classroom is about — from "You've got to get to this number on a quiz," to a deeper form of moral and intellectual education?
Vital City: I saw something the other day on social media saying that the acquisition of knowledge in the age of AI is going to distinguish kids from one another — that AI is so authoritative-seeming that you need actual grounding to know when to take it seriously and when not, and the only thing that gives you that is knowledge. Do you buy that the rise of AI actually places a premium on knowledge acquisition in K–12 education, or is that just spin?
Josh Tyrangiel: I don't know, honestly. It's too early to tell. What we're going to see is the high flyers are going to be just fine. The self-motivated kids are going to use these tools to get themselves higher up and learn more.
The risk, as always, is in the middle and the bottom of the class. The temptation to use these tools to just get by is going to be off the charts, and all of us can relate to that. We all have something that feels like drudgery in our lives, and we'd love to find hacks to get away from it. The vulnerability of those kids to missing out on the experience of struggle and of practice is significant.
What I discovered in my reporting is that when you take an AI tool like Khanmigo — an AI tutor that’s a joint venture between Khan Academy and OpenAI and you just put it into schools in front of students, the reaction's sort of like, "Mm, I have lots of software." They've got laptops and open tabs, and this is just one more thing. The only difference is that it won't give you the answer. The tool was designed in the style of a tutor, and it's supposed to bring you along at your own skill level, in the proximal zone of development. So if a kid is a high flyer and wants practice, it's great. There aren't that many high flyers.
But what I did see is that teachers were able to use the tool in far more productive, and in some cases, really revolutionary ways. At its heart, it's an LLM designed to the curriculum of your district. I watched a couple of teachers who had come to the realization through COVID that standing in front of the classroom, broadcasting knowledge to their students and then assessing their receptivity to it, is really not working anymore. Kids are distracted. Kids learn in very different ways.
So they reinvented their teaching styles on the fly, and turned things into all sorts of collaborative labs, assessments through group conversation. That kind of transition ordinarily takes years of trial and error but some of these teachers are using Khanmigo to do it the day before class or between classes.
Not every teacher. But some have begun to figure out that the tool is far more powerful for them than it is for their students. The teachers' union today put out a statement basically saying they want no AI tools in schools, particularly in younger grades. I get it. We're going to see a series of pendulum-swinging responses over the next couple of years: let's ban it all, let's bring it back, let's find a middle ground. But what I saw left me pretty hopeful that AI can help teachers change things.
Vital City: If you had to write a rule for how students should or shouldn't be able to interact with AI in the classroom, would you write a rigid rule at this point, or would it be more flexible?
Josh Tyrangiel: You'd have to be flexible, because we just don't know enough. First of all, every superintendent will tell you they have no idea what can happen in an individual classroom. l think the variation from classroom to classroom and teacher to teacher is a feature, not a bug. If we look at it that way, maybe we can rapidly figure out what's working.
You don't want to treat your schools as laboratories haphazardly. So much of ed tech, to me — the flagrant disregard for what's actually supposed to happen in the classroom in service of somebody's brilliant tech idea — is just appalling. That's why it fails, time and again. You design a system and expect the students and teachers to adjust to it, as opposed to having a system that works at the rhythm of the classroom.
If I were in the very challenging job of designing AI policy for education, I'd say we need a systematic way of experimenting and rapidly apprehending what works in what classes, so we have examples. In this Indiana school district I visited — it's small, run by a ferocious superintendent — she was committed to two things. One, rapid experimentation: getting AI into the classroom so that students understood it was there, so that teachers understood it was there, so that they felt like they weren't King Canute holding back the waves. Two, a massive commitment to professional development, where teachers come in and say, "I learned this, and it does this really, really well. And oh, by the way, it does this kind of poorly, so I adjusted my class."
That is a lot to expect of teachers and of kids. When we talk about the current environment where people really resent AI, I think a large part of the reason, justifiably, is that this tech has arrived and we're all being told, "Change everything." It's hard to change everything. I like the way I do certain things in my life. At the same time, we kind of have to rise to the challenge, because if we don't, we now have 25 years of experience that teaches us that when you're passive in the wave of technology, you generally get the worst out of it. Social media — we were all so optimistic, we were all so dazzled and we ended up with crap that undermined the republic.
I know it stinks to have to adjust and change who you are, but there is upside here, and leadership has to be ready to motivate people, to incentivize people to actually use it for the good — because otherwise we know what the bad looks like.
Vital City: Let's transition to government service delivery more broadly. There are fears that AI's going to replace white-collar jobs, knowledge jobs, en masse. There are also hopes that AI can make taxpayer dollars go further. That's actually a version of the same idea — maybe we don't need 300,000, or however many, New York City public employees if 250,000 equipped with AI tools can do the job as well. What is the right way to strike a balance here? We don’t want to strike fear in the hearts of every public service employee that their job is on the chopping block, but we do want to find efficiencies and maybe even help government do better.
Josh Tyrangiel: It's hard to call out the IRS as a paragon of wise management when it comes to technology. But I did a ton of reporting around what the IRS was up to with its technology stack. There are certain rules the federal government has about tasks that are inherently governmental. They have so many regulations around what they can do, and they also have so much political pressure, because so many elected members and so many voters actually don't want the IRS to be successful.
But what they were able to carry out, very quietly, was to explain to the thousands of engineers inside the IRS that these tools were important. They were going to make their lives easier. They were going to make the IRS function better. They were going to open up possibilities for more pride in the organization — that by using tools that could do the work faster and cleaner, and they'd create a better customer experience.
They oriented people around two things. One, used properly, the tools were going to be successful. Two — and this is true for everything I explored — none of this stuff works without humans in the loop. One of the great myths of AI is that you just roll it out, it works itself into the enterprise or into the corporation, and all of a sudden everything's great. It turns out that's just not true. All of this stuff needs really wise Sherpas — first to get it accepted into the organization and into its systems, but then to calibrate how people use it.
