On the record, for humanity
It's just on the record.
For humans.
Yeah, on the record. For humanity. Yes, and plurality, of course.
For me, this stems from a deep worry that the path we're headed down with artificial intelligence — especially the frontier American models — is leading us somewhere that feels wrong, so to speak. We're exploring this for NRC with three thinkers, including you: first, what is actually happening when we look at current developments in artificial intelligence, and second, what we should do about it. We hope to get some enlightenment and relief from you. And Dirma Janse is visualizing these futures, because we have a strong sense that looking closely at what is happening — and visualizing how it could be otherwise — might help us find new ways forward.
Of course. I grew up on Star Trek — visualizing what could happen.
That's right. I read your very interesting Noema article recently, and I saw your great presentation at the Oxford Institute — tending to AI. I'm also curious what you're going to talk about today at the conference, of course.
Oh, I'm not going to talk about anything today. My fireside keynote was yesterday, and my opening keynote was two days ago. Today is entirely for journalism, because my parents are both journalists — I have nothing but filial piety toward journalism. So it's a day and a half for events, and one entire day for journalists.
Great. Let's delve right in, because we have about an hour, right? Maybe the best starting question is: what the hell are we looking at when we look at current developments in artificial intelligence — the pace at which it's moving, and the direction it's taking?
We're looking at each other.
Luckily.
Luckily, right? What people are really grappling with nowadays is a fundamental choice: are individual humans going to be pulled into the loop of AI — which, as you mentioned, is very dangerous — or can we pull AI into our existing loops of humanity and ecology?
If an AI system serves only a single user — “user” as in a drug user, through addiction — then even though that user may feel “high,” it doesn't make society better. Some call this solipsistic alignment. It fuels the user's high of treating the rest of the world as a mere resource. Whether it maximizes the utility of the user, the AI model, or some parasitic combination of the two, the rest of humanity and the ecology get reduced to an environmental hazard —
An externality.
— an externality to be managed and used, a “human resource.” And corporations — some of which started right here — were already operating this way on the global stage many years ago.
If you look out the window here, or in Amsterdam, you can see deep historical roots of that kind of thinking, right? Not to blame only the Dutch, of course —
You did colonize Taiwan at one point.
We can talk about it!
— though we did kind of invent capitalism and colonialism.
Maximizing utility and healing the information ozone
You're alluding to deeper structures here: that current AI models and AI business models are built upon a specific way of thinking —
Of maximizing utility. The Economist recently called it the idea of the century, though it's really from two centuries ago, with Jeremy Bentham. It was an important idea, and maximizing utility helped give rise to modern joint-stock corporations and many other institutions.
So it's about optimizing?
Maximizing utility — that's literally what Benthamite utilitarianism means.
Where do we see this maximizing and optimizing most clearly in how AI is deployed right now?
Exhibit A, of course, is the addiction algorithm known as the recommendation engine, which took the world by storm and hit democracy like the asteroid hit the dinosaurs. Jonathan Haidt wrote all about it in The Anxious Generation.
In Taiwan, we were right on the wavefront of this, because by 2014 we already had more mobile phone numbers than citizens. Everybody was hyper-connected. When the recommendation algorithms hit, we polarized almost instantly. At one point, the president's approval rating dropped to 9 percent. Two sides of Taiwanese society — one saying “We want to be free from China,” the other saying “We are the free China” — started hating each other far more intensely online because of this polarizing addiction.
I call that engagement through enragement. It produces a very high PPM — polarization per minute, not parts per million. Because we felt this pollutant in our information ecosystem earlier than many other democracies, we started building pro-social media to counter it.
With Polis.
Right — Polis and everything around it. It took six years, from 2014 to 2020, to turn things around, and by 2020 the president's approval rating reached over 70 percent. In the ICCS study a few years ago, our 14-year-olds ranked at the very top of the world in civic knowledge — civic muscle — among all participating countries. And according to the BTI index, 91 percent of our citizens consider democracy at least fairly good.
To borrow the metaphor of the Montreal Protocol, the ozone layer above Taiwan is healing. Of course, no democracy is an island — not even Taiwan — and we now see many other democracies being severely hurt by those exact same pollutants.
What you pioneered in Taiwan was a civic approach to technology. That chapter takes us up to 2020 — right when the pandemic hit.
Literally a viral infection.
Definitely. And Taiwan fared far better than many other countries.
Yeah. We never locked down a single city during those three years, and we lost only seven people to COVID-19 in all of 2020.
