The Thinking That Survives

There is a sentence I have heard so many times in the last three years that I have stopped hearing it. You will know it too. “Jobs will go, but new ones will come; people will be retrained.” It is meant to reassure, and it usually does. The trouble is in the small word at the end. Retrained. Note the passive voice. Something will be done to people, and on the far side of it they will be employable again, slotted into roles that do not yet have names. The promise treats workers as components that can be swapped between sockets once someone updates their firmware.

I want to take that promise seriously, because the people saying it are not fools and the broad shape of it has been right before. New technologies really have destroyed categories of work and created others. But the comforting version skips over the part that matters most, and the part that, as a trainer, I watch fail in real rooms with real people. It assumes that the learning involved is the easy kind. It is not. The cognitive-science evidence points the other way; and once you see the mechanism underneath the slogan, you stop asking whether the new jobs will appear and start asking who will actually be able to do them.

My answer, after just over a year of running these workshops, is this. The people who thrive will not be the ones waiting to be shown. They will be the ones who already hold, or can build, a handful of hard-to-train modes of thinking; and the rest of this essay is an attempt to name those modes, defend them, and then, because I am a trainer and not a prophet, say something about how you might grow them.

The mechanism under the slogan

Start with the word “retrained” and ask what it is quietly assuming. It assumes transfer of learning; that what you learn in one setting will carry across to a different one. This is one of the oldest and best-studied problems in education, and the news is not good for the slogan.

Perkins and Salomon (1992) drew a distinction that has held up for nearly four decades. They separated low-road transfer from high-road transfer. Low-road transfer is the automatic, reflexive triggering of a well-practised routine in a situation that closely resembles the one you practised in; you learned to drive a car, you get into a slightly different car, and your hands already know what to do. High-road transfer is different in kind. It “depends on mindful abstraction from the context of learning or application and a deliberate search for connections,” and it “demands time for exploration and the investment of mental effort.” It is the kind you need when the new situation does not look like the old one, when you have to reach up to a principle and carry it across a gap to somewhere it was never rehearsed.

I see the low-road version every week, and it has a tell. In an Excel class it is the upskiller who writes down the menu path and the exact composition of the VLOOKUP formula that does the thing, word for word, like a spell that must be cast in precisely the right order. They have the artefact; what they do not have, or will not risk, is the lateral move; opening the formula up, breaking it on purpose, kicking the problem around with Copilot until they arrive at the answer themselves. They have learned the path. They have not learned the territory, and the moment the spreadsheet poses a question that is one step off the recorded one, they are stuck. That is near transfer, perfectly intact and perfectly brittle.

The cybersecurity version is the one my other favourite field wrings its hands over relentlessly. Think of the person who waits for the resident tech expert in their life to show them how to switch on multi-factor authentication, rather than sitting with the small irritation of it, logging into one device now demands juggling two, and asking why and pondering could this be a clue about how it works? I have never met a single person who uses MFA without complaint until they understand why it works; and the reason it is so easy to resent is that the logic is not as obvious as locking the car door in the driveway. The threat it defends against is invisible and abstract. So the person who only ever had it switched on for them never crosses from doing to understanding, and they are the same person who will later believe, sincerely, that a “hacked” Facebook account can simply be walked back with a phone call to the aforementioned tech expert in their life, because they never built the model of how any of it actually works.

Here is the uncomfortable finding under all three; the driver, the Excel scribe, the reluctant MFA user. Conventional teaching reliably produces the first kind of knowing and reliably fails to produce the second. Perkins and Salomon warned that ordinary instruction “often fail[s] to establish the conditions either for reflexive or mindful transfer.” Detterman, surveying the same territory, was blunter still; far transfer is rare and genuinely hard to produce on demand. These are the modes of thinking that leave people dazed and stuck the day Microsoft rearranges the furniture in the ribbon; and, more worryingly, the modes that let someone believe a compromised account credential is a thing you can politely revoke from those naughty hackers.