Vital City: Stephen Goldsmith wrote a piece for Vital City saying that "accountable discretion" is the model — that AI actually makes it possible for a bureaucrat, who once upon a time simply had an enormous stack of rules to follow, to have a more flexible way of navigating them, with a lot of information that can lead to smarter human judgment.
Josh Tyrangiel: I think that's right. When you look at New York City, it's largely governed by massive stacks of paper, some of them dating back 100 years or more. For a lot of employees in city agencies, most of what they're doing is consulting a different kind of code. The law is a kind of code. Regulations are a kind of code. For people to dive into this antiquated code and then figure out how it applies to the real world is very complicated.
The reason AI is so powerful with things like the law and regs and tax is it's basically the same thing — a series of rules. So it can manage document comparison. It can take a permit application, look at the law as currently stated, and give you a recommendation rather quickly. What it can't do is know the details. It can't know little things about the geography. It can't walk the street. There are all these ways where the discretion of the employee becomes so much more valuable and the drudgery is outsourced.
It's not going to be true in every case, and I'm very sensitive to the fact that some people feed their families based on their tolerance for drudgery, and we want to be careful about that. But what I saw time and again is that the tech is more than ready to help us improve government. I don't sense yet that leadership understands that and is willing to have the pretty complicated conversations with employees about how we're going to navigate this together. But I'm very convinced the technology can do the job.
Vital City: If the government is making a decision on a potential sanction — or for permission to do something — that involves 20 or 25 steps, and at step 17 the AI makes an error, how would we know, if humans aren't checking every step of the way? And if they are checking, doesn’t that defeat the purpose of AI?
Josh Tyrangiel: You're baselining your expectation on a perfect human system. What we've seen in the federal government is that a very high rate of food stamp applications were rejected incorrectly as recently as a couple of years ago. So can we get better than that? I think we can. You're not going to have a perfect system. You are always going to have errors within AI. There's no question about it. But you can certainly reduce them, and I also think you can reduce the time to judgment.
In New York City, every single person complains about permitting, no matter what business they're in — from food carts to buildings. Everybody has the same experience: the paperwork is cumbersome, often they need expediters and a single mistake will slow the process or send it back to the beginning. There's really no reason we cannot take everything we know about permitting and create a portal for people to test out their initial pass, so they can find out what errors they made — or may possibly have made — before submission.
If we can get better, cleaner paperwork to the people making decisions about permits, think about the thousands of hours saved. Think about the productivity, the speed to building, the bias toward action, as opposed to this pervasive air of entropy. It’s true that we're going to have some errors with AI. But we have errors now.
Vital City: In New York City today, we have a transit system that, by law and contract, requires two people to run every single subway train. At the same time, we have Waymo trying to get in on the streets, which uses zero people in every car. A lot of people who worry about congestion have very well-intentioned objections to Waymo's entry. They argue Waymo could be dangerous — even though plenty of research says that driverless cars are safer than cars driven by people. Should New York lift its current ban on driverless cars? And if it does, what's the path?
Josh Tyrangiel: It's a big question, and the answer comes down to one word: politics. Politics doesn't have to be rational, and oftentimes the lack of rationality is helpful — because New York is a vibe. The mayor's top job is to understand the vibe. If what the politician understands is, "Oh, there is resentment, rational or irrational, to these cars with no drivers, but people are comfortable with two operators on every train, even though trains all around the world operate with zero conductors" — that's where the system collides with the human beings.
That's why so much of this really revolves around the people doing the work. The tech is complicated. But what human beings do is assess the other human beings on the other side of the equation. When we look at these labs, we ask, "Do I trust Sam Altman? Do I trust him more than Dario Amodei?" It's natural for us to focus on those human beings in making our judgment about the tech, in the same way that it will be natural for politicians to focus on the human beings if they implement the regulations.
Whether we want it or not, human-centered AI is still going to have to be the answer, because you can't just plug this stuff in and expect it to work. So the burden is going to be on people's assessment of the tolerance for change, of the appetite for change.
Vital City: What does the right balance look like on the roads? If you save some number of lives by replacing human-driven cars with driverless cars, that's a good thing. On the other hand, maybe you've lost a much larger number of driver jobs. How do you put those inputs in and figure out whether it's a good or a bad thing?
Josh Tyrangiel: Some of this is economics. When we get this kind of general purpose technology and this much change, ideally it comes in at a pace at which you can maximize the economic benefit while minimizing the harm. The benefit in this case being greater productivity; the harm being mass displacement of labor.
It's really hard to do, because we live in a capitalist system where there are a lot of incentives to get the drivers out, because drivers are a cost center. For the city, that's a disaster — tens of thousands of jobs. Across the country, multi-millions. So we want this to happen at a natural rate of adjustment, so that people's lives aren't so displaced.
But there are two forces at work: the people who are actually doing this work and who are going to have to change, and the people forcing the change upon the country. I hate to keep reverting to politics, but it's our best hope for managing this change. We know what happens if the tech guys are left to their own devices. They're going to force change, and we've had 25 years, particularly in this century, of watching the consequences. People are very, very angry about it.
What I would expect — and we've already seen some of this in the city, but I would expect it nationally — is that AI, as it pervades the economy, is going to be the number one issue, possibly as soon as 2026, certainly by 2028. People are beginning to understand just how many ripple effects it has. When it comes to good policy, I think we have good answers to how we might want to implement it. But it is going to run into the brick wall of human beings, and for that it requires knowledgeable politicians.
Disclosure: Josh Tyrangiel's wife and partner, Sarah Feinberg, is the former president of New York City Transit.
