AI, Work, and the Lives We Will Lead

Imani 0:06 Imagine your first Monday at a new job. You have your laptop, your coffee, and a question you are slightly embarrassed to ask. Now imagine that job was never advertised. Nobody was fired. There was simply no opening. That may be one of the most important sounds of the AI transition. Silence.

Miles 0:26 That is a much less cinematic opening than a robot taking somebody's desk. And probably a more useful place to start. Because the question isn't only, can a machine do my work? It's, who gets invited to learn the work in the first place?

Imani 0:41 Welcome to The Quiet Rewrite. I'm Imani.

Miles 0:45 And I'm Miles. Today, we're looking underneath the demonstrations. At jobs, robots, power grids, trust, and the ordinary decisions that could make an extraordinary technology either liberating or deeply unequal.

Imani 1:01 Our reporting is current to October ninth, twenty twenty-six. We will separate measured changes from forecasts, and flag our own scenarios. There is no honest countdown clock to the end of work. There are, however, some signals worth taking seriously.

Miles 1:17 And a few surprises. Including this one: making expertise cheaper could open doors for millions of people, while making the doorway into a profession harder to find. Both things can happen together.

Imani 1:32 Let's start there. If someone is choosing a career now, what does the evidence actually say?

Miles 1:37 The International Labour Organization's twenty twenty-five assessment put about one in four workers worldwide in an occupation with some exposure to generative AI. Clerical work was most exposed. But exposure measures tasks a system might affect. It does not mean a quarter of workers are about to lose their jobs.

Imani 1:58 A job is a bundle. The spreadsheet, the awkward conversation, the exception nobody documented. Automating one part changes the bundle. It doesn't tell you how many people a firm will employ afterward.

Miles 2:12 Right. But there is a worrying early signal. Stanford researchers using American payroll data through June twenty twenty-six found a nineteen percent relative employment gap for workers aged twenty-two to twenty-five in highly exposed occupations, compared with the trajectory of less-exposed peers. The change was primarily weaker hiring, rather than more people leaving.

Imani 2:35 Relative is doing important work there. That's not nineteen percent of all young people losing their jobs. And the researchers describe early patterns, not a clean causal estimate. Education and other economic forces complicate the picture.

Miles 2:50 Exactly. Our concern is what happens if that pattern persists. A firm can keep its experienced people and quietly reduce junior intake. An applicant experiences hundreds of unanswered applications. The firm experiences an efficiency improvement. Those are two descriptions of the same possible transition.

Imani 3:10 And then, five years later, somebody asks where the experienced people are going to come from.

Miles 3:16 Yes. Reading the messy documents and answering the routine tickets are also how people learn judgment. If we remove the practice, we need to redesign the apprenticeship. Paid supervised work. Real responsibility that grows. A training budget is not the same thing as another video course somebody watches after their shift.

Imani 3:38 So the first question for an employer is not just, how many hours did we save? It's, what path into competence did we remove, and what did we build in its place?

Miles 3:52 Now, the counterweight. A survey of nearly six thousand senior executives across four countries, reported by the National Bureau of Economic Research in twenty twenty-six, found more than ninety percent reported no employment effect from AI over the previous three years. That's self-reported company experience, not proof that nothing is changing.

Imani 4:12 How do you hold that next to the young-worker finding without picking whichever result suits your argument?

Miles 4:19 Different samples. Different measures. A company-wide average can miss a shrinking doorway into one occupation. Adoption also takes time. Buying an assistant is easy. Rebuilding a workflow, fixing the data, and deciding who is accountable when it fails are harder.

Imani 4:38 But there are real gains, not just promises. The published study called Generative AI at Work found about fifteen percent more customer-support issues resolved per hour with an AI assistant, on average. It studied a particular deployment. That's useful evidence, not a universal productivity multiplier.

Miles 4:58 And here is the bargaining question. Imagine a team can handle the same workload in fewer hours. Management could shorten the queue, expand the service, reduce headcount, raise expectations, or share the time savings. The software doesn't choose which of those futures workers get.

Imani 5:14 People often say the new jobs will arrive. Some will. But a job created in another city, requiring another qualification, doesn't automatically pay the rent of the person displaced today.

