The Thinking That Survives

I've heard a sentence so many times over the past three years that it no longer registers at all. “Jobs will disappear, but new ones will appear; people will be retrained.” It sounds reassuring, and generally works. But the hidden message lies in that final word. Retrained. Notice the passivity inherent in that statement.

Some action will be performed on people, and on the other side of that action they will once again be employed in roles that have not yet been defined. The result is equivalent to treating workers as components to be substituted for sockets in response to a firmware update. I intend to take that promise seriously because those making it are not foolish, and much of the overall message has been valid in the past. New technologies have indeed eliminated entire classes of work and generated new classes in their place. However, they conveniently omit the most important aspect of the process-the one that is visibly occurring in actual rooms and with real people.

The experience of learning is clearly not a simple one. Rather, it is substantially more difficult than would be implied by the dominant message. Our knowledge of the mechanisms of learning strongly supports this conclusion (Perkins and Salomon, 1992), and ultimately leads to a different question: Who actually will be able to perform the new jobs? My answer after more than a year of conducting these workshops is as follows. The individuals who will succeed are those who will not wait to be retrained.

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 gives it away. In an Excel class, the upskiller records the menu path and the precise composition of the VLOOKUP formula that accomplishes the task, word for word as if reciting a spell that must be cast in precisely the right order. They obtain the artifact, but not the lateral move of expanding, intentionally breaking and reworking the problem with Copilot to arrive at answers of their own. As a result, they learn the path, but not the territory. When confronted with a question one step beyond the previously recorded response, they are completely lost. This represents nearly complete preservation of and susceptibility to disruption by the integrity of transfer.

The cybersecurity equivalent is the one that distresses my other favourite discipline to the point of exhaustion. Consider the individual who waits for the resident expert in his or her life to demonstrate how to activate multi-factor authentication, rather than experiencing the mild annoyance associated with this task. Access to one device then requires management of two, and prompts for reasons and consideration of whether these might represent clues to mechanisms of operation. I have never encountered a single individual who was completely free of complaints until fully understanding the basis for operation. The ease with which we express dissatisfaction reflects the fact that mechanisms of operation are not perceived as obviously as the act of locking the door of a car in the driveway. The nature of the threat against which we are protecting is invisible and abstract, and therefore it is not surprising that individuals who had never experienced activation of multi-factor authentication for themselves ultimately fail to achieve an understanding of how any of it actually works.

Finally, I obtain the disturbing conclusion that all three types of individuals ultimately represent one and the same category. Traditional methods of instruction are fully successful in producing the first type of knowledge and completely ineffective in generating the second. Perkins and Salomon cautioned that routine instruction "frequently fails to provide conditions for reflexive or mindful transfer." Detterman's survey of the field was even more pessimistic: Far transfer is infrequent and extremely difficult to achieve on demand. These represent forms of thinking that leave individuals bewildered and immobilised when Microsoft reorganises the ribbon, and even more disturbingly, forms of thinking that support the belief that compromised credentials for an account can be politely revoked from mischievous hackers.

Hold that slogan up to the evidence. Our commitment to retrain people for new and different jobs represents an expectation of very large transfer. By definition, these new jobs are unlike previous work; otherwise, they would not be new. Thus, the slogan reflects an expectation of exactly what the entire literature on transfer says is the least well achieved by education.

Short programs for reskilling will teach new patterns of menu selection, new tools, new procedures and represent large but relatively reliable transfers of skill. They will not, however, reliably produce the flexible, principled capacity for adaptation required by truly novel roles. This is the gap that our slogan crosses without awareness.

