Vollständiges Transkript von Snow on demand

21 00:03:08.000 --> 00:03:18.980 The catch is that the AI-assisted stories were more similar to each other. Each person got better, and the group as a whole got more uniform. Everyone skied on the same snow. 22 00:03:19.640 --> 00:03:22.200 That's the snow cannon. Perfectly even. 23 00:03:22.820 --> 00:03:26.680 Is there a way to use the cannon that doesn't make everything even? 24 00:03:27.340 --> 00:03:45.960 I think so. Don't ask it for the best idea – ask it to irritate you. Give me the strongest argument against my plan. Show me how a beekeeper would solve this. Combine my product with something from a completely different field. We do that in workshops, and suddenly the ideas stop looking alike. 25 00:03:46.520 --> 00:03:58.680 That's actually close to Rosa. He says resonance needs something foreign, something that doesn't simply confirm you. A machine that irritates – I hadn't considered that. That's a point for you. 26 00:03:59.260 --> 00:04:03.100 I'm writing that on my office wall. "Krzysztof gave me a point." 27 00:04:03.880 --> 00:04:06.880 Half a point. It still depends on who uses it. 28 00:04:07.480 --> 00:04:12.940 Krzysztof, the waiting – is that just romanticism, or does it actually do something? 29 00:04:13.540 --> 00:04:50.040 It does something, and this is a very old idea. In nineteen twenty-six, an English social psychologist, Graham Wallas, described four stages of creative thinking. Preparation, incubation, illumination, verification. Incubation is the strange one – you stop working on the problem, and something keeps working on it without you. And modern research confirms it. A big review of more than a hundred studies found that setting a problem aside before coming back improves the solutions. Especially for problems that need a new angle. 30 00:04:50.660 --> 00:04:57.220 I believe that. My best ideas come on the train, and not when I'm prompting. When I'm looking out of the window. 31 00:04:57.820 --> 00:05:09.240 The essay ends with a question that I think is made for you, Francesca. Could an AI support incubation – for example by deliberately waiting before it answers? 32 00:05:09.900 --> 00:05:29.080 My guess: yes, but not by waiting – people would just switch to another tool. I'd design it differently. A "sleep on it" mode. You describe your problem in the evening, and the machine doesn't answer. It asks you two questions back. And in the morning it shows you your own notes first, and only then its ideas. 33 00:05:29.760 --> 00:05:34.940 That's – uh – that's paternalistic. A machine that decides you need to wait. 34 00:05:35.560 --> 00:05:42.660 You choose the mode. Nobody forces you. Like the alarm clock – you set it in the evening because you know who you'll be in the morning. 35 00:05:44.280 --> 00:05:49.740 Hm. All right. The alarm clock I accept. I set mine every night and hate it every morning. 36 00:05:50.420 --> 00:05:51.520 So you'd use it? 37 00:05:52.200 --> 00:05:58.220 I'd use it. I wouldn't trust the

company that sells it to keep the waiting in. Waiting doesn't sell. 38 00:06:03.132 --> 00:06:11.832 Second topic, and it's about reading. Or not reading. Krzysztof, you have shelves of books you haven't read, I assume? 39 00:06:12.412 --> 00:06:29.032 Walls. Rosa has a word for this too – he describes how we buy books, download articles, save links, and feel that we somehow own the knowledge. But owning a book is not reading it. And reading is not the same as making it your own. 40 00:06:30.512 --> 00:06:36.452 My browser has two hundred open tabs. Each one is a small promise to my future self. 41 00:06:37.012 --> 00:06:38.912 And your future self ignores you. 42 00:06:39.452 --> 00:06:42.412 My future self has two hundred and one tabs. 43 00:06:43.092 --> 00:06:48.572 Now with AI, we can have everything summarised. Is that the solution or the problem? 44 00:06:49.272 --> 00:07:13.072 Both, honestly. There's a set of experiments with more than ten thousand participants – people learned about a topic either from an AI summary or by clicking through ordinary search results. Same facts. The summary group ended up with shallower knowledge, and when they gave advice to others afterwards, it was shorter and more generic. The effort of searching made the knowledge stick. 45 00:07:13.652 --> 00:07:15.232 The effort is the point. 46 00:07:15.852 --> 00:07:44.512 But here I have to defend summaries. Herbert Simon said it in nineteen seventy-one: a wealth of information creates a poverty of attention. Nobody can read everything. Scholars have always used abstracts, reviews, someone else's notes. And for many people the alternative to a summary is not the original – it's nothing. My mother reads English slowly. Without translations and summaries, most of what I work on would be closed to her. 47 00:07:45.092 --> 00:07:52.012 That's fair. I've read Rosa in German, which is a privilege. Most of my students read him in summaries of summaries. 