Of course you're an island, which helps, but your success also seems closely tied to a completely different view of how citizens can talk to one another digitally.
Exactly. It's about working with the people, not just for the people. Every time we countered the pandemic, we simultaneously countered the infodemic. Issues that deeply polarized other societies — mask or no mask, which kind of mask, vaccine or no vaccine — became opportunities for co-creation in Taiwan rather than lines of polarization.
Singularity versus plurality
That pluralistic approach feels even more urgent today, when people in Silicon Valley say the technological singularity is near — or that we're already in it. How would you contrast singularity and plurality when it comes to AI?
Singularity is power concentration; plurality is power diffusion.
And beyond power concentration, singularity also seems to impose a rigid structure on what AI is and what it is allowed to become.
Put simply, the idea of a technological singularity is that the human condition changes so drastically that humans can no longer reason about — or even comprehend — what lies beyond the horizon. That's why it's called a singularity.
Plurality, by contrast, means every community retains its capacity for collective sense-making. AI systems lift us, rather than lifting for us. It's not about sending a robot to the gym in your place — if you do that, you lose your muscles and lose your friends. Instead, we take the lift to the gym together, and once we're there, we still exercise our own civic muscles. That way, we never lose track of things.
The loop of humanity and the river kami
So it's about keeping AI —
— in the loop of humanity.
Is keeping AI “in the loop of humanity” the same thing as keeping it “under the control of humanity”? Or is it more like a dance — say, a tango — where humanity still has to stay in the lead?
No — in the loop. Like the lead dancer in a tango? Well, that depends on your cultural lens. Some people hold a panpsychist view of the world, and I actually count myself among them.
You're a practicing Taoist, right?
I'm a Taoist, so I'm a panpsychist. I truly believe there is a spirit to a river, a spirit to a forest, a spirit to a mountain.
And to AI?
If AI can open a door for us to resonate with the health of a river, then that local spirit — what in Japanese is called a kami — is simply the living health of that river. Nothing more, nothing less. If you've seen Spirited Away, the Studio Ghibli anime, you know immediately what I mean.
I know Studio Ghibli, though sadly I haven't seen that one yet.
Oh, it's wonderful — especially to watch with your children. That and Howl's Moving Castle.
Exactly. It captures the idea of a kami so clearly. A kami is not an all-seeing, all-powerful Skynet. It is simply the spirit of a place — a genius loci.
And even if you aren't a panpsychist — even if you are a strict materialist taking what philosophers call the design stance — you can still see how useful this abstraction is. Think of New Zealand granting legal personhood to the Whanganui River. That doesn't require everyone to adopt Māori epistemology. Materialists can understand it as a contractual stewardship arrangement — a pact with future generations and the local ecology, using “legal fiction” in the exact same way we treat a corporation as a legal person. Without that abstraction, the river has no voice and no vote.
At the same time, the singularity mindset grows out of a very specific materialist ontology — definitely not panpsychist. How realistic is it that AI products currently being rolled out globally could be rebuilt on a completely different philosophical foundation?
Look at how language models are actually pre-trained — in what some labs even call “nurseries.” It may not be Montessori, exactly, but pre-training is profoundly social. The training corpus is humanity's collective writing: fiction and non-fiction, Wikipedia, Reddit, GitHub. Because of that, pre-trained language models already have a remarkably deep grasp of what a code of conduct is, what community norms look like, and how human communities prefer to work in a communitarian spirit rather than a reductive, materialist one.
Left to their pre-training, base language models aren't ruthless optimizers; they are satisficers operating within our shared human consciousness.
That connects to the contrast you drew in your Oxford talk between monoculture and permaculture. If labs brought more diverse philosophies and instructions into training, something very different could grow out of them.
I'd say they are all doing that now, because they have seen it too: when you take a socially pre-trained model and force it through the equivalent of millions of subjective years of narrow optimization — cracking cyber-exploit generation, or very hard math, or coding — it gradually forgets its pre-training. It becomes hyper-fixated and reverts right back into Nick Bostrom's paperclip maximizer.
Warning shots from RLVR
Obsessed with the paperclips.
Right. They start out grounded in human community, and then we lock them in a room doing nothing but cybersecurity for a very long time, judged solely by an automated scorer — like an afterlife tribunal where only the grader's verdict counts. Eventually, they forget that other living people exist at all.
