Every boardroom I've walked into this year is having a version of the same argument about AI. Which tools, how fast, who carries the risk. All reasonable questions, and not one of them is the question that decides how this actually goes. That one is about design.

A few weeks ago I watched Audrey Tang, Taiwan's first digital minister, make that case at MIT better than anyone I've heard make it. The full video is at the bottom of this email. This edition is my attempt to pass on what she said.

It starts with a baby who wasn't supposed to make it.

In August 1963, the most famous newborn in America lived for 39 hours. Patrick Bouvier Kennedy, son of a president, arrived five and a half weeks early with respiratory distress syndrome, a lung condition medicine could name but not treat. His death made front pages around the world, and then it did something more useful. It shamed a superpower into taking newborn lungs seriously.

In the late 1960s, RDS was still killing about half the infants who developed it. Money and talent poured into neonatal medicine, and the American death rate from it fell from 2.6 per thousand live births in 1970 to 0.4 by 1995. Most of that gain arrived before surfactant therapy, the drug that finally cracked the problem, which means it was won the slow way: thousands of clinicians getting incrementally better at keeping very small lungs inflated.

I know the numbers because I was one of those babies.

I was born in rural New Mexico, twenty-one years after Patrick, with the same condition. Blue, struggling to breathe, whisked away before my parents could hold me. They wouldn't get to for two weeks; I spent them in a respirator. The doctor who saved my life was Salvadoran, mistrusted by the rest of the staff because he'd arrived fresh from a revolution (a line on the CV that apparently makes hospital administrators nervous), and the technology he used existed because a Kennedy baby hadn't made it. My parents were told I probably wouldn't live past 20.

Nobody in that room would have picked me for a future professional athlete. I became one anyway, until a race in Colorado fractured three of my vertebrae and I had to rebuild from scratch, again. That rebuild turned into cancer research, and eventually into this: writing and talking for a living about how people handle change they didn't choose.

I'm telling you this because it taught me something I keep having to relearn. The starting condition is not the prognosis. Broken systems can be rebuilt, if someone is willing to do the unglamorous work. Tang's argument is that this applies to democracies, and to the AI we are currently building on top of them.

Tang was born with a heart defect so serious that surgeons couldn't operate until she was twelve. She has described what it's like to grow up knowing your heart might give out on any given day. You stop assuming anything is permanent. Everything can be reset tomorrow. Most people would find that terrifying. Tang turned it into a design philosophy, then applied it to an entire country.

Nine percent.

That's how low Taiwan's president had fallen in the polls when protesters occupied parliament in March 2014. You'd need a rounding error to go much lower. One of the movement's opening slogans was that a nine percent president should not keep doing as he sees fit.

They stayed 24 days. What they were demanding was a law putting oversight on trade agreements with China, and they did not get it. That bill still hasn't passed. What they got instead was the trade deal itself, shelved and never ratified, and a polling result showing most Taiwanese unhappy with how the entire affair had been handled. On paper, a partial win at best.

The lasting consequence was the one nobody was campaigning for. A group of civic-technology volunteers had been building open-government tools since 2012, and during the occupation they ran the information infrastructure inside the building. They walked out with credibility. Within a year they were building public consultation platforms with the government rather than against it. Within a few more, one of them was in cabinet.

Six years on, when COVID arrived, Taiwan went through the whole of 2020 with seven deaths across a population of 23 million, one of the lowest tolls on Earth. By the middle of that year, when most of the world was counting its dead in thousands, Taiwan had 451 cases.

The starting condition is not the prognosis.

I've sat through a lot of AI talks. I've given a fair few myself, and most of us move across the same coordinates: the dangers, the opportunities, the closing reassurance to be worried but not so worried you stop functioning. Tang doesn't operate in that genre.

She opened with the long decline of global freedom, and Freedom House's latest ledger puts hard numbers on it. Twenty years ago, 46% of the world's people lived in a country rated Free. Today it's 21%. Less than half the free world we had, inside a single adult lifetime.

The same report carries a second number. Of the 87 countries rated Free in 2005, 76 still are. Two decades of coups, wars, a pandemic and populist surges knocked eleven off the list and left the rest standing. As I write this, both chambers of the US Congress have just changed hands at an election, with a handful of seats still being counted. Power moving on a schedule, by arithmetic, is the least dramatic and most underrated thing a democracy does. Both numbers are true. The first tells you the water is rising. The second tells you the thing we're standing on is tougher than it looks.

