Sixteen experienced developers agreed to an experiment. They chose real tasks from their own repositories, codebases they had worked on for an average of five years. A researcher randomised each task: use AI, or work without it. Before starting, the developers predicted AI would make them about 24 percent faster.
They finished 19 percent slower.
The important number is the one that came afterwards. When the study ended, those same developers estimated that AI had sped them up by about 20 percent. They had lived through the slowdown, recorded their screens while it happened, and still could not feel it.
Treat the slowdown itself as provisional. METR, the nonprofit that ran the trial, now labels that 2025 result historical, and a later cohort using newer tools landed close to break even. Tools improve. What has not improved is the thing the study accidentally measured: the distance between how good work feels and how good it is.
AI closed the gap on production almost overnight. It did nothing at all to the gap on evaluation.
That distance is the most valuable real estate in the labour market right now, and almost nobody is trained to occupy it.
For most of economic history, the expensive part of knowledge work was making the thing. Research, drafting, analysis, design, code: all slow, all costly, and being able to do any of it competently was enough to be paid. That cost is collapsing. What does not collapse is the cost of deciding whether the thing is right, whether it was the right thing to make in the first place, and who answers for it when it turns out not to be.
When answers become cheap, judgment becomes expensive.
This article is not going to tell you that AI will never replace you. That claim is comforting, unfalsifiable, and possibly untrue. It is going to do something more useful: look at what the current evidence says about which human capabilities are gaining value as machine capability rises, explain why the economics work that way, and give you a specific method for building each one.
What follows
- Judgment: knowing whether an answer is worth acting on
- Problem framing: knowing what deserves to be asked
- Taste: choosing well when everything is available
- Emotional intelligence: being believed, not just sounding warm
- Trust: the scarce asset in a synthetic world
- Adaptability: becoming a beginner on purpose
- Agency: owning what the machine produced
First, what the evidence actually says
A great deal of writing about AI and jobs collapses seven different questions into one. It is worth pulling them apart, because they have different answers.
Capability is what a model can do in a test. Adoption is what people actually use it for. Exposure is how much of a job consists of tasks a model could plausibly do. Task automation is how much of that is actually being handed over. Displacement is people losing work. Transformation is the job changing while the person keeps it. And prediction is guessing. Most alarming headlines take a number from one column and print it under the heading of another.
Start with the widest measurement available. The International Labour Organization and Poland’s NASK scored nearly 30,000 occupational tasks and mapped them across more than 140 countries. Their 2025 finding: roughly one in four workers worldwide is in an occupation with some generative AI exposure, but only 3.3 percent of global employment sits in the highest exposure band. Exposure runs at about 34 percent of employment in high income countries and 11 percent in low income ones. Their conclusion is blunt and important: because most occupations contain tasks that still require human input, transformation of jobs is the likely outcome, not mass redundancy.
Exposure is not displacement. But it is not nothing either.
Where displacement is showing up
Stanford’s Digital Economy Lab has been tracking payroll records from ADP covering millions of American workers. Their August 2026 revision of Canaries in the Coal Mine reports no evidence of widespread, economy wide job displacement. It also reports that employment of workers aged 22 to 25 in the most AI exposed occupations now sits about 19 percent below where it would be had it kept pace with similarly aged workers in less exposed jobs. The gap has widened steadily since they first documented it in 2025. It operates mainly through reduced hiring rather than layoffs, and it is concentrated in occupations where AI usage substitutes for human tasks. Where AI mainly complements workers, employment is flat or rising, especially for experienced people.
Not everyone agrees on the cause. The Economic Innovation Group has argued the pattern reflects the sharpest interest rate tightening cycle in four decades rather than AI. The Stanford authors tested that: the most rate sensitive occupations, such as construction, have among the lowest AI exposure. The disagreement is live and worth watching.
Read those two findings together and the shape becomes clear. The dividing line in this labour market is not industry, seniority, or even skill level in the usual sense. It is whether the AI in your workflow substitutes for what you do or amplifies it. That is the difference between being the person who directs the work and being the person who was the work.
Which raises the obvious question: what puts you on the amplifying side?
Three large datasets point in the same direction, from three different angles.