48 00:07:52.652 --> 00:07:53.912 So where's the line? 49 00:07:54.572 --> 00:07:58.932 My line: summaries for choosing what to read – not for knowing. 50 00:07:59.652 --> 00:08:28.172 I'll take that line. And here's something we actually do, which answers one of the essay's questions about reading rituals. Some of our meetings now start with ten minutes of silence. Everyone reads the same short document – on paper if they like. Then we talk. Amazon has done this for years with their six-page memos. At first it feels like school. After a month, people say it's the only time they actually read anything at work. 51 00:08:28.752 --> 00:08:40.572 There's a question in the essay that worries me professionally. If everyone can produce text that looks expert – how do I, as a journalist, recognise an expert? 52 00:08:41.252 --> 00:08:57.452 My guess: the oral exam comes back. Ask them to explain it without the screen. Ask the follow-up question, and the one after that. Text can be generated. A second follow-up question in a live conversation cannot – not yet. 53 00:08:58.052 --> 00:09:01.512 Which is, by the way, exactly what you do on the radio. 54 00:09:02.472 --> 00:09:07.052 Thank you. I always suspected radio would be useful again one day. 55 00:09:07.792 --> 00:09:13.872 As a radio person, I find that almost moving. Ten minutes of silence in a meeting. 56 00:09:14.432 --> 00:09:18.172 It's a small act of resistance. I'd put it in the constitution. 57 00:09:22.552 --> 00:09:28.532 Third essay, and my favourite title. The red pullover. Krzysztof? 58 00:09:30.172 --> 00:09:55.792 It's Rosa's example for something called appropriation – making something your own. A pullover your grandmother knitted for you, or one you knitted yourself with all the mistakes, is yours in a way a bought one isn't. Even if the bought one is warmer. You fought with it. Ideas are the same. An idea you wrestled with – you know where its weak spots are, you know why it matters. 59 00:09:56.372 --> 00:09:58.512 And here comes the brain scan study, right? 60 00:09:59.112 --> 00:09:59.792 Here it comes. 61 00:10:00.432 --> 00:10:20.092 The MIT Media Lab had students write essays over several sessions – some with ChatGPT, some with a search engine, some with only their brain. They measured brain activity. The ChatGPT group showed the least engagement. And afterwards many of them couldn't quote a single sentence from the essay they had just written. 62 00:10:20.671 --> 00:10:22.252 They couldn't quote themselves. 63 00:10:22.872 --> 00:10:40.572 They couldn't quote themselves. But – and this is important – it was a small study, fifty-four students, timed essays like in a school exam. And there was a group that first wrote alone and only then used the AI. Those people were very engaged – actually more than before. 64 00:10:41.212 --> 00:10:42.532 So the order matters. 65 00:10:43.132 --> 00:10:46.792 The order matters. First the pullover, then the machine. 66 00:10:47.392 --> 00:10:50.392 There's a famous experiment with furniture, isn't there? 67 00:10:51.052 --> 00:11:04.532 The IKEA effect. Michael Norton and colleagues. People who assembled an IKEA box themselves valued it much more than the same box already assembled. Even with crooked shelves. Labour leads to love. 68 00:11:06.172 --> 00:11:09.152 I have a very crooked IKEA shelf that I love deeply. 69 00:11:09.792 --> 00:11:15.112 Francesca, isn't there a counter-argument – that we've always thought with tools? 70 00:11:15.772 --> 00:11:39.452 Yes, and it's a strong one. The philosopher Andy Clark – he's the one who came up with the "extended mind" in the nineties – says our basic nature is to think with things outside our heads. Notebooks, calculators, maps. Nobody says your diary makes your thoughts less yours. He even describes a model trained on his own writing, whose answers start to feel like his own voice. 71 00:11:40.172 --> 00:11:47.792 That's exactly what frightens me. A machine that sounds so much like you that you can't tell what