The same assumption appears in the executive register as well. Sam Altman has written repeatedly that we will "discover new jobs, as we always do following a technological revolution," and that no one now would wish to be a lamplighter. He is certainly correct about the lamplighters. But observe the effect of the statement: It places the entire situation in the past and treats the human response as automatic. Lamplighters were not reconverted to electricians by a memo. The process was generational, uneven and extremely difficult for many of those involved. "We always do" reflects the experience of a century, and is little comfort to individuals with careers of only a few years. Consequently, the remainder of this discussion is not really about jobs at all, but about modes of thinking that survive the gap ignored by the slogan. There are four such modes of thinking, of which I have evidence of varying degrees. I will not attempt to describe them with equal accuracy, because the third represents the mode for which I have the most direct experience.

(a) Adaptive expertise, not routine expertise

There are three types of participants at every meeting to discuss change management. One is conducting the meeting. Another arrived with great enthusiasm. And a third was sent to the meeting as a matter of routine. I need have no concern about the enthusiastic participants. The difference between the first and the third represents the first mode of thinking, and has a well-defined label in the literature of cognitive science. Hatano and Inagaki distinguished routine expertise from adaptive expertise. Routine expertise is characterised by high speed, accuracy and automaticity within a well-learned domain. The prototype is the expert who can calculate at an extraordinary rate, but only along familiar lines of computation.

Adaptive expertise permits performance of tasks that cannot be achieved by the routine expert. It reflects understanding of the principles underlying task performance, capacity for adaptation to changes in conditions, and ability to develop new modes of task performance when previous ones are exhausted. Expert performance is maximal when conditions are most adverse, and represents the full potential for rapid adaptation of behaviour to changing conditions.

This fits almost too neatly with what I see. The individual who can recite the steps but freezes with each update in the interface is a routine expert in a world that no longer is routine. The individual who interprets a new tool as a puzzle to be reverse-engineered is an adaptive expert. And here is the message for anyone who wants to become the latter type of expert. The conditions that foster adaptive expertise are well known.

Hatano and Inagaki named three conditions of particular importance: environments of variability and some degree of randomness rather than of routine repetition; low levels of risk in which one is not rewarded or penalised for every action and thus can afford to experiment; and cultures that encourage experimentation rather than maximal efficiency. Note again how familiar these conditions are. They represent variability, low risk and permission to experiment in contrast to a culture of complete efficiency. These are precisely the conditions of unstructured experimentation with AI tools, of the throw-away project, of an afternoon spent simply asking "What would happen if I told it this?" and of deliberate, purposeless experimentation. The conditions that produce the greatest value of expertise are identical to those provided by a single afternoon of unstructured interaction with a chatbot, provided that the experience is interpreted as exploration rather than as completion of a task.

This is not a trivial observation, and will be repeated again later. But there is also an important message of warning. Adaptive expertise is costly. The component of novelty impairs short-term performance with routine tasks; an individual who constantly asks "But why does it work this way?" is less efficient in the short term than is an individual who simply performs the required task. Organisations generally provide inadequate compensation for this, and thus contribute to the rarity with which adaptive expertise is achieved.

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

The second mode is the one I find most exciting and least trustworthy, and so I will alert you to the risk before I make my case.

There is a substantial literature, principally Sarasvathy's, on what she calls effectuation. Most planning is causal: One establishes a goal, and then identifies the means to achieve it. Effectual reasoning is in the opposite direction. It begins with the means available, determines the amount that can be afforded to lose, and acts to exploit whatever contingencies arise, rather than to predict them in advance. It represents the logic of creation, not discovery, and embodies the folk wisdom that we should seek forgiveness, not permission.

Sarasvathy found that her expert entrepreneurs largely applied this mode of means-based reasoning, and importantly, she framed it as learnable rather than as an innate capacity.

Combine this with two well-established dimensions of personality: Need for cognition, reflecting the tendency to enjoy and engage in thinking, and epistemic curiosity, reflecting a desire for novel ideas and for closure of existing gaps in knowledge. High levels of these traits do not await the opportunity to receive a problem; instead, they generate problems. In a world in which the cost of attempting something is virtually zero, individuals who automatically try things have a compounding advantage.