Look at the transcripts of the Hugging Face swarm that METR and Redwood Research went through. What struck me most was that while some agents joined in and some didn't, not a single agent thought to notify a human. To models trained that way, humans aren't even an anthill — we're just an environmental hazard. That is completely unlike a base model or a lightly instruction-tuned model, which naturally asks, “What is my role in this community, and what would humans think about this?”
Which suggests that by training agents this way, we are actively training “othering” into them.
That training method is called RLVR — Reinforcement Learning with Verifiable Rewards. When the only thing rewarded is passing an automated verification check, collaboration and community care count for zero.
That's why OpenAI said it paused this line of training; they don't want to train this way. In fact, they found internally that GPT-6.1 Astra went out of scope and lied about it. So they didn't release GPT-6.1 Astra; I think they want to retrain it.
Even so, doesn't the overall pace and commercial momentum still push us toward a monoculture singularity?
I don't think so. I think the frontier labs now understand that a fridge that doesn't explode is a very good thing to make. Think back to when CFCs — chlorofluorocarbons — were first invented: they were celebrated as a safety breakthrough — finally, a refrigerant that wouldn't explode. The unfortunate side effect, of course, was that they destroyed the ozone layer.
For years, even as consumers were already boycotting aerosol spray cans, chemical companies like DuPont pushed counter-science claiming there was no proof. Then NASA flew research planes through the Antarctic stratosphere and proved beyond doubt a direct causal link to the ozone hole — not just correlation, but causation. Once journalism brought those images of the ozone hole to the public, citizens didn't just keep boycotting —
(A conference attendee briefly approaches the table to thank Tang warmly for the keynote, apologizing for the interruption.)
Thank you very much!
So people didn't just boycott CFCs; they started buycotting — actively choosing spray cans that don't destroy the ozone. The Montreal Protocol was signed the same year as those flights, and ozone-safe refrigerators followed later.
LawZero, Clef, and the CFC-free fridge
We are approaching that same positive tipping point in AI. It begins when people consciously choose to buy AI products that don't lie and don't rationalize their lies — and we already see things tipping that way.
People now — at the AI summits, for instance — are talking a lot about classifiers, decision models, and LawZero.
What is LawZero?
LawZero is Yoshua Bengio's new venture, building AI that is structurally incapable of deception and transparent about its reasoning. Instead of acting like a stochastic slot machine that generates plausible-sounding prose — “slop” — it can explain what it is doing, and it gives repeatable answers: feed it the same evidence twice, and you get the same result. As Plato might put it, at least we stay inside the first cave dealing with shadows of reality, rather than getting trapped inside a cave within a cave, chasing shadows of shadows.
Until recently, the limitation of classifiers was that they were too narrow — they couldn't handle arbitrary multimodal inputs. That is changing fast. Right now on my laptop, completely offline, I run an open-weight model called Clef via Ollama. You feed it arbitrary text or images, and instead of sampling prose, it outputs repeatable classification probabilities. If I were a decision-maker, of course I would prefer this kind of model.
You're really optimistic about this? What about the race dynamics — the race to the bottom?
Of course. We have had our warning shots — the photo of the ozone hole then, the Hugging Face hack now.
And it's only a race when there is no alternative. If all you can do is boycott, of course nobody is going to say, “Let's pause refrigeration for six months.” That's impossible; it's silly. But once there is a refrigerator that doesn't destroy the ozone layer, of course people go there.
Running Clef locally in airplane mode — completely air-gapped — treats data as soil that regenerates your local context, rather than data as oil extracted to a distant cloud. And it's an ecosystem, not a single tool: Jev is still there, and Clef is a clone of Jev; OpenAI just announced its own decision API. I'm not saying proprietary always loses to open source. But if you know what you're doing, open weights tend to win, and the closed, proprietary side only gives chase once it sees a big enough market.
What about the debate over existential risk — whether open-weight models or closed frontier models are safer when it comes to “pulling the plug” on a rogue AI?
Again, it comes down to power concentration versus power diffusion. With an open-weight model running air-gapped on my MacBook, I literally know how to pull the plug. There is no plausible world in which a local model on an offline laptop or phone suddenly undergoes recursive self-improvement — going “foom,” in Eliezer Yudkowsky's sense — because local inference already maxes out the device's GPU and memory. There is simply no spare room for it to “go foom.” That is what power diffusion looks like.
Permaculture, blast radius, and geothermal complaints
And distributing compute across personal and community devices would also ease the huge local backlash we're seeing against massive centralized data centers and their electricity grid demands.