you thought and what it thought. 72 00:11:48.352 --> 00:11:59.692 Or a machine that sounds like you because it learned from everything you ever struggled with. The struggle doesn't disappear, it moves – from writing sentences to choosing, cutting, contradicting. 73 00:11:59.692 --> 00:12:00.972 If you still do the cutting. 74 00:12:01.532 --> 00:12:04.032 If you still do the cutting. Okay. Agreed. 75 00:12:04.680 --> 00:12:22.740 Let me add something from my own job. I learned interviewing by doing it badly for years. My first interviews were terrible – I asked three questions at once, I interrupted at the wrong moment. Nobody could have summarised that for me. I had to knit that pullover myself. 76 00:12:27.320 --> 00:12:40.980 The fourth essay is the most personal one. Chatbots as company. Someone to talk to who is always available, always patient, never tired. Krzysztof – is that an answer? 77 00:12:41.680 --> 00:13:05.140 Rosa would say: it's an echo. For Rosa, resonance means that something answers you with its own voice – and that both sides are changed by the encounter. A friend who disagrees with you, a book that resists you. A chatbot that is designed to be agreeable gives you your own voice back, only nicer. And nothing is at stake for it. 78 00:13:06.540 --> 00:13:22.560 I want to be careful here, because people really do feel less lonely. There are experiments from Harvard – a fifteen-minute conversation with an AI companion reduced loneliness about as much as a conversation with another person. More than watching videos. 79 00:13:23.200 --> 00:13:24.520 For fifteen minutes. 80 00:13:25.140 --> 00:13:37.680 In a week-long study, too. And think about who's on the other end. An old man in a village whose children live in another country. The alternative to the chatbot isn't a friend. It's silence. 81 00:13:38.420 --> 00:13:45.800 That – I don't dispute. The alternative is often silence. That's an indictment of us, not of the chatbot. 82 00:13:46.440 --> 00:13:47.300 And the risks? 83 00:13:47.940 --> 00:14:14.699 OpenAI and MIT looked at this together last year – millions of conversations and a four-week trial with nearly a thousand people. For most people, loneliness went down slightly. But the people who used it the most – the really heavy users – were lonelier and more dependent. And the researchers say themselves they can't tell which way round it goes. Do lonely people use it more, or does it make you lonely? 84 00:14:15.260 --> 00:14:36.380 The American psychologist Sherry Turkle has been warning about this for fifteen years. "Alone together", she called it – we expect more from technology and less from each other. Her worry was never that machines are bad company. It was that they're easy company. And easy company makes the difficult kind look unnecessary. 85 00:14:36.920 --> 00:14:45.300 There's an open question in the essay I'd like to try. Could a chatbot be designed to bring people towards other people, instead of replacing them? 86 00:14:45.880 --> 00:14:46.380 Go on. 87 00:14:47.060 --> 00:15:04.140 My guess: yes, and it's not even hard. Imagine it notices you've talked to it every night for two weeks about the same worry. And one evening it asks: "Who else could you tell this to?" Not as a lecture. As a real question. Maybe it even helps you write the message to your sister. 88 00:15:04.740 --> 00:15:11.600 Like a good barman. He listens, he pours you one more, and at some point he says: go home, call your wife. 89 00:15:13.280 --> 00:15:14.900 Exactly like a good barman. 90 00:15:15.500 --> 00:15:19.600 And would a company build a barman who sends its best customers home? 91 00:15:20.780 --> 00:15:28.380 That's – that's the business question. Some would, if the law or the reputation made them. Most wouldn't by themselves. 92 00:15:29.060 --> 00:15:33.520 So again we agree on the design and disagree on whether anyone will build it. 93 00:15:34.120 --> 00:15:35.760 That seems to be our whole season. 94 00:15:36.520 --> 00:15:50.580 I'll admit something. My son talks to a chatbot about his exams. Not instead of me – I hope – but at two in the morning, when I'm asleep. And I don't know whether to be relieved or worried. 95 00:15:51.280 --> 00:15:53.780 Both? I think both is right. 96 00:15:54.460 --> 00:15:58.300 I would ask him one thing. Does it ever tell him he's wrong? 97 00:15:59.440 --> 00:16:02.940 I'll ask him. He'll say I'm being a sociologist. 98 00:16:03.560 --> 00:16:06.400 That's the worst thing a son can say to a mother. 99 00:16:10.980 --> 00:16:17.760 Last essay, and it's the one about judgment. Krzysztof, Rosa's newest book has a strange title. 100 00:16:18.460 --> 00:16:51.880 "Situation and Constellation". The idea is this: in a situation, you have to judge. You look at the person in front of you, you weigh things, you decide. In a constellation, the decision has been made in advance – by a form, a rule, a scoring system. You only have to tick the box. And generative AI is a constellation machine for almost everything. How to answer the email, how to give feedback, how to word the diagnosis. The next step is always already suggested. 101 00:16:52.420 --> 00:17:15.180 And sometimes the suggestion is better than the judgment. That's the uncomfortable fact here. There's a famous study with American bail decisions – Jon Kleinberg and colleagues. A model using only the information the judges had could have reduced crime considerably without jailing more people. Or kept the same crime rate and jailed far fewer. 102 00:17:15.820 --> 00:17:17.740 I know the study. It's very good. 103 00:17:18.260 --> 00:17:33.980 And human judgment is noisy. Kahneman and two colleagues wrote a whole book about it – "Noise". The same case gets very different decisions depending on the judge, the day, whether it's before or after lunch. Discretion is also where prejudice lives. 104 00:17:34.580 --> 00:17:57.720 Agreed, and Rosa admits that rules can serve fairness. But

