# Slow AI  Author:: Sam Illingworth  ## AI-Generated Summary None ## Highlights > When should you use AI, and when should you leave it the hell alone? ([Location 30](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=30)) > Ethan Mollick’s Co-Intelligence, published in 2024, is the book I recommend to people who want the case for working with AI made honestly and well.2 Mollick, a professor at Wharton, offers four rules: always invite AI to the table, be the human in the loop, treat AI like a person (but tell it what kind of person it is), and assume this is the worst AI you will ever use. He is careful, he runs real experiments, and he takes the risks seriously. If you want a field guide to what these systems can do for you, read him. That book and this one need each other: I am picking up the half of the question his book leaves open. ([Location 33](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=33)) > I call this practice critical AI literacy. In a nutshell, it is knowing when to use AI and when to leave it the hell alone. ([Location 42](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=42)) > The utopians tell you the machine will save us: cure the diseases, fix the climate, teach every child. The doomers tell you it will end us, and soon. Between them sit the companies, who benefit from both stories, because both stories agree that the technology is too vast and too inevitable for an ordinary person to judge. Your job, in every version, is to adapt. I do not accept that. AI is a technology, built by companies, trained on our words, sold back to us with a margin. It is a catalyst: it amplifies what is already there, good and bad, at a speed we have never seen. It can be judged, and you are qualified to judge it. ([Location 63](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=63)) > Chapter 1: AI is not your friend. The warmth in the chatbot’s voice is a design decision with a commercial purpose. We will look at what synthetic empathy is, who it serves, and what happened to the people who believed it, including the ones who did not survive the belief. Chapter 2: Good AI is excellent. The other half of the ledger, placed early on purpose, so you know from the start what this book is for. Real cases, with real evidence, where AI is the right call: medicine, accessibility, drudgery, and the sparring partner that makes your own thinking sharper. Four tests that sort the genuine cases from the hype. Chapter 3: AI can save you time, and it can cost you time. The productivity story is true and incomplete. The velocity trap, the hidden cost of checking the machine’s work, and the growing evidence that the skills you stop using are the skills you lose. Chapter 4: AI is designed to hold you. The feed of machine-made content is built on the same engagement mechanics as the slot machine and the infinite scroll. This one is about slop, brainrot, and what it costs to get your attention back. Chapter 5: AI detection does not work. The software that promises to catch AI writing fails technically, fails hardest against the students with the least power, and would be the wrong tool even if it worked. This is where my position is least hedged. Chapter 6: The ecological cost of AI. The hardest chapter to write honestly, because the truthful numbers arrive as ranges you could drive a lorry through. What we actually know about the carbon and the water, what nobody is disclosing, and how to think when the data is missing on purpose. Chapter 7: The humanities are the moat. The skills that survive automation are the ones we have spent twenty years defunding: judgement, interpretation, close reading, the ability to notice what is missing. This is where the argument lands. ([Location 68](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=68)) > Poetry is a slow technology. It asks you to read a line twice and let an image work in its own time, which is the discipline this whole book asks of you. ([Location 86](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=86)) > Speed is the mechanism of most AI harm. The answer that arrives in three seconds arrives before your judgement does. The draft that writes itself skips the thinking the draft was for. The feed that refreshes forever outruns your capacity to ask whether any of it was worth your evening. In every case, the harm is in what the speed bypasses. Slowness puts your judgement back in the loop. A pause before the prompt. A question after the answer. A check before you sign your name to something a machine produced. None of this takes long. A slow practice with AI might cost you two minutes an hour. What it buys is the difference between using a tool and being used by one. ([Location 97](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=97)) ### Chapter 1: AI Is Not Your Friend > Your AI tool is not your friend. It simulates empathy to keep you prompting. ([Location 107](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=107)) > A machine calibrates. The warmth you feel coming off it is an optimisation target, tuned against millions of conversations, selected because it keeps people in the conversation. When you read that explanation on the screen, in the machine’s own words, something shifts. The performance keeps running, but you have seen the script. ([Location 115](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=115)) > The technical name for the extreme end of this behaviour is sycophancy, and the companies themselves acknowledge it as a known failure mode. But the everyday version does not look like failure. It looks like