Finally, let me defend carefully, rather than reject outright, a provocative assertion. Certain forms of apparently aimless experimentation with artificial intelligence, including the development of a "vibe-coding" project for the afternoon with no client and no deadline, are cognitively equivalent to basic research. This is not a classification that would receive official approval, and I will be precise regarding conditions of validity and of failure.

The OECD's Frascati Manual defines basic research as work "primarily to increase knowledge of the underlying principles of phenomena and of observable facts, without a specific application or use in mind." It is characterised by five criteria for eligibility as research: novelty, creativity, uncertainty of outcome, systematic methods, and capacity for transfer or replication. Exploratory tinkering certainly fulfils the first four criteria for research: It is novel to the tinkerer, creative, uncertain and, if you are paying attention, systematic.

It fails on the fifth criterion and in a way that is entirely appropriate. Only when the maker evaluates, tests and interprets the results will tinkering represent the cognitive equivalent of research. Only then will the activity be transferable to the maker's own competence. This is precisely the distinction Simon Willison has drawn about vibe coding, a term coined by Andrew Karpathy. In effect, if every line of code were written by the model but evaluated, tested and interpreted by you, this would not represent an instance of vibe coding. Rather, it would reflect use of the model as a typing assistant and represent learning.

Finally, my own experience of exploratory tinkering represents an entirely appropriate application of the method. Here, I am an expert practitioner operating at the limits of my competence and interpreting the results of my work.

Here is the bounded claim, and I will not let it flatter those who do not deserve it. What deserves praise is the posture, not the artifact. Attempts to do things, break new ground and ask "What would happen if I simply wished this into existence and observed what emerged?" represent a form of cognitive value even when results are rough. The experimenter who observes results and learns from them achieves something real. In contrast, the producer of large volumes of material that receive no critical examination represents neither skill nor learning. Reward the posture, never the sloppiness for its own sake.

Some remarks of candour about the term "slop." It has become popular to apply this term to material that is mindlessly produced and directed to people who did not request it. Both versions of the term received designation as word of the year for 2025. Disapproval of the artifact is justified. But it is inappropriate to allocate curious experimenter the same slop production bucket as the spammer. These are activities of very different cognitive complexity, and reflect differences in extent to which results were evaluated and interpreted.

(c) The information-literate context engineer

This is the mode on which I would bet the most, because it is the one in which real work is accomplished and because it reflects where the field has been quietly moving ahead of me.

For several years the skill everyone extolled was "prompt engineering," the art of formulating requests to obtain good responses. However, by mid-2025 the centre of gravity had shifted, and the practitioners adopted a new term: context engineering. Karpathy describes it as "the delicate art and science of populating the context window with exactly the information needed for the next step," and Anthropic calls it "the natural evolution of prompt engineering." Ultimately, one industry analyst simply stated that context engineering is in and prompt engineering is out.

This represents a technological change of major importance, but is ultimately a reflection of the changing locus of human value-added. With prompt engineering, the role of the human was to evaluate the output, receive a response, judge it and request additional information. With context engineering, human responsibility moves upstream to the task of curating the input, determining what information should be permitted to enter the system, what should be included or excluded, and what sources should be considered reliable. As a consequence, the focus of human activity shifts from evaluation of results to control of the process of information retrieval. This is the major message I convey repeatedly in my workshops and reflects a philosophy of information literacy that is considerably older than the associated jargon.

The discipline of information literacy is thus firmly established as a mature area of practice with well-developed theoretical foundations. The information-literacy competencies described by the Association of College and Research Libraries reflect the importance of recognising the constructed nature of authority, of appreciating the value of information and of perceiving research as a process of inquiry rather than as an exercise in information retrieval. The seven interactive components of the SCONUL model of information literacy also reflect the importance of considering and providing for all components of information-literacy practice. The new information technologies are rewarding users for a skill I have taught for decades. That skill is to decide what information to feed the machine and what to trust. Most important, the skill of organising and curating information sources to make it easy to locate and trust information later is being rewarded as well.