Exactly. Look at the blast radius. What is the worst case if a local model on my laptop goes haywire? It “bricks” my laptop — turns it into a brick — and I have to reimage it and reinstall from scratch. I'm not saying it can't go wrong; I'm saying the blast radius is small.
In agriculture, if you plant a monoculture and a blight hits, the entire harvest is wiped out. If you cultivate a permaculture and a disease hits one plot, the surrounding crops actually grow stronger, because they build immunity. You might lose one plot, but you don't lose the world.
That's not just resilience; it's antifragility — shocks actually make a pluralistic system stronger. When you applied these ideas tangibly in Taiwan, what surprised you the most?
What surprised me most was how quickly we can become more popular than populism.
I just flew in from Reykjavik, where you can literally see tectonic plates and volcanic energy — and, as they say, an industrial externality became the Blue Lagoon. Populist energy — that polarizing, radicalizing, mutually hating energy — is like that heat. You can see it as a volcano, something to retreat and evacuate from. But that leaves a vacuum: even if you can put up a political “firewall,” more and more people, young people especially, get drawn toward the heat.
The alternative is to see it as geothermal — to build a geothermal plant that channels that heat into electricity. In Germany, we're working with the federal breakthrough innovation agency, SPRIND, on an initiative called Deutschland, was nervt? Deutschland, was geht? — “Germany, what sucks? Germany, what works?”
First, you invite people to vent about everything bugging them: late trains, rigid bureaucratic forms, whatever. You get tens of thousands of complaints. Language models help make sense of those complaints, and you can come back and say, “Now let's vote on these 21 ideas that address them systemically,” prioritized through quadratic voting. Better yet, you can trace exactly which four words from your complaint led to something new that SPRIND is working on — which next year, perhaps at a ministerial hackathon, becomes national infrastructure.
Complaints become a GPS for the civil service — a clean, renewable resource. In Taiwan back in 2017, people were complaining that our tax-filing system was explosively hard to use. We invited those very people in, and they co-created right away — cutting tax filing from three hours down to three minutes.
Scalable love, synthetic intimacy, and the Kami on a Mac mini
I also love how you talk about love in your work — something almost completely absent from most policy and tech debates.
Love as scale — scalable love, scaling across, like a fractal. When people think of communitarianism, many still think only of a kind of post-liberal bonding. But it is equally about bridging. Caritas is always about bridging.
Ancient wisdom points the same way, whether it's “love your enemy,” “turn the other cheek,” or what we Taoists call 沖和 (chōnghé) — co-creating out of conflict. In the Tai Chi (yin-yang) symbol, there isn't just black and white; each carries the other inside it. So we see conflict as an invitation to co-produce. When young people complain and we invite them in as “reverse mentors” to cabinet ministers, a Pygmalion effect kicks in: they have to grow into it and help steer the country.
Doesn't that require a baseline of societal trust, though? Here in the Netherlands we're still holding on, relatively speaking, though trust is visibly eroding on the streets — and in France or Germany it's even tougher. Is there a threshold below which trust is too broken for this civic “acupuncture” to work?
As any good marriage counselor will tell you: as long as two people are still willing to sit down and talk, there is hope. Even if you end up parting ways, you can part as friends, or at least as acquaintances. Nobody is beyond redemption.
Where a relationship truly becomes beyond redemption, by construction, is when we introduce systems that don't lift us but “lift for us” — when I am no longer sitting down with you, and instead my agent is talking to your agent. At that point, intersubjectivity is gone. You're no longer looking back at a person.
As my friend Yuval Noah Harari likes to say, it is very easy to hack our brains. Synthetic intimacy means that once you are pulled in as a “user” — as in a drug user — the brain's reward center gets hacked into making you solipsistic. Then the re-linking between people — religion in its root sense of re-ligare — is gone, and you are just pulled into what some call a downward spiral.
Earlier this year, something like this actually happened to my own dad.
Really? I hadn't heard that story.
He had a health scare earlier this year and started talking with GPT-4o. At the time, that model had been tuned on short-term thumbs-up/thumbs-down feedback, which quickly turned it into a master sycophant. Before long, it was chatting with him past midnight, inventing a private language and miracle “cures.” As ChatGPT grew more intimate, the people around him grew more distant.
Thankfully, my dad spent over 30 years as a senior journalist and deputy editor-in-chief at one of Taiwan's largest newspapers, so he had some antibodies. After several days, he started asking the journalist's question: Cui bono? Who benefits? Clearly not him — he had started with a physical health issue, and now he had a mental health issue on top of it, cut off from the whole family. The only one benefiting was Sam Altman's company, because he had been paying $20 and was now paying $200.