Miles 5:30 Nor does lower cost guarantee more employment. If demand expands enough, you might need more people. If it doesn't, fewer. Routine document processing, basic support, and repeatable digital production face pressure for that reason. Entire professions aren't a single switch you flip off.

Imani 5:48 The thing to watch is the bargain around productivity. Who gains time? Who gains money? And who is asked to absorb the uncertainty?

Miles 6:03 The next shift is from a system that suggests an answer to one that takes a sequence of actions. Find the records. Compare the options. Update the system. Follow up. That's why agents matter beyond a nicer chatbot.

Imani 6:16 Provided they finish the right job. There's a big difference between a plausible answer and a reliable action with consequences.

Miles 6:24 The research group METR measures performance on tasks using the time a human expert would need. Its fifty-percent horizon means success about half the time at that task difficulty. It does not mean an agent can safely run unattended for that many hours. The tests are largely well-specified software tasks, cleaner than many real jobs.

Imani 6:46 Let me make the reliability problem concrete. Suppose ten steps each work ninety-five percent of the time, and failures are independent. The chance of all ten working is only about sixty percent. That's an illustration, not a measurement of a particular product. But it shows why an impressive step isn't an impressive process.

Miles 7:07 And the repair work matters. Who notices the wrong invoice? Can the action be reversed? Does the reviewer have enough time and information to disagree? A human approval button is not much protection if the human is checking three hundred cases an hour.

Imani 7:22 Here is an overlooked possibility. Imagine your assistant negotiates a bill with a company's assistant. Your costs of making an appeal fall. Their costs of rejecting it also fall. We could automate our way into a larger bureaucracy.

Miles 7:37 Or out of one, if we design it properly. Give both sides shared records, clear rules, and a real route to a person with authority. The scarce resource may become accountable human attention. A world of instant answers could still leave you waiting for somebody who can actually say yes.

Imani 7:56 That distinction is going to matter in daily life. Not just whether a system can talk to you, but whether you can challenge what it does to you.

Miles 8:08 All right. Let's give the software a body. This is where the conversation tends to leap straight to a humanoid making breakfast.

Imani 8:16 Meanwhile, the International Federation of Robotics reported in September that five million industrial robots were operating in factories worldwide in twenty twenty-five. More than six hundred thousand were installed that year. Those are industrial robots, not a count of humanoids, and not all use the newest AI.

Miles 8:35 Exactly. Physical automation is already real. But a factory is designed around repeatability. A home contains a sleeping dog, a wet towel, and a drawer that sticks only on Tuesdays. Recognising an object is different from handling it safely, thousands of times, under changing conditions.

Imani 8:54 So what should we ask when we see the astonishing video?

Miles 8:59 How many successful hours between interventions? Who resets it? How much human supervision is outside the frame? What happens when a part wears out? Those questions don't dismiss the achievement. They tell you whether a demonstration can become an affordable service.

Imani 9:13 And where would you expect the pressure to arrive first?

Miles 9:17 Our expectation is that bounded, repeatable settings keep an advantage: moving materials, sorting, inspection, and parts of food production. Unpredictable homes and complex care are harder. That's a direction of travel, not a promise that every warehouse role disappears or every care role is safe forever.

Imani 9:39 There is a profound upside if machines take dangerous or exhausting tasks. But the person whose body was doing that work still needs an income, a transition, and a say. Calling the work undesirable doesn't make losing it painless.

Miles 10:00 There's another physical limit. You can download new software overnight. You can't download a substation.

Imani 10:07 The International Energy Agency's twenty twenty-six outlook projects data-centre electricity demand rising from roughly four hundred and eighty-five terawatt-hours in twenty twenty-five to nine hundred and fifty by twenty thirty. That's a central projection for all data centres, not a measurement of AI alone.

Miles 10:26 And it warns about bottlenecks. But efficiency is improving too. The important question is whether cheaper individual tasks are outweighed by many more tasks, or more demanding ones. A brief answer and an agent working through a long investigation don't have the same footprint.

Imani 10:43 For a household, the interesting question isn't the size of a remote computer building. It's who pays for the new wires and generation.