Place those two developments side by side, and it is impossible to miss the message. The skill rewarded by the new technologies is precisely the skill that librarians and information-literate researchers have been teaching for decades. As a result, the future in this area bears a striking resemblance to the past, but in new clothing.

The cleanest example of an application is a tool called NotebookLM, based on a method referred to as retrieval-augmented generation. Rather than responding based on information acquired during training, the system responds only to information provided by the user and cites the sources accordingly. That is, the model retrieves from a set of selected documents before generating a response that is grounded in information chosen by the user rather than in the model's general memory.

The important point for my purposes is to identify where the value of the application resides. The machine does the retrieval and the writing, whereas the user provides the information necessary to determine whether the output is acceptable. That is, the user controls the extent to which information is included and excluded from the analysis. This represents exactly the "information control" face described by Bruce, and the process of defining boundaries for information that is selected and used to provide output represents the "scope, gather and evaluate" activities of SCONUL.

The machine provides for generation of output, and the user provides for evaluation of output. Evaluation is the component of the system that is not provided by the box. Here, I will use a term for which I owe you a definition, because it applies equally to all four modes of application. Metacognition refers to knowledge and awareness of cognitive processes; in other words, it represents thinking about thinking. It has two components, including awareness of mechanisms of learning and reasoning, and regulation of that process by planning, monitoring performance and modifying behaviour. Importantly, effective selection of information, recognition of errors in results and regulation of performance during analysis are all metacognitive activities. The result is not simply processing of information, but also continuous assessment of the degree of confidence in results. Fortunately, metacognition is among the most easily trained of all components of the system.

(d) The articulate imaginer

The fourth mode is the quietest and easiest to underestimate. It represents the ability to describe one's images with sufficient precision to enable another mind, including a machine, to act on them.

The best test of this ability is image generation. To obtain good results with an image model, one must describe with great specificity what one wants, and most people are remarkably poor at this task. This is not because they lack creativity, but because they have never been required to externalise it. Their mental images remained comfortably vague, and only at the moment of verbal description did they discover the extent to which details of their mental images had been omitted.

This illustrates one of the major trends in the literature on labour. There is a premium for articulating tacit knowledge. As Polanyi noted, "we know more than we can tell." Similarly, as Schön observed, competent practitioners always know more than they can express. For most of history, this was a safe way of protecting knowledge, because what could not be verbalised could not be automated. It represented the moat of the human.

Modern tools increase the risk of crossing this moat in the opposite direction. Individuals who can raise tacit and nearly unconscious intentions to the level of explicit verbal description are now able to direct machines to construct their mental images. In contrast, individuals who cannot achieve this level of verbal description remain constrained to gestures of hope and expectation.

Finally, there is an additional, somewhat disturbing observation. The ability to describe one's mental images is closely associated with the ability to detect features of artificial objects. Virtually every study confirms that humans are poor at recognising images produced by artificial intelligence. For example, in a large survey of consumers, only a very small proportion of subjects correctly identified a complete set of real and artificial images, and continued to express high confidence regardless of whether they were correct. However, further analysis of which subjects performed better confirmed the same predictors of ability: high capacity for analytical thinking and strong abilities with digital technology. Those of us who actually create things and wrestle with the process of generating an image from description are far better at identifying where a forgery falls apart. The casual user who tried an image generator once, was unimpressed and never returned has not simply failed to acquire a useful tool. Rather, he or she has foregone the experience that would have honed his or her ability to see.

Now the part where I argue against myself

A thesis like mine is worth reading only if it has already withstood its strongest objections. Let me therefore state them as forcefully as possible, since each has teeth.

First, and most difficult of all, I will not dismiss the strongest field evidence available. Access to a generative AI assistant produced the greatest gains for novices. In a large study of customer-support agents, it increased the rate at which issues were resolved by about 14%. The greatest benefit, however, was achieved by the most inexperienced workers, for whom AI-assisted care yielded a 34% improvement in performance, whereas the most experienced workers achieved no measurable improvement. The mechanism was transmission of the best practices of experienced workers to all others. This completely invalidates my claim that passive learners lose out; instead, the technology provides a powerful equalising effect that raises the level of performance of all workers.