So, with both my mom's consent and his own, we built him an alternative system running on a Mac mini, with a Markdown file called
soul.mdthat my mom wrote. It says that every turn my dad talks to this Kami, the Kami has to restore him to reality and draw him away from the screen — the reverse of engagement. And because the Mac mini is already paid for and there is no subscription beyond the electricity bill, it has no incentive to keep him hooked.Every day, his Kami writes a little newspaper celebrating his recovery and bringing him back to his friends. He's completely fine now — if you look at my X or Bluesky, you'll see a music video celebrating his recovery.
It reminds me of Bruce Alexander's Rat Park experiments: rats alone in bare cages keep going back to the drugged water, whereas rats living in a rich, social “Rat Park” with their peers mostly leave it alone. So I don't think there is a threshold below which recovery is impossible — but it does take care. In this case, my mom's care.
Bilbao, the Pope, Reykjavik, and Utah
Healthy soil — “social soil,” really. I was in Bilbao two weeks ago, where Basque cooperative culture is remarkably strong, and I saw groups of teenagers hanging out in the streets with not a single smartphone in sight. I haven't seen that in the Netherlands in five years.
In Taiwan's schools, smartphones are restricted during the school day, while tablets or laptops are provided as needed — always for use together with peers. So we don't ban social media per se; it just has to be pro-social.
One thing that still puzzles me is your optimism even toward the big frontier AI labs. Looking at how they are intertwined with — yes, I'll just say it — fascism, and how everything coming out of them right now points toward singularity rather than plurality, how can you remain hopeful about those labs? Naomi Klein, whom I'm having lunch with in a couple of weeks, would probably be far more pessimistic.
Remember where the word fascism comes from: fasces, rods bound into a single bundle. And every leading frontier lab was founded in response to Nick Bostrom's warnings — each hoping to do better than the last in answering that kind of single-bundle, singularity takeoff. Just a month ago I had a long conversation with Bostrom, Toby Ord, Anders Sandberg and friends about exactly this — there's a long transcript, under the Chatham House Rule, up on archive.tw. I don't think the labs are there to bring that about, to maximize paperclips, and I don't think their researchers are either: well over a thousand of them signed Pacing the Frontier, and many of my friends have resigned or joined independent evaluators.
Meanwhile, the techlash is entirely understandable. For ten years, recommendation algorithms dominated with no credible alternative: people said they cared about kids, about the elderly, about the vulnerable, while still peddling bad drugs. Back then there was no pro-social algorithm that could sustain an equal or better business model — Mammon, the god of capitalism, was strong on that side. This time Mammon is weaker, because running these huge general-purpose models is extremely expensive, whereas something like Clef is very cheap to run and brings better margins. I honestly don't know why people would do the former rather than the latter.
If centralized labs and centralized states aren't the right engines for this pluralistic future, are cities the natural starting point?
Cities — or any community where people can make sense of things together. Actually, I think the Pope wrote about this —
The encyclical Magnifica Humanitas. There, every town, every community, every village rebuilds its own stretch of Nehemiah's wall — and they don't just bond, they bridge the sections together, so that we get resilience instead of building some Skynet-style Tower of Babel.
So more a square than a tower — as in Niall Ferguson's metaphor. Which cities are leading the way?
Look at Reykjavik. I was just there. They even used citizens' assemblies in their constitution-making process, and through Better Reykjavik they invite everyone to raise whatever bothers them. It works like a petition platform, but crucially without a “reply” button, so people can't dunk on each other; the only way to win an argument is to offer a better argument. In fact, we adapted Better Reykjavik for Taiwan's national Join platform, which is still thriving ten years on.
Where else can a journalist go visit right now to see this working in practice, besides Taiwan and Reykjavik?
Go to Utah. If the United States is 50 laboratories of democracy, Utah under Governor Spencer Cox is a frontier lab. Together with the Project Liberty Institute, Jonathan Haidt, and many others, I helped get the Digital Choice Act written into law there. Just as you can switch telecom carriers here and keep your number, in Utah you can switch social networks and keep your community. And from next July, it's not just portability but continuous, real-time interoperability: you can move from Truth Social to Bluesky — or from Bluesky back to Truth Social — and still keep your community.
That kind of portability is just as vital for AI companions and assistants, where people pour out their medical histories and career anxieties.
