The Secret to Understanding AI

In the before times—before machines could hallucinate, before compute was a noun—it was not uncommon to go several weeks without someone telling me the world was about to end. Similarly, a whole season might pass without anyone assuring me that it was also, simultaneously, about to become perfect.

That particular luxury died on November 30, 2022, when OpenAI released ChatGPT to the public. What followed was less a news cycle than a weather event—a tropical depression that would not budge. Within weeks, millions of people had their first experience with generative AI. Within months, every major technology company had announced its own version of a large language model, or a partnership, or a pivot. Venture capital arrived drooling. Most people in tech think about money, but AI-profit projections are different—like CFO fan fiction, written in Excel. In 2023, the McKinsey Global Institute estimated that $4.4 trillion in annual corporate profits could be up for grabs from generative AI alone. Morgan Stanley estimated $40 trillion more in operational efficiencies. The words artificial intelligence went from obscurity to a constant hum, present in every earnings call, every school-board meeting, and far too many arguments at dinner tables.

Yet for all of the noise, a simple question stayed unanswered: What exactly was this new technology going to do for people? Not for corporations or the billionaires who aspired to become trillionaires, but for people with mortgages and sick parents and children struggling to learn things.

Answers, when they came, were either so enormous as to be meaningless or so specific as to seem beside the point: AI would cure cancer and write your text messages. AI would create deadly superviruses and drain all meaning from our existence.

I got to know some of the people delivering these competing prophecies, and they had a lot of overlapping traits. Brilliance, certainty; delight at being players in a turbulent drama. A hairball of motives.

Accelerationists—the cure-cancer people—were often in charge of, or funded by, or praying to be funded by the companies whose products they were predicting would save civilization. Doomers—the extinction people—were then led by Elon Musk, who sued OpenAI to try to reclaim its founding mission as a nonprofit serving humanity. (Although a more plausible read was that he wanted to hobble his archnemesis and former partner, Sam Altman, long enough for his own AI start-up, xAI, to catch up.)

Matteo Wong and Lila Shroff: Silicon Valley is in a frenzy over bots that build themselves

Geniuses, rivalries, clashing ideologies—all lovely ingredients for a writer like me to work with. But documenting a state of confusion isn’t the same as providing clarity, and after months of talking with the assorted zealots, I was getting a little loopy myself. I needed someone who could see the technology clearly—not as a salvation or a catastrophe or a Powerball ticket, but as a tool.

Danny Hillis was one of the first people on the internet, back when it was still called the ARPANET and the community of users was so small that he knew all of the other Dannys online. His work on parallel processing led to the creation of cloud computing, which laid a foundation for the rise of artificial intelligence. Danny listened to me rant about the AI industry with sympathy and bemusement. He’s seen every gold rush in Silicon Valley, and his heart rate is as steady as the Buddha’s. When I arrived at my exasperated coda—“Danny, what is AI actually good for?”—he was ready.

“Try to imagine the tech without the tech companies,” he told me.

To my embarrassment, it had not previously occurred to me that one could do that.

Danny was certain that an AI counterculture had to be out there, beyond the tech megalopolises, full of people experimenting with AI in ways more meaningful than the latest chatbot-calendar integration. Why not write about them?

Not long after, I discovered whole tribes of people who were tinkering with artificial intelligence to make things that matter—education, health care, government, human connection—work better. A Cleveland Clinic cardiologist was using AI to make lifesaving heart scans available to everyone; teachers in an Indiana school district were finding new ways to engage with students; technocrats were bringing their deeply unglamorous government agencies into modernity; a former physicist was racing to build AI-powered translation for nonverbal autistic kids, including her son.

Like the accelerationists, these people are plenty frustrated with bureaucracies and ideas that have aged into obsolescence. But they don’t believe in the techno-optimist philosophy known as “Move fast and break things,” because they don’t want to break things; they want to fix things. They had run into a problem that defied conventional solutions, and were stubborn or desperate enough—or just cared enough—to keep going, even if it meant having to learn more about technology than they had ever wanted to.

The downsides of AI are real: misuse, malfunction, the temptation to replace people instead of teaching them new skills. It’s easy to understand why some people would prefer that AI just go away; no one is in the market for another existential risk. But here’s the thing about defensive crouches: They don’t actually stop anything. They just ensure that you get whacked in the back of the head. The people in the AI counterculture have figured out that the only effective response to a transformative technology is not to hide from it but to get your hands dirty and make it work to preserve and improve the things you care about. That’s not naive optimism—it’s enlightened self-interest.

A week before the 2024 presidential election, I went to Washington, D.C., for the least sexy reason: I’d heard that the IRS was up to something. Let me rephrase. People who work inside the tight circle of government information technology kept whispering the equivalent of Psst. Y’know what’s going on at the IRS? When I would answer that I did not, they’d smile and tease me with rumors of some secret AI Fight Club inside the federal government that may or may not exist. Who could say?