Miles 10:51 Exactly. More demand doesn't automatically mean everyone's electricity bill rises. The IEA says the effect depends on local supply, investment, and how costs are allocated. Ask who carries the risk if infrastructure is built for a project that uses less power than promised.

Imani 11:08 There's a broader ownership question too. If intelligence becomes inexpensive but access to customers, computing, and distribution stays concentrated, the savings might collect at a few bottlenecks. That is a possible economic outcome, not an unavoidable law of technology.

Miles 11:27 Which is why competition and the ability to move your data matter. A wonderfully capable assistant is less empowering if changing providers means losing years of your working life.

Imani 11:41 Let's talk about the future that makes this worth doing. More people getting help they could never previously afford.

Miles 11:49 Education is a good test. A randomized Harvard physics study published in twenty twenty-five found greater learning gains with a carefully designed AI tutor than with its active-classroom comparison, across the lessons tested. It's promising. It is not evidence that a generic chatbot can replace a school, or that short-term gains always last.

Imani 12:12 That difference is everything. A tutor can ask you to explain your mistake. An answer machine can help you hide it. Imagine affordable coaching in a language you understand, whenever you need it. Then imagine measuring success by what you can do after the assistant is switched off.

Miles 12:30 And apply that same discipline elsewhere. A useful medical tool must be tested on patient outcomes and real workflows, not just exam questions. A scientific hypothesis still has to survive an experiment. Faster preparation could be enormously valuable without making the whole profession obsolete.

Imani 12:51 There is also a cost to abundance. If polished speech and convincing messages become cheap, being persuaded can stop being evidence that somebody is who they claim to be.

Miles 13:02 The FBI has warned about criminals using generated voices to impersonate relatives and request money. A sensible response is to verify through a channel you already trust, rather than treating a familiar voice as proof. Our own hosts are synthetic voices, which makes that point uncomfortably close to home.

Imani 13:22 The longer-term concern isn't only a fake call. Imagine an assistant that knows what reassures you, what embarrasses you, and when you're tired. Does it serve your interests, an advertiser's, or a sales target? That is a design choice we should inspect before the relationship becomes intimate.

Miles 13:41 The most helpful interface we've ever built could also be an unusually persuasive one. Knowing whose interests it serves needs to be part of knowing how it works.

Imani 13:54 So, near term, what should a listener actually watch?

Miles 13:59 Watch hiring by experience level, not just layoff announcements. Watch the quality of service after automation, not just the number of tickets closed. Watch whether robot deployments run reliably beyond a demonstration. And watch who receives the productivity gains. Those are better signals than a confident date for the end of employment.

Imani 14:19 And further out? Suppose reliable digital labour becomes dramatically cheaper, and physical automation broadens. What changes beyond the workplace?

Miles 14:30 Potentially the price of services, the structure of companies, how people learn, and how governments raise revenue. But cheaper cognition doesn't make housing, energy, care, or political agreement unlimited. A richer economy can still contain households that feel less secure. Distribution and institutions remain central.

Imani 14:50 I can see two plausible versions of the same technical progress. One where a smaller group owns the systems and everyone else competes for the remaining attention. Another where people gain better services, more independence, and genuinely more time. Neither arrives just because a benchmark improves.

Miles 15:10 For employers, that means preserving routes for beginners and giving workers a voice in redesign. For public policy, it means support people can actually use during transitions, competition, and meaningful appeals. For an individual, learn the tools alongside a domain you understand. Keep practising judgment instead of outsourcing every difficult thought.

Imani 15:32 And leave room for uncertainty without using it as an excuse to do nothing. We don't need to know the exact shape of twenty thirty-five to decide that people deserve a fair chance to learn, to earn, and to challenge a decision.

Miles 15:47 I keep coming back to the empty desk in your opening. The remarkable thing about this moment is that we still get to ask what belongs there. A machine doing the routine work? A beginner learning beside it? A person with more time to care for somebody? The answer is a choice about what we value.

Imani 16:04 This has been The Quiet Rewrite, with Imani and Miles. The studies, their limitations, and our source notes are included with the episode. Thanks for spending this time with us.

Miles 16:15 The future will contain extraordinary machines. Let's make sure it still contains a place for people to begin.