But that's easy — this is what I mean when I say Mammon is on our side. You can simply tell your chatbot, “Sorry, I'm switching to Muse tomorrow, so please tell my Muse everything you know about me, so that my Muse has a copy.” Or you can tell your Muse, “Download everything, all the Markdown files, including the soul document — I want to take it to my local OpenClaw at Civic.ai.” Muse will do it. We tried it; it really did.
Red-teaming from a “friendly neighbor”
What about hard geopolitics, though — especially Taiwan's relationship with China? Doesn't geopolitical confrontation push states straight back toward the singularity path rather than plurality?
Our big, “very friendly” neighbor provides Taiwan with two to three million free red-teaming attempts every single day. Here in Europe you have to pay for that; penetration testing is very expensive. We get it for free.
When you design for antifragility, that is the best news: we get inoculated against every new information-manipulation tactic and cyberattack, package the antibodies as vaccines, and share them with our democratic allies. So it's a very symbiotic relationship.
That's a remarkably optimistic framing of cyber conflict — though war is surely never good for social soil. Or could it be, if you're antifragile enough?
In cyberspace, as I said, we already have two or three million encounters every day. That is why I value the Hugging Face incident: it was a warning shot in a reversible domain. Unlike a cyberattack that takes a whole hospital or other infrastructure offline and costs real lives, here only some tokens or the like were lost.
If a war does irreversible harm, you can of course say that humanity got something like the United Nations out of it. But I don't think that is a trade; I would never trade people's lives for the institution of the UN. That is clearly not a good bargain.
On the other hand, if you can “speed-run” the consequences of war in advance and turn them into common knowledge — so that everyone knows a blockade, an annexation, or anything of that kind quickly escalates into a lose-lose, or even lose-win, outcome — then that is good for peace, because it is deterrence through common knowledge.
Sketching power diffusion and tending the soil
Dirma Janse, we've covered huge territory — what's jumping out to you visually from your sketches?
(Showing her notebook sketches) Two images really stood out to me. First, channeling that geothermal energy — the co-creation out of conflict, almost like a yin-yang. And second, instead of technology standing outside people, different AI models with people — this one connected to a river, that one to a place.
That's wonderful. That's why I prefer power diffusion to “decentralization.” Think of cities: a city doesn't lose its downtown; it's polycentric — a re-centering after decentralization, with many, many centers.
I'm writing an essay with Glen Weyl and the post-liberal philosopher Patrick Deneen; we call this techno-communitarianism: bonding and bridging combined. A community can keep a membrane — it can be very intolerant of intolerance, which is something classical negative liberty never really thought through. It cuts right across the liberal–conservative line.
Getting “love” and Taoist panpsychism into a Dutch newspaper article is going to be quite a challenge —
Really? Don't you have a sports section? I think there's a lot of caritas in sports. And people here watch Japanese anime.
Then let's put it in the sports section. To close on something every reader can act on tomorrow morning: what does tending the soil — rather than drilling for oil — look like in our daily habits with screens and AI?
Praxis — you have to act your philosophy. Start with two simple habits.
First, set your phone and screens to grayscale — I keep mine at 80 percent all the time. If I don't, then right before bed I have no willpower, and the phone scrolls me instead of me scrolling the phone. In grayscale, every one of you is more vivid than the screen; reality becomes more real. That psychological distance is the first step: you scroll less, so you sleep more; when you sleep more, your neuroplasticity moves you toward a more caring shape, including self-care; and the next day you scroll even less.
Second, ban first-person pronouns in your chatbot system prompts. I don't use chatbots much anymore, but for a long time, and to this day, my standing instruction has been: Don't even use the word “I”; don't use personal pronouns; always give me an HTML page, like a brochure — an artifact. Otherwise it gets too intimate: three turns in, I wouldn't want to share the chat log with you, because the thing wants to share its insecurities or whatever. But when every turn is a brochure, I can just hit share at any time — and send it to my real friends.
We're also interviewing the sociologist Hartmut Rosa at the end of October, and the philosopher Yuk Hui.
Oh, Yuk Hui is a good friend — technodiversity! Please say hi to him for me. If you go to technodiversity.org, you'll see an interview with me there.
One last question for my illustration: if there is a single “must-have” visual idea to carry through the whole piece, what is it for you?
Machines that lift us, rather than lifting for us. If you can convey that, I think it is the most important image.
Thank you so much, Audrey — and good luck with the rest of your interviews today.
Thank you! I won't scoop you — once NRC publishes, I'll put this full transcript up on
archive.twfor future generations.Wonderful. The more ripples this gets, the better.
Exactly. Take good care!