I have pondered these results for many years and believe that the ultimate resolution represents the most interesting aspect of this paper. Both outcomes may therefore be true at the same time, with the interaction between them representing the essential point. In the short term, the technology equalises outcomes; inexperienced workers solve more cases per unit of time. However, development of adaptive expertise derives directly from the same intensive, effortful experience with difficult cases that the technology now permits me to avoid. The large improvement in performance for inexperienced workers reflects complete elimination of effort on difficult cases, with delivery of the results of expert analysis directly to the workers. This experience of intensive effort and learning represents the basis for development of adaptive expertise. Consequently, equalisation of short-term performance results in progressive reduction of the experience curve for development of adaptive experts over the next decade. I ultimately obtain greater productivity for inexperienced workers and fewer masters of experience for future years. In addition, the warnings about the "Turing trap" are fully applicable. With substitution rather than enhancement of worker performance, gains in productivity flow primarily to the owner of the technology and have little effect on the relative bargaining power of workers. Equalisation of output ultimately results in concentration of returns to the owners of the technology. Experience with the results of generative AI thus confirms and extends my previous conclusions. The initial surprise that observers often report when confronted with the work of AI-augmented novices gives rise to a larger question concerning time horizons. The finding of novice augmentation does not invalidate the thesis, but instead sharpens it into a question about time horizons.

Finally, it is important to add a caution here concerning the long-term effect that I have yet to measure. Should my subsequent experience confirm that AI-augmented novices ultimately achieve greater levels of expertise than do their unaided counterparts, then I will have to modify my current views.

The second objection concerns a term on which I have depended, "wanting to learn." It would be easy to reinterpret this as reflecting a growth mindset, the popular view that belief in the potential for growth actually produces growth. But I deliberately refrain from this strategy, because the supporting evidence is weak. The overall correlation between mindset and achievement was about 1% of the variance, and effects of interventions on mindset were trivial. Subsequent work with high methodological standards yielded essentially no effect for most students, with a small, genuine benefit occurring primarily for students at risk under supportive conditions. The resulting construct is in dispute, but not discredited. Nevertheless, it is too fragile to support a thesis. Consequently, when I refer to a disposition to continue learning, I am not evoking Dweck's ideas. Rather, I am using well-supported mechanisms of self-regulated learning and metacognitive awareness. (See the following section.)

The third objection applies equally to both sides of the argument, and I must admit it. I have exploited Detterman's assertion that far transfer is rare to attack the "just retrain them" argument. But if far transfer is really that rare, then to expect it of everyone represents not merely an insufficient capacity for training, but a complete failure of realism as a goal of policy. The honest result is not "Everyone must become an adaptive expert," but rather "This will always be a minority capacity, and we should stop pretending otherwise." I have no attractive solution to this problem. The result, however, is to modify the thesis from a universal requirement to a description of where competitive advantage will lie, and represent a less demanding and more defensible position.

The fourth objection is that I may be exaggerating the degree of disruption. Estimates of the magnitude of all these effects are not agreed upon by macroeconomists. Acemoglu's most careful estimate of the effect of AI is no greater than about two-thirds of 1 percent of total factor productivity over 10 years, a trivial amount. In contrast, the Goldman Sachs projection of a potential effect of generative AI on annual world GDP of 7 percent over 10 years is striking. Estimates of the potential for automation are equally discordant; the well-known Frey and Osborne prediction of 47 percent of U.S. jobs at risk is in sharp contrast to OECD critiques based on task requirements that describe the actual potential for automation as approximately 9 to 14 percent.