It seemed unlikely the IRS was working on a supercool, supersecret AI project, because the IRS runs on ancient tech and has never once flirted with being cool. As for secrecy, I had entered its headquarters to meet then-Commissioner Danny Werfel within two weeks of requesting an interview. But after a few minutes in Werfel’s waiting room, I began to wonder. Dull-blue carpet. Walls the color of cafeteria pudding. The room’s center of antigravity—its un–focal point—was a faux-mahogany cabinet displaying unloved plaques and seasonal gourds. I had never been in a place so perfectly optimized to kill all curiosity. If a diabolical genius were hiding an incredible AI project, this is the anteroom he’d build.

Werfel is trim and boyish, and he welcomed me into his office with the slightly besieged air of someone used to getting kneecapped whenever he stands. Werfel knew what I wanted to discuss, and cautiously allowed that “there’s a trajectory for artificial intelligence that has a net positive impact on society and government.” But he raised a hand to indicate he would go no further: complications first.

The IRS is bound by rules about “inherently governmental” functions and cannot simply replace its employees with AI. It has a duty to serve all taxpayers equally, whether they file on smartphones or with pencil and paper, so imposing chatbots on them isn’t an option. In any case, the IRS has some of the strictest privacy and cybersecurity requirements in the world, and many AI products don’t meet them.

Werfel sidestepped politics—commissioners are appointed to a five-year term that is intended to span presidencies—while acknowledging that the IRS is inherently political. From 2010 to 2021, as the annual flow of tax returns increased by 15 million, its budget was slashed by more than 22 percent. As a result, crucial IT infrastructure had been on life support. Recent cuts were driven by Republicans, but the IRS has always been the essential part of government from which everyone recoils—the body politic’s colon. Since its creation, in 1862, only one president, John F. Kennedy, has visited its headquarters.

“The other thing,” Werfel continued, “is that a bureaucracy like the IRS doesn’t move in 180-degree turns. We move in five-degree turns. And that’s just understanding the biorhythm of our bureaucracy.”

This was such a colossal bummer that I almost missed his pivot.

After cataloging the reasons that it’s nearly impossible for the IRS to use AI, Werfel quietly began to list some of the ways in which the IRS was already using AI. Natural-language processing was speeding taxpayers through call centers and getting them to the right human representative. Large language models, including GPT-4 and Meta’s Llama, were being tested to assist with code generation. Bespoke AI was helping employees spot complex tax-evasion schemes. Most impressive of all, AI was assisting in the translation of the IRS Individual Master File (IMF)—the massive Kennedy-era database that contains not just the tax records of every American, but every change ever made to those tax records—into modern software languages. The IMF is the white whale of obsolete government technology; the team that drags it into the present should be given its own national park.

By design these were incremental changes, only to be whispered about. In the perverse environment of Washington, where the IRS is somehow both the most neglected agency and the most abused, Werfel was shrinking its AI efforts to invisibility, using the perception of the IRS as slow and boring and technologically hopeless as cover for his effort to transcend that perception.

To be clear, Werfel did not admit to any of this. The first rule of Fight Club, et cetera. The closest thing to a slipup was when he said, “The IRS has launched more digital tools in the last two years than we launched in the previous 20, and it’s possible AI can help us move faster than that in the future.” But that hardly counts as swagger.

Maybe he was a diabolical genius, although traditionally we call a person willing to swallow their ego and navigate hardship in the service of their country a patriot.

Danny Werfel was obliged to suppress most signs of exuberance—and committed to the bit. Further down the chain, Kaschit Pandya, then the IRS’s chief technology officer, had the freedom to get excited. “The opportunities with AI are endless,” he told me.

For most of the 21st century, IRS customer-service reps would get calls from taxpayers, listen to their questions, and use an internal search engine that had indexed thousands of pages of Internal Revenue Manuals in hopes of finding answers. “Very kludgy,” Pandya said.

Pandya’s team used AI to restructure these dense manuals, making them easier to search and navigate. Now when a taxpayer calls, representatives can find answers almost instantly in language that makes sense to non-accountants. This was one of many service improvements noticed by the national taxpayer advocate, who informed Congress in 2024 that “despair has turned to cautious optimism.”

“A whole bunch of IRM manuals, it’s not the sexiest thing,” Pandya said. “But when you call us, and our customer-service reps can get answers faster, that’s a modernization journey too.”

Pandya is the first chief technology officer I’d ever heard use the phrase modernization journey. It’s the equivalent of meeting a brain surgeon who talks about chakras, and Pandya modestly agrees that he’s unique in his field.

After college, Pandya worked in consulting; clients said that he was great at deploying new technology—and horrible at explaining it. He went back to school for an M.B.A. and loaded up on communications courses. “I used to say, ‘Here’s some tech, and here’s what it does,’” Pandya said. “But it didn’t resonate until I could explain why you should care, why it impacts you, how it can be transformative.”

At the IRS, Pandya splits his time between tracking the latest AI developments and explaining them to the people who have to live with the consequences. He’s learned that the second part is harder. AI can mean the difference between a taxi and a bullet train—an obvious improvement, unless you feed your family by driving a taxi. So he’s become a skilled empath. “We can’t get to the target if you don’t come along with us on the journey,” he said he tells people. “The intent isn’t simply to extract knowledge from you. It is to broaden your portfolio of available skills, and make it so that you are the reason why we succeeded, not the tech underlying the effort.”