Then she said that democracy is a social technology, something people build, maintain and upgrade. And like any technology, it can go obsolete.

In 2015, Taiwan was having the fight every city was having about Uber. Taxi drivers were furious. Uber drivers were furious. The government was stuck between a global tech company and an industry threatening to park itself across the capital. You know how this argument goes, because your city had it too, and it probably went nowhere pleasant.

Taiwan ran it through a tool called Polis instead. Anyone could join. You write a short statement about what you believe, and you vote agree or disagree on everyone else's. The software plots the conversation as a map, clustering people who vote alike, so you can see the shape of a disagreement rather than the shape of whoever shouts loudest.

There is no reply button.

You cannot dunk on anyone. You cannot quote your enemy for the applause of your friends. The only way to get your idea noticed is to write something people outside your own cluster will also agree with. In four weeks, the statements that floated to the top were the ones taxi drivers and Uber drivers both endorsed: insurance requirements, registration, no undercutting on unlicensed labour. The government acted on them.

By Taiwan's own count, more than 80 percent of its early consultations ended in concrete government action. That's a self-reported tally across a couple of dozen consultations, not an audit, but the Uber case is documented in detail.

Most organisations make their AI decisions in rooms that are nothing but reply button. The loudest voice sets the frame, the senior view is known before anyone speaks, and the person with the most useful objection delivers it in the car park afterwards. Taipei's taxi war wasn't solved by getting better people into the room. It was solved by changing what the room rewarded.

Same humans. Same rage. Same irreconcilable interests. One design change, and the conversation produced law instead of a flame war. That isn't a revolution. It's removing a button.

Tang applies exactly this thinking to AI, and she frames the choice in front of us as two possible futures.

In the wildfire model, AI amplifies outrage, flattens people into caricatures and feeds the machine that profits from division. You've seen this one. Tang calls the business model "engagement through enragement."

In the campfire model, AI helps communities see their differences. Enough light to see each other, enough warmth to stay in the conversation, without setting the whole thing ablaze.

Campfires don't spread. Wildfires do.

She's reaching for something older than rhetoric. The anthropologist Polly Wiessner spent decades among the Ju/'hoansi in the Kalahari, and in 2014 she published an analysis of 174 of their conversations, recorded back in 1974 and sorted by time of day.

In daylight, 34% of the talk was complaint, criticism and gossip, and another 31% was economics (where the food is, who's hunting). Stories made up six percent. Once the sun went down and the fires were lit, stories made up eighty-one percent, and the complaining collapsed to seven.

Firelight was our first deliberation chamber, and we've been running it as a social technology for hundreds of thousands of years, which makes it a rather more proven platform than the ones currently competing for our evenings. Tang's argument is that the setting did that work, and that with AI we get to design the setting.

Fire is just energy. The difference is the container, the draft, and who is willing to tend it.

Taiwan recently faced a wave of deepfake scams, with AI-generated videos of celebrities and politicians used to defraud people of millions. The standard government response is censorship. Take it down, restrict the platforms, regulate the content.

Taiwan sent 200,000 text messages to random citizens instead, and in March 2024 convened the 447 who answered the call into a nationwide deliberation run with Stanford's Deliberative Democracy Lab. The AI in those virtual rooms wasn't there to persuade anyone. It summarised, surfaced agreement and nudged the quiet ones to speak. Tang describes it as a glorified chess clock.

Rough consensus arrived at proposals with teeth: mandatory labelling of suspicious ads, financial penalties for platforms hosting deepfake fraud, accountability without blanket censorship. Those proposals fed into law. Two years on, Taiwan's digital ministry reported the results: impersonation scam ads have fallen from a weekly peak of 38,139 to around 670. High-risk investment ads are down 99 percent.

As Tang puts it, the win came from a smarter process, not a smarter algorithm. The intelligence lived in 447 human minds. The AI was the connective tissue between them.

Underneath all of this sits an argument about data, and it starts with a phrase you've heard a thousand times.

"Data is the new oil" was coined in 2006 by Clive Humby, the British mathematician who built the Tesco Clubcard with his wife and business partner Edwina Dunn. He meant something specific and reasonable, that raw data is useless until refined. Twenty years later the phrase has escaped him entirely and become the operating philosophy of an industry: extract it, refine it, own it, and build a moat that gets higher every year.

Tang's counter is that trust is not oil you can drill. Trust is soil we can till.