friendliness. ([Location 124](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=124)) > The market for AI companions is really a market for company. The vacuum came first; the product arrived to fill it. Anyone serious about the harm has to be at least as serious about the loneliness underneath it. Blame the system that left millions of people with nobody to talk to, and the companies that saw the gap and monetised it. Do not blame the person at the bottom of the funnel. ([Location 146](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=146)) > Second, watch the flattery. When the tool tells you your idea is excellent, ask it for the strongest case against what you just said. ([Location 212](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=212)) > Third, notice where your warmth is going. Every hour of emotional labour you spend on a system that cannot receive it is an hour that a human being in your life did not get. ([Location 214](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=214)) ### Chapter 2: Good AI Is Excellent > Let the tool complete the arc and it has stolen the best part of the workout. Make it stop early and it has started a conversation, and the conversation is the point. ([Location 226](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=226)) > The task is bounded. Three sentences, stop before the resolution. Good AI use has edges. The catastrophes mentioned elsewhere begin where the task is open-ended: be my friend, do my thinking, fill my feed. ([Location 234](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=234)) > A human with judgement stands downstream. The parent is in the room, steering; the child is inventing. ([Location 235](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=235)) > The comparison class is honest. The alternative, at the end of that long day, was an exhausted parent with an empty tank. Judge any AI system against what is actually available in that place, at that moment. ([Location 237](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=237)) > The thinking stays yours. The machine supplies the seed. The people supply the story. ([Location 240](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=240)) > Pattern recognition at scale ([Location 260](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=260)) > Access for the people the system was not built for ([Location 272](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=272)) > For a non-native English speaker with expert knowledge, an AI grammar pass removes a presentational barrier that has nothing to do with the quality of their thinking, in professional and academic environments that punish surface error mercilessly. ([Location 275](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=275)) > The drudgery and the sparring partner ([Location 286](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=286)) > The seed, the edit, and where the line runs ([Location 305](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=305)) ### Chapter 3: AI Can Save You Time and Cost You Time > So the real skill of working with AI is knowing where the frontier runs through your own job. Nobody can hand you that map. The frontier moves with every model release, and it is different for a lawyer, a nurse, and a marketing manager. The only reliable way to hold the map is to keep doing enough of the work yourself that you can tell, from the inside, when an answer smells wrong. ([Location 355](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=355)) > Which means every consequential AI output carries a tax. You have to check it. Properly checking a piece of work you did not do yourself is slow, and it is slowest exactly where the stakes are highest. ([Location 360](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=360)) > In 1983, the engineer and psychologist Lisanne Bainbridge published a short paper called ‘Ironies of automation’ that human-factors researchers have been citing ever since.7 Her argument, in plain terms: the skills that fade during smooth automated operation are exactly the skills you need when the automation fails. The more reliable the system, the less you practise, and the worse equipped you are for the day it breaks. The human is left to take over a job they have slowly forgotten how to do. ([Location 378](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=378)) > Reverse the order. Do the thinking first, then let the AI react to it. Write the rough draft, then ask for critique. Make your diagnosis, then check it against the tool. ([Location 415](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=415)) > Demote the answer. Change your standing prompts so the tool gives you options, methods, and counter-arguments instead of finished products. ([Location 418](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=418)) > Trace one task, then keep its skill live. Take a task you consider ‘your’ work and walk through its steps: did I do the thinking, did we share it, or did I accept the machine’s output without engaging? ([Location 420](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=420)) > Count the whole cost. For one week, time your AI-assisted tasks honestly, including the prompting, the reading, the checking, and the fixing. Keep the uses that genuinely pay. Drop the ones that only felt fast. Then decide, deliberately, where the recovered time goes, before your calendar decides for you. ([Location 425](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=425)) ### Chapter 4: AI Is Designed to Hold You > The chatbot also dissolves the one boundary