Now hold the slogan up against that. “We will retrain people for new and different jobs” is a promise of far transfer. The new jobs, by definition, do not resemble the old ones; if they did, they would not be new. So the slogan is promising precisely the thing that the entire transfer literature says education delivers least reliably. A short reskilling course can teach you a new menu path, a new tool, a new procedure; that is near transfer, and we are good at it. It cannot reliably manufacture the supple, principle-level adaptability that a genuinely novel role demands. That is the gap the slogan steps over without noticing.

You can hear the same assumption in the executive register. Sam Altman has written, more than once, that we will “discover new jobs, we always do after a technological revolution,” and that nobody now wishes they were a lamplighter. He is probably right about the lamplighters. But notice what the sentence does; it moves the whole question into the past tense and treats the human side as automatic. The lamplighters did not get reissued as electricians by a memo. The transition was generational, uneven, and brutal for many of the people inside it. “We always do” is true at the scale of a century and cold comfort at the scale of a career.

So the rest of this is not really about jobs. It is about the kind of thinking that survives the gap the slogan ignores. I count four modes. They are not equally well-evidenced and I will not pretend they are; the third is the one I would defend hardest, because it is the one I have actually watched work.

(a) Adaptive expertise, not routine expertise

There are three kinds of people in every change-management workshop. One is running it. One arrived excited. And one was sent to it. We need not worry about the excited; it is the gap between the first and the third that marks the first mode of thinking, and cognitive science has a name for it.

Hatano and Inagaki distinguished routine expertise from adaptive expertise. The routine expert is fast, accurate, and automatic within a known domain; the classic example is the abacus master who can compute at extraordinary speed but only along familiar lines. The adaptive expert can do something the routine expert cannot. They understand why the procedure works, they can modify it when conditions change, and they can invent new procedures when the old ones run out. The routine expert is brilliant until the ground shifts; the adaptive expert is at their best precisely when it does.

This maps almost too neatly onto what I see. The person who can recite the steps but freezes when the interface updates is a routine expert in a world that has stopped being routine. The person who treats a new tool as a puzzle to be reverse-engineered is an adaptive one. And here is the part that matters for anyone who wants to become the second kind. The conditions that grow adaptive expertise are known. Hatano and Inagaki named three; environments with variability and a degree of randomness rather than rote repetition; low-stakes settings where you are not being graded or rewarded on every move, so you can afford to experiment; and group cultures that value experimentation over raw efficiency.

Look at that list again, because it should be familiar. Variability, low stakes, permission to experiment over a culture of pure efficiency. Those are also, almost exactly, the conditions of unstructured tinkering with AI tools; the throwaway project, the “what happens if I ask it this” afternoon, the deliberately pointless experiment. The conditions that build the most valuable form of expertise are the same conditions that an idle hour with a chatbot can provide, if you treat the hour as exploration rather than as a task to be completed. That is not a small observation, and I will come back to it.

A fair warning sits inside this, though. Adaptive expertise has a cost. The innovation dimension can depress short-term routine performance; the person who is always asking “but why does it work this way” is, in the short run, slower than the person who just does the thing. Organisations often rationally under-reward this, which is part of why it is rarer than it should be.

(b) The creator, the inventor, the one who acts under uncertainty

The second mode is the one I find most exciting and trust least, so I will flag the risk before I make the case.

There is a body of work in entrepreneurship, principally Sarasvathy’s, on what she calls effectuation. Most planning logic is causal; you set a goal, then assemble the means to reach it. Effectual logic runs the other way. It starts from the means you already have, decides how much you can afford to lose, and acts; it exploits whatever contingencies turn up rather than predicting them in advance. It is the logic of creation rather than discovery, and the folk version of it is “ask forgiveness, not permission.” Sarasvathy found that her expert entrepreneurs used this means-driven logic most of the time, and, importantly, she framed it as learnable rather than innate.