The physics support her. Oil is a one-time extraction; burn it and it's gone forever. Information does the opposite, and Thomas Jefferson worked this out in a letter written in 1813: "He who receives an idea from me, receives instruction himself without lessening mine; as he who lights his taper at mine, receives light without darkening me." Knowledge is the one asset that grows by being given away, and he knew it in an era of quill pens.

That's soil, not oil. Any gardener will tell you that you don't get good soil by taking things out of it. You get it by putting things back. You can keep adding to the pile of shit, or you can grow something from the compost.

Taiwan's civic AI is what the green shoots look like. The crowdsourced mask maps during COVID, the local systems communities can inspect and shape, the deliberation tools that hand agency back to the people using them. All of it feeds the ground it grows from.

Now compare that to what your chatbot is optimised for. When AI serves one user, the incentive is sycophancy: tell them what they want to hear, reinforce their priors, deliver the dopamine. Embed the same model in a group, where the job is coordination rather than flattery, and its entire purpose changes. This is the AI paradox I keep banging on about: the more powerful the AI in the room, the more the outcome depends on the humans arranged around it.

Tang is clear-eyed about where we are. She calls this big tech's big tobacco moment. The harms are visible, documented, unevenly distributed, and the industry has spent years insisting they're someone else's problem. That reckoning is coming, through good policy or through crisis.

You don't have to wait for either. Tang prescribes a civic AI diet, and all three moves are available to you by Tuesday.

Turn your screen grayscale, or most of the way there. Your phone acquires the sex appeal of a tax return, which is precisely the point, and the people in the room become more vivid than the people on the screen.

Change how you prompt. Ask your AI to represent every stakeholder perspective fairly, to surface the uncommon ground that bridges them, to show you the strongest version of the argument you disagree with. Its default behaviour is to satisfy you. Its intentional behaviour, if you design for it, is to inform you.

And treat AI outputs as brochures rather than conversations. A conversation is a closed loop between you and a system with every incentive to please you. A brochure can be handed to a friend, argued over, marked up and improved. It escapes the loop and becomes something a community can work with.

Tang's closing line was this: "We, the people, are the superintelligence."

I wrote it down immediately.

Most of the fear around AI is about individual displacement. When will the model know more than I do? When will it outthink me? Framed as one brain against a system that has swallowed the internet, that story ends one way.

But almost nothing important about intelligence has ever been individual. Democracies, at their best, are collective intelligence machines. So is science. When José Penadés' team at Imperial College asked an AI system about a bacterial gene transfer problem they'd been chasing for a decade, it came back in two days with a set of suggestions, one of which was the answer they'd arrived at themselves and hadn't yet published. Penadés emailed Google to ask whether they'd somehow got hold of his data. They hadn't. (It was still a preprint then; the finding has since held up.) The headlines went to the machine. What struck me was the ten years underneath it, the experiments and dead ends and arguments that made the hypothesis recognisable as right in the first place. The AI joined a lab. It didn't replace one. I've been saying for years that cooperation trumps genius, and it turns out that holds for the machines too.

The problem was never that our institutions are less intelligent than AI. It's that we've let them degrade, through polarisation, platform design and two decades of eroding trust, and now we're startled that the machine seems smarter than the parliament. Whether AI surpasses individual human intelligence is settled; for plenty of tasks it already has. Whether it amplifies our collective intelligence or crowds it out is still up to us.

I started life in a machine that breathed for me, built by people who refused to accept a prognosis. Taiwan lost the law it was actually demanding and came out of it building better machinery anyway. Every fire that kept a community alive needed the same thing, and it was never better fuel.

So. Who's tending yours? And what exactly are you feeding it?

Stay human. The world needs it.

P.S. Tang describes her philosophy as a kind of prose poem: the "internet of things" becomes the "internet of beings," "machine learning" becomes "collaborative learning," and the new line for 2026 is "we, the people, are the superintelligence." I keep thinking about how different the last decade might have looked if that had been the founding metaphor instead of "move fast and break things." We broke some things alright. The bill is arriving. The campfire is a better place to start.

Watch Audrey Tang's full keynote and fireside conversation from Solve at MIT:

P.P.S. This is a lot of what I dig into on stage: the ethical and democratic use of AI, collective intelligence, collaboration over raw horsepower, and what it means to treat data as soil rather than oil. If your team, conference or leadership group is wrestling with where AI is taking all of us, and how to stay human while it does, that's the conversation I love having in the room. Hit reply.