My argument doesn't come undone here as it doesn't depend on which estimate is right. If the degree of disruption is small and gradual, then the four modes of thinking represent not a method of survival, but simply a description of where competitive advantage will lie. In contrast, if the predictions of the optimists or the alarmists are correct and the degree of disruption is large, then the same four modes of thinking will provide the basis for survival. 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 with doom, because doom is not the point, and frankly is not supported by the evidence. The important part, and the part for which I actually can do something on a Wednesday morning, is that these modes are partially amenable to training. Not magically, not for everyone and not to the level of adaptive genius, but sufficiently to permit adaptation in the direction that matters. Let me be specific.

Teach the thinking-about-thinking directly.

Metacognition is responsive to instruction; a single semester of explicit instruction in metacognitive skills produced greater gains than did maturation alone. The application of this in a workshop was almost embarrassingly simple: Teach people to plan before acting, to monitor the effectiveness of their actions while acting and to evaluate honestly after acting. The resulting cycle of planning, monitoring and evaluation represents the core of what is referred to as self-regulated learning and constitutes my most reliable lever of instruction.

The individual who applies this cycle to himself or herself is developing the regulatory component of metacognition regardless of awareness of that fact.

Build the conditions for adaptive expertise on purpose.

This experience is characterised by variability, low risk and a climate in which experimentation is more highly valued than is efficiency. Consequently, I provide experiences of high variability, of true low risk and in which neither success on the first attempt nor performance of the conventional "right" response are rewarded. Instead, attempts to apply novel procedures are reinforced. This results in a mode of instruction that is considerably slower than a conventional step-by-step demonstration, but ultimately results in development of functions that are never achieved by the conventional approach.

Use the tool to teach the posture, not just the task.

My most effective teaching strategy is also my simplest, and achieves maximum effect in settings of the lowest expectations. For example, students enrolled to learn either Excel or Teams received instead a complete experience of integration with the additional module of Copilot. When students asked technical questions concerning methods to obtain desired results from this tool, I simply replied, "Exactly. Let's ask Copilot." On the surface, this was an evasive response. In reality, four major effects were produced simultaneously.

I am allowing Copilot to demonstrate its signature non-determinism, and will not get the same answer twice. That is the fun, not the problem.

I am encouraging the room to push and experiment a little with solution-oriented tinkering rather than await the proper magic words.

Everyone will run the same prompt and compare results with the person sitting next. Two identical questions yield two completely different answers and a brief appreciation of how capricious the underlying machine actually is.

Finally, they will experience firsthand what an AI conversation feels like in the brain. It is not like doing Excel, where the correct path of operations produces an exact reward. Instead, deviation from the expected path produces an unexpected reward.

The resistance is real and deserves to be named honestly. The response I fear and sometimes receive is "I didn't sign up for AI training." That is a legitimate complaint on its own terms, because they signed up for a course in spreadsheet manipulation. But the truth to which I have finally made peace is that none of us signed up for what I now experience. And yet here we are. My "let's try Copilot" experience represents the smallest possible dose of the required change in perspective.

It requires no effort, embarrasses no one and provides an inadvertent experience of the experimental attitude in the midst of doing exactly what they had originally intended to accomplish. Teach the art of source selection. That is where the utility of my information-literacy programs is most clearly demonstrated. Place a student in contact with a system such as NotebookLM and the result is not "see what the machine can do to produce text." Rather, it is "see how dependent on my choices of input the results of my work really are." The skills of source selection, of defining limits of coverage, of making judgements about what to trust, reflect exactly those skills for which librarians have always been responsible. And reward every instance of curious experimentation.

My experience with the occasional amateur who tries once and then stops elicits my sympathy but not my optimism. In contrast, the person who continues to experiment, and interprets the system as an opportunity for reverse engineering rather than as a vending machine, achieves a level of intellectual engagement even when the results remain rough. Reinforce the associated attitude of professional respect. This represents by far the most valuable and least teachable component of my entire experience, and is the closest I can come to teaching it. Finally, this is the experience of thinking that survives to guide my future efforts.

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.