Pair that with two well-studied traits; need for cognition, defined as “the tendency for an individual to engage in and enjoy thinking,” and epistemic curiosity, the appetite for new ideas and for closing gaps in what you know. People high in these do not wait to be handed a problem; they go looking. In a world where the cost of trying something has collapsed to near zero, the person who reflexively tries things has an advantage that compounds.

Now the provocation, which I want to defend carefully rather than throw out and run. I think a certain kind of aimless AI tinkering, the throwaway “vibe-coding” project, the afternoon spent generating things for no client and no deadline, is cognitively analogous to basic research. This is my analogy, not a classification anyone official would sign off on, and I want to be exact about where it holds and where it breaks.

The OECD’s Frascati Manual defines basic research as work “undertaken primarily to acquire new knowledge of the underlying foundation of phenomena and observable facts, without any particular application or use in view”; curiosity-driven, exploratory, fundamental. It sets five criteria for what counts as research at all; it must be novel, creative, uncertain in outcome, systematic, and transferable or reproducible. Exploratory tinkering clearly satisfies the first four. It is novel to the tinkerer, creative, uncertain, and, if you are paying attention, systematic.

Where it breaks is the fifth criterion, and the break is the whole point. Tinkering only counts as the cognitive equivalent of research when the maker reviews, tests and understands what came out; when the activity is transferable into the maker’s own competence. This is exactly the line Simon Willison drew about vibe coding. If a model wrote every line but you reviewed, tested and understood it all, that is not really vibe coding; that is using the model as a typing assistant, and you have learned something. Andrew Karpathy’s own tinkering, the thing he was describing when he coined the phrase, is the legitimate case; he is an expert playing at the edge of his competence and watching what happens.

So here is the bounded claim, and I will not let it flatter anyone who does not earn it. The thing worth praising is the posture, not the artefact. Trying things, breaking ground, asking “what if I simply wished this into existence and saw what came back”; that posture is cognitively valuable even when the output is rough. The maker who experiments, inspects the result and learns from it is doing something real. The maker who generates volume and accepts it unexamined is not, however polished it looks. Reward the posture; never the slop for its own sake.

A point of honesty about the word “slop.” It was popularised in 2024 as a term for content that is mindlessly generated and pushed at people who did not ask for it; both Merriam-Webster and Australia’s own Macquarie Dictionary made versions of it their word of the year for 2025. The disdain is fair when aimed at the artefact. It is unfair when it sweeps up the curious experimenter along with the spammer; those are different people doing cognitively different things, and the difference is whether anyone looked at the output and thought about it.

(c) The information-literate context engineer

This is the mode I would stake the most on, because it is the one I have watched do real work, and because it is the one the field has been quietly migrating toward in front of us.

For a couple of years the skill everyone talked about was “prompt engineering”; the art of phrasing a request to get a good answer. In the middle of 2025 the centre of gravity shifted, and the practitioners gave the new thing a name. Context engineering. Karpathy describes it as “the delicate art and science of filling the context window with just the right information for the next step.” Anthropic calls it “the natural progression of prompt engineering.” One industry analyst put it flatly; context engineering is in, prompt engineering is out.

The shift sounds technical but it is really a shift in where the human adds value, and that is the part I care about. With prompt engineering, the human’s job was to critique the output; you got an answer, you judged it, you asked again. With context engineering, the human’s job moves upstream, to curating the input; deciding what the system should be allowed to see, what to feed it, what to leave out, what sources to trust. The value migrates from judging what came out to shaping what goes in. This is the insight I keep coming back to in workshops, and it is older than the jargon. The discipline that has always been about judging, scoping and curating sources has a name. It is information literacy.

Information literacy is not a buzzword; it is a mature field with frameworks behind it. The Association of College and Research Libraries built one around the idea that authority is constructed and contextual, that information has value, that research is inquiry rather than retrieval. Britain’s SCONUL model lays out seven non-linear pillars; identify, scope, plan, gather, evaluate, manage, present. And, closest to home, Christine Bruce’s Seven Faces of Information Literacy, an Australian model out of Griffith University, includes a face she calls “information control”; the skill of organising and curating sources so you can find and trust them later.