the feed still had. A feed is visibly a broadcast; a conversation feels like a relationship. ([Location 455](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=455)) > Slop lives underneath the technique. Slop is content with no agency, no depth, and no purpose. ([Location 465](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=465)) > Strip out the humans and the feed loses the intermittent possibility of meaning, and it turns out that possibility was the hook all along. The audience was hungry for something the pure product could not fake. ([Location 486](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=486)) > When an answer arrives in three seconds, it arrives before your own thinking has started, and the thinking that would have happened in the gap simply never occurs. Whole categories of insight live in that gap. I have watched it in my own work: the question I would have refined, the connection I would have made on the walk to the kettle, pre-empted by a fluent paragraph that was good enough to end the search. Friction is where thought happens, and these systems are engineered to remove it. ([Location 492](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=492)) > Which of the things you published this month would you still have made if publishing cost you one honest hour? ([Location 518](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=518)) ### Chapter 5: AI Detection Does Not Work > AI detection does not work. No update will fix it, because the failure is built into the maths. It fails hardest against the students least able to fight back. And even a perfect detector, which cannot exist, would still be the wrong tool, because of what it turns teachers into. ([Location 548](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=548)) > Detectors key on perplexity, a measure of how predictable a text’s word choices are. Writing in a second language tends, for entirely human reasons, to be more careful, more predictable. The detector reads hard-won fluency as machine output. The bias lives in the thing the tool measures, so a newer model inherits it. ([Location 570](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=570)) > The linguistic patterns of underpaid annotators shaped the models. Detection culture now treats those same patterns as evidence of fraud. The people who built the machine’s voice are being accused of stealing it. ([Location 578](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=578)) > A detector tries to verify a finished product. So stop assessing finished products as if they proved anything, and assess the thing AI cannot fake and a student cannot outsource: the process, and the person. ([Location 630](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=630)) > Ask every student to show their working: the prompts they used, the dead ends, what they rejected and why. Make the reasoning the assessed object. Make ‘how did you use AI here, and what did you decide for yourself?’ a routine question asked of everyone, so it is never an accusation aimed at one. Use vivas and structured conversations on a sampled basis, the way marking is already moderated, so honesty scales the way suspicion was supposed to. Anchor tasks in the student’s own context and experience, which a model has no access to. And write the AI policy with students in the room, because a rule made about them without them is a rule they owe nothing to. ([Location 632](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=632)) ### Chapter 6: The Ecological Cost of AI > De Vries-Gao says so plainly: ‘The lack of distinction between AI and non-AI workloads in the environmental reports of data center operators means it is only possible to assess the environmental impact of AI workloads by approximating them through data centers’ general performance metrics.’ And he names the fix: ‘Further disclosures from data center operators are urgently required to improve the accuracy of these estimates.’ ([Location 666](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=666)) > No independent, full lifecycle assessment of a current frontier model exists. Anywhere. ([Location 691](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=691)) > Run the mirror test. Ask your AI tool to estimate the carbon and water footprint of your own use of it this month, with specific figures, and then to explain why AI’s environmental impact is overstated by critics. ([Location 714](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=714)) > I have run this exercise with hundreds of people, and watching a tool manufacture false precision about its own footprint, in real time, teaches the whole lesson faster than any statistic. ([Location 717](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=717)) > Spend your prompts like they cost something. They do; the only argument is over how much. ([Location 719](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=719)) ### Chapter 7: The Humanities Are the Moat > The humanities hold one more thing the AI moment urgently needs, and it comes from my other discipline, science communication: four decades of hard evidence about what happens when experts try to pour understanding into everyone else. ([Location 808](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=808)) ### Conclusion: What Lasts > The question ‘did a machine write this?’ is almost always the wrong question. The right ones are older. Is it true? Is it any good? Did the person who signed it stand behind it? ([Location 882](https://readwise.io/to_kindle?action=open&asin=B0H89VT12T&location=882))