there's an irony – an engineer, Lisanne Bainbridge, described it in nineteen eighty-three, about pilots and power plants. The "ironies of automation". When the system takes over the routine, the human is left with the rare, difficult cases – and they're out of practice exactly for those. 105 00:17:58.380 --> 00:18:03.400 The pilot who hasn't flown by hand for months, and then the autopilot gives up. 106 00:18:03.980 --> 00:18:06.240 Exactly. Judgment is a muscle. 107 00:18:06.936 --> 00:18:39.096 And let me give you an example of a constellation from my own company, because I don't want to sound like the algorithm's lawyer. Last year many of our managers started drafting performance reviews with AI. Perfectly reasonable – they're long, they're painful to write. And after one cycle, HR noticed something: all the reviews sounded the same. Same structure, same phrases, "demonstrates strong ownership". You couldn't tell from the text whether the manager had actually watched the person work. 108 00:18:39.736 --> 00:18:40.396 What did you do? 109 00:18:41.056 --> 00:18:54.416 We changed one thing. The manager has to write the first three sentences alone – what this person did this year that nobody else did. Then the machine can help with the rest. Suddenly the reviews were different again. 110 00:18:54.976 --> 00:18:58.076 The situation came back. In three sentences. 111 00:18:58.676 --> 00:19:03.456 In three sentences. It's never the whole document. It's the first move. 112 00:19:04.116 --> 00:19:15.676 The essay asks a hard question, and I'd like you both to try. Where should societies insist on human judgment even when the algorithm is statistically better? 113 00:19:16.336 --> 00:19:36.416 My hunch: wherever the person affected has the right to look someone in the eye. A sentence in court, a child's school report, a decision about a refugee. Not because the human is more accurate. Because a judgment you can't argue with, can't question face to face, isn't a judgment – it's an administration. 114 00:19:36.996 --> 00:19:49.836 That's a beautiful principle and I don't think it survives contact with reality. If the algorithm sends fewer innocent people to prison, how do I explain to the extra people in prison that it was for the sake of eye contact? 115 00:19:50.896 --> 00:19:56.096 That's the hardest objection. I – yes. I don't have a clean answer to that. 116 00:19:56.696 --> 00:20:14.136 My guess would be different. Let the algorithm propose, and make the human write down why they follow or don't follow it. Every time. Then the judgment stays exercised, and the human can still look you in the eye – with a reason, not with "the system says". 117 00:20:14.716 --> 00:20:20.536 A discretion budget. Hm. I could live with that – as long as deviating isn't punished. 118 00:20:21.176 --> 00:20:32.516 That's the real question, yes. In most companies, the person who overrules the system and is wrong gets fired. The person who follows the system and is wrong gets a shrug. 119 00:20:33.176 --> 00:20:35.816 That's a sentence I'll be thinking about for a while. 120 00:20:36.696 --> 00:21:21.436 Let me try to gather today. You both agree that availability gives us real things – ideas when we're stuck, summaries when we have no time, company when there's no one, and sometimes better decisions. And you both agree that it takes something: the waiting where ideas grow, the effort that makes knowledge ours, the friend who disagrees, the practice of judging. Where you differ is what to do about it. Francesca would design the machine to give some of it back – a "sleep on it" mode, the dissent written down, the reason for every override. Krzysztof doubts that what doesn't sell will survive in the product. 121 00:21:22.076 --> 00:21:24.276 I doubt it. But I'd like to be wrong. 122 00:21:24.876 --> 00:21:25.996 I'll send you the beta. 123 00:21:26.636 --> 00:21:39.236 What stays open is a question I'll leave with you. Think of the last idea you were truly proud of. How long did you have to wait for it – and would you have got it if you hadn't had to? 124 00:21:39.916 --> 00:21:56.716 The essays and all the sources are on ideation.zone. Next time is the end of our season: predicted futures, mindfulness as an accelerant, and happiness at the edge of control. Krzysztof, Francesca – thank you. 125 00:21:57.376 --> 00:21:58.456 Thank you, Elisabeth. 126 00:21:59.056 --> 00:22:00.856 Grazie – see you next time.