Put those two things side by side and the point is hard to miss. The skill the new tools are rewarding, deciding what to feed the machine and what to trust, is the skill librarians and information-literacy researchers have been teaching for decades. The future, in this corner at least, looks a great deal like the past, wearing newer clothes.

The cleanest worked example is a tool called NotebookLM. It is built on what is called retrieval-augmented generation, or RAG; rather than answering from whatever it absorbed in training, the system answers only from a body of sources you give it, and it cites them. (That is what RAG means in plain terms; the model “retrieves” from a curated set of documents before it “generates” its answer, so the answer is grounded in material you chose rather than in the model’s general memory.) The point for our purposes is where the value sits. The tool does the retrieving and the writing. The human does the part that determines whether the output is any good; selecting the sources, scoping the corpus, judging what belongs in it. That is Bruce’s “information control” face exactly; that is SCONUL’s “scope, gather, evaluate.” The machine handles generation. The human handles judgement. And judgement is the thing that does not come in the box.

This is also where I will use a word I owe you a definition for, because it runs through all four modes. Metacognition is, in the original phrasing, “knowledge and cognition about cognitive phenomena”; more plainly, it is thinking about your own thinking. It has two halves; knowing things about how you learn and reason, and regulating that reasoning as you go, by planning, monitoring whether it is working, and adjusting. The reason it matters here is that curating an input well, deciding whether a source belongs, noticing when an answer smells wrong, is a metacognitive act. You are not just processing information; you are monitoring your own confidence in it. The good news, which I will return to, is that metacognition is among the more trainable of the things on this list.

(d) The articulate imaginer

The fourth mode is the quietest and easiest to underrate. It is the skill of describing what you picture; of turning something half-formed in your head into words precise enough that another mind, including a machine, can act on them.

The clearest test case is image generation. To get something good out of an image model you have to say what you want, in detail, and most people are startlingly bad at this. Not because they lack imagination but because they have never had to externalise it; the picture stayed comfortably vague inside their own head, and the moment they have to render it in language they discover how little of it was actually specified.

This connects to one of the deeper currents in the labour literature; the premium on articulating tacit knowledge. Polanyi’s famous line is that “we can know more than we can tell,” and Schön made a related point about skilled practitioners; “competent practitioners usually know more than they can say.” For most of history that was a safe place to keep knowledge, precisely because what could not be articulated could not be automated; it was the human’s moat. What the current tools do is raise the reward for crossing that moat in the other direction. The person who can drag tacit, half-felt intention up into explicit language can now direct a machine to build it. The person who cannot is stuck pointing and hoping.

There is a sharper, slightly unsettling finding underneath this. The skill of articulating what you imagine seems to be tied to the skill of detecting what is fake. Study after study finds that humans are poor at spotting AI-generated images; in one large consumer survey only a tiny fraction of people correctly identified a full set of real and synthetic images, while their confidence stayed high regardless of whether they were right. But the researchers who dug into who does better kept finding the same predictors; higher analytical thinking and stronger digital skills. The people who have actually made things, who have wrestled with describing an image into existence, are better at seeing where a fake comes apart. The free-tier dabbler who tried an image generator once, was unimpressed, and never came back has not just missed a tool; they have skipped the practice that would have sharpened their eye.

Now the part where I argue against myself

A thesis like this is only worth reading if it has already met its strongest objections, so let me put them as forcefully as I can, because each of them has real teeth.

The first is the hardest, and I will not wave it away. The best field evidence we have says generative AI helps novices most. In a large study of customer-support agents, access to an AI assistant raised resolved-issues-per-hour by around fourteen per cent on average; but the gain for the least experienced workers was about thirty-four per cent, while the most skilled barely benefited. The mechanism was that the AI spread the best workers’ practices to everyone else. On its face this demolishes my “passive learners lose” story; if anything the tool is a great leveller that lifts the bottom.