Empathy has practical limits, though; at some point, the systems just have to work. The IRS divides its technology into two tiers. Tier one refers to crucial stuff, such as the IMF. Tier two encompasses all of the programs and machines that integrate with tier one—everything including smaller databases and fraud detection and the taxpayer online-account portal—but that aren’t part of the tax record.

The IMF is the IRS’s master database—software originally built decades ago that runs on a mainframe, a kind of industrial-strength computer designed to process massive amounts of data reliably and securely. Mainframes are designed to be up and running almost 100 percent of the time, making them ideal for securely managing sensitive government data. (Seventy percent of Fortune 500 companies—airlines, banks—also rely on mainframes.) “Our hardware gets updated every two to three years—it’s not outdated,” Pandya said. “What makes it seem old is the software. The system was originally built 60 or 70 years ago using programming languages like COBOL and ALC, and those are still what run the IMF today,” he said, referring to Common Business-Oriented Language and Assembly Language Code.

In a vacuum, there’s nothing wrong with COBOL and ALC. They grind away inside the mainframe efficiently. But not everything is a mainframe, and most of the software in tier two—and the world—is coded in languages that prioritize usability and interoperability with other software. That’s turned COBOL and ALC into the equivalent of Sanskrit—perfectly useful if you happen to know a bunch of other people who speak Sanskrit, and pretty isolating if you don’t. Plus, COBOL and ALC engineers are retiring, and dying, faster than the IRS can replace them.

If a customer-service agent using modern tier-two software wants to look at a taxpayer record on graying tier one, they have to navigate multiple systems or wait while middleware, which is exactly what it sounds like, translates the request. That’s usually what’s happening while the IRS’s signature hold music is slowly lobotomizing you.

In 2014, the IRS began a 10-year process to replace the IMF’s 2 million lines of code. By law, there could be no disruption to tax filing or the 400 IRS processes that rely on the IMF—“ripping and replacing was not an option,” Pandya said. “And there was no tool out there that easily converts from the old to the new. What that meant is: We had to use an approach called ‘pair programming.’ Literally, you, COBOL, and an ALC programmer sit here next to me and tell me what this thing is doing, and I will work on creating a similar logic in the modern version of this language.” But eventually, somehow, by November 2024, 90 percent of the IMF was shiny and new.

Next up is the migration of the equally monstrous Business Master File, and it will not take 10 years. “This,” Pandya said, “is where AI gets really exciting for us.”

AI tools such as Llama, Claude, and ChatGPT can digest COBOL and ALC and create pseudo-code. It’s not a one-for-one translation machine. It’s an AI assistant that extracts the logic of the original code and gives human developers a foundation to build on. But what took months on the IMF project, AI is doing in days.

Rogé Karma: So, about that AI bubble

These tools also automate documentation, the process by which software engineers are supposed to—but rarely do—note all of their thinking so that future engineers can modify or maintain the code. “When I talk to people outside of work and say we’re using AI so our developers can save two hours a week on documentation, they’re like, ‘So what?’ But it matters!” Pandya told me. “When we have 500 or 1,000 developers, all of a sudden, two extra hours a week turns into some real development progress that we can make at a much faster rate.”

Migrating these master files is a once-in-a-lifetime test—the CTO equivalent of restoring Notre Dame. But there is a type of person who finds all of this—civil servants, upskilling, rule-following, empathy—insufferable. Not just inefficient, but offensive. To them, the government is a failed company that never goes out of business, and every public employee is complicit in its mediocrity. This type of person does not believe in incrementalism. They believe in chainsaws, in moving fast and breaking things, especially if the things are slow, unionized, and taxpayer-funded.

This type of person was reelected to the presidency the day after Pandya and I spoke.

Werfel resigned on Inauguration Day. The IRS cycled through four acting commissioners in three months, and lost its chief financial officer, chief risk officer, and chief privacy officer, along with thousands of employees who took buyouts and walked out the door. In March 2025, the IRS told the Government Accountability Office that it had paused its modernization programs because it was reevaluating its priorities.

Despite DOGE, Pandya (whose title is now chief information officer) and his colleagues proved that careful, unglamorous AI adoption can move a bureaucracy toward something better. AI is still young and weird—a puppy that will read the Quran in Portuguese and eat the TV remote. But the tech is catching up to its hype, and every day it gets easier, faster, and a little less strange. If we don’t shape AI for good, in our government and in our daily lives, it will be shaped by people who don’t know or care about our problems. If we don’t teach it what matters, someone else will teach it what’s profitable. The choice isn’t between a world with AI and a world without it. The choice is between AI designed by people who think fixing things is worth the trouble, and AI designed by people who think breaking things is more efficient.

This essay was excerpted from Josh Tyrangiel’s forthcoming book, AI for Good: How Real People Are Using Artificial Intelligence to Fix Things That Matter.