I have sat with this finding for a long time and I think the resolution is the most interesting thing in this essay, so I want to give it room rather than rush past it. Both things can be true at once, and the link between them is the point. In the short run, the tool levels; the novice resolves more tickets today. But adaptive expertise, the first mode, grows out of exactly the deep, effortful practice that the tool now lets you skip. The thirty-four per cent gain comes from not having to struggle through the hard cases yourself, because the machine hands you the expert’s answer. The struggle was never just inefficiency; it was the experience curve, the thing from which deep expertise is built. So the short-run levelling may be quietly flattening the very curve that produces the adaptive experts of the next decade. You get a more productive novice today and, possibly, fewer masters tomorrow. And on top of that, the Turing Trap warning applies; when a tool substitutes rather than complements, the gains tend to flow to whoever owns the tool, not to the worker’s bargaining power. The leveller can raise everyone’s output and still concentrate the value. The novice-augmentation finding does not refute the thesis. It sharpens it into a question about time horizons.

(Now, it is worth pinning a caveat here that this is an argument about a long-run effect we cannot yet measure and I’m keenly aware that a longitudinal study showing AI-augmented novices going on to deeper expertise would force me to revise it.)

The second objection is about a word I have leaned on; “wanting to learn.” It would be easy to dress this up as growth mindset, the popular idea that believing your abilities can grow makes them grow. I am deliberately not doing that, because the evidence is weak. A large meta-analysis found the correlation between mindset and achievement was about one per cent of the variance, and that the average effect of mindset interventions was tiny; later high-quality work found effects close to nothing, with a small, real benefit surviving mainly for at-risk students under supportive conditions. The construct is contested, not debunked, but it is far too thin a reed to hang a thesis on. So when I talk about the disposition to keep learning, I am not appealing to Dweck; I am appealing to the better-evidenced machinery of self-regulated learning and metacognition, which is the next section.

The third objection cuts both ways, and intellectual honesty demands I admit it. I have used Detterman’s point that far transfer is rare as a weapon against the “just retrain them” slogan. But if far transfer really is that rare, then demanding it of everyone is not just under-supplied by training programs; it may be flatly unrealistic as a policy goal. Maybe the honest conclusion is not “everyone must become an adaptive expert” but “this will always be a minority capacity, and we should stop pretending otherwise.” I do not have a palatable answer to that. I think it tempers the thesis from a universal demand into a description of where advantage will sit, which is a more modest and more defensible claim.

The fourth objection is that I may be overstating the whole disruption. The macro-economists do not agree on the size of any of this. Acemoglu’s careful estimate puts AI’s effect at no more than about two-thirds of one per cent of total factor productivity over a decade; a rounding error. Against that sit Goldman Sachs, projecting that generative AI could raise annual global GDP by seven per cent over ten years, and McKinsey, putting the annual prize at trillions. The occupational estimates are just as split; the famous Frey and Osborne figure of forty-seven per cent of US jobs at risk sits against OECD task-based critiques that put the genuinely automatable share closer to nine to fourteen per cent.

Here is why I am not undone by this. My argument does not depend on which estimate is right. If Acemoglu is right and the disruption is slow and modest, then the four modes of thinking are not a survival kit; they are simply where the advantage sits, and the essay softens from “you will drown” to “you will get ahead.” If the optimists or the alarmists are right and the disruption is large, the same four modes are what keep you afloat. The cognitive case holds across the whole range of macro outcomes, because even small, task-level changes reward the person who can adapt, curate, create and articulate over the person who is waiting to be shown. The labour numbers tell you how urgent this is. They do not change what it is.

What a trainer does about it

I will not end on the doom, because the doom is not the point and, frankly, it is not even what the evidence supports. The useful part, the part I can actually do something with on a Wednesday morning, is that these modes are partly trainable. Not magically, not for everyone, not to the level of an adaptive genius; but movably, in the direction that matters. Let me be concrete.

Teach the thinking-about-thinking directly. Metacognition responds to instruction; in one longitudinal study, a single semester of direct metacognitive teaching produced the largest gains, larger than maturation alone. The practical version of this in a workshop is almost embarrassingly simple. Make people plan before they act, monitor whether it is working while they act, and evaluate honestly afterward. That loop, plan, monitor, evaluate, is the core of what is called self-regulated learning, and it is the most reliable lever I have. The person who runs that loop on themselves is building the regulation half of metacognition whether they know the word or not.

Build the conditions for adaptive expertise on purpose. Recall Hatano and Inagaki’s three; variability, low stakes, a culture that rewards experiment over efficiency. So I design practice that varies, that is genuinely low-stakes, where nobody is graded on getting it right first time, and where trying the weird thing is praised rather than the fast thing. Case-based learning, where you meet many different messy situations rather than drilling one clean procedure, is the form this takes. It is slower than a step-by-step demo and it builds something the demo never can.

Use the tool to teach the posture, not just the task. My most reliable trick is also the simplest, and it works best in the rooms least expecting it; the Microsoft suite classes, where people came to learn Excel or Teams and find Copilot bolted on as a module they did not ask for. Someone asks a technical question, “how do I get this tool to do such-and-such,” and instead of answering I say, “right, let’s ask Copilot.” On the surface I am dodging the question. Underneath, four things are happening at once. First, I am letting Copilot showcase its trademark non-determinism; it will not give the same answer twice, and that is the feature, not the bug. Second, I am inviting the room to push it, to do a little solution-oriented tinkering rather than wait for the correct incantation. Third, and this is the one I love, I have everyone run the same prompt and then turn to the person beside them and compare what came back; two identical questions, two different answers, and a small jolt of understanding about what kind of capricious machine this actually is. Fourth, I am giving them a chance to feel what an AI conversation is like in the brain, because it does not feel like doing Excel; Excel rewards the exact path, and this rewards the willingness to deviate from one.

The resistance is real and worth naming honestly. The feedback I dread, and sometimes get, is “I didn’t sign up for AI training.” It is a fair complaint on its own terms; they signed up for a spreadsheet course. But the truth I have made a kind of peace with is that, with the exception of a handful of very wealthy tech entrepreneurs, none of us signed up for this. Yet here we are. The “let’s ask Copilot” move is the smallest possible dose of the mindset shift; it costs nothing, it embarrasses no one, and it lets the irritated spreadsheet-enamoured attendee discover the experimental posture almost by accident, in the middle of doing the thing they actually came for.

Make source-curation a teachable act. This is where the information-literacy frameworks earn their keep. Sit someone down with a RAG tool like NotebookLM and the lesson is not “look what the machine can write”; it is “look how much the answer depends on what you chose to feed it.” Curating the corpus, scoping it, deciding what to trust, that is the skill, and it is the same skill the librarians have always taught.

And reward the curious experiment, every time you see it. The free-tier dabbler who tries once and quits has my sympathy but not my optimism. The person who keeps poking, who treats the tool as a thing to be reverse-engineered rather than a vending machine, is doing the cognitively valuable thing even when their output is rough. Praise the posture. It is the most precious and least teachable item on the whole list, and the closest I can come to teaching it is to make sure it is the behaviour the room respects.

So that is where I land, after just over a year of these workshops. The “new jobs will come” promise is not a lie. It is just incomplete in the one place that matters; it tells you the jobs will exist and stays silent on whether you will be able to think your way into them. The difference is not between people who will be retrained and people who will not. It is between being shown and learning to see. The first is done to you. The second you have to build. And the four modes above, adaptive expertise, the creator’s posture, the curator’s judgement, and the articulate imagination, are, as best the evidence can tell us, the materials you build it from.

That is the thinking that survives. The rest is firmware.