WEF Future of Jobs 2025
Microsoft Work Trend Index 2026
Anthropic Economic Index, June 2026
Microsoft’s 2026 research, based on trillions of anonymised productivity signals and a survey of 20,000 AI users across ten countries, describes four patterns of human and machine collaboration that have emerged first in software and are now spreading across other functions: author, where you produce and call on AI for a line here and there; editor, where AI drafts and you approve; director, where you write a specification and hand off whole tasks; and orchestrator, where you design a system of agents running in parallel and handle the exceptions. Their summary of what changes as you move along that scale is the single most useful sentence in the whole report: what declines is tactical, step by step execution done by humans, and what rises is the need for humans to set direction, define standards, and evaluate outcomes.
Anthropic’s usage data supports the same reading from the other end. Its June 2026 Economic Index found that in conversations mapped to higher wage occupations, the model produced more output per turn and the human took more turns. More machine involvement did not mean less human involvement. The researchers note that when the human stays engaged in the highest value tasks, the pattern looks labour augmenting rather than labour displacing.
One caution before we go further. The Microsoft and WEF numbers are surveys: they measure what employers and workers believe, not what has happened. The Anthropic figures come from a self selected population of people who already use Claude. The Stanford payroll data is the hardest evidence in the set, and even its authors are careful about causation. Anyone who tells you the future of work is settled is selling something.
What follows are seven capabilities that sit on the expensive side of the line the evidence keeps drawing. They are ordered by widening radius: from judging a single output, to framing a single problem, to choosing among many options, to one other person, to many people, to yourself over time, and finally to the whole system you are responsible for.
SKILL 01
Judgment: knowing whether an answer is worth acting on
Getting cheaper
Producing a fluent, confident, well formatted answer.
Getting expensive
Knowing whether that answer should change what you do.
Why it stays scarceThe system that generated the output cannot be the final check on the output. Verification has to come from somewhere with independent access to reality.
Generating an answer and judging an answer are different operations, and only one of them just got cheap.
For most of working life, fluency was a usable proxy for competence. A polished brief with fifteen citations was expensive to fake, so its polish carried information. That proxy has broken. Polish is now free, and it arrives with no signal attached about whether the underlying claim is true.
The clearest documentation of what happens next comes from courts, because courts write everything down. Legal researcher Damien Charlotin maintains a public database of decisions in which a judge found that a party had relied on AI fabricated material. It listed roughly 200 cases in mid 2025. By August 2026 it had passed 1,900. Penalties have run from a 5,000 dollar sanction in the first well known case, Mata v. Avianca in 2023, to roughly 110,000 dollars in a single 2026 matter, along with dismissals, public reprimands and the first licence suspensions.
Three patterns repeat across those rulings, and none of them are really about law.
First, the failure is almost always retrieval, not reasoning. Models are far weaker at locating authority that exists than at analysing authority once you give it to them. Fabricated citations are formatted perfectly, which is precisely why they survive review.
Second, purpose built professional tools hallucinate too. Several sanctioned filings came from paid legal AI products, not consumer chatbots.
Third, and most instructive, the harshest penalties land on the cover up rather than the error. Courts have converged on a standard that generalises to every profession: no document should carry a claim the responsible person has not personally checked.
The duty did not move when the tool arrived. Only the volume of unchecked output did.
This is not a lawyer problem. It is an analyst pasting a market size into a board deck, a student citing a study that does not exist, a founder repeating a competitor statistic that a model invented, a clinician accepting a summary of a paper nobody opened.
There is a subtler failure mode that matters more in ordinary work. Often there is no factual error at all. There are simply four defensible answers, and only one of them fits your situation.
Ask a model which marketing channel to cut, and it will give you a well argued answer. Ask again with a different attribution window and it will give you a different well argued answer, equally well argued. Both are technically correct. Choosing the window that matches how your business actually makes money is not a retrieval task. It is judgment, and it depends on knowing things about your company that were never written down anywhere the model could read.
Harvard Business School researchers found a version of this with management consultants using GPT-4. Inside the range of tasks where the model was strong, consultants improved substantially. Outside it, they got worse than the control group, because plausible output is very hard to argue with when you have no independent way to check it.
Which points at where judgment actually comes from. It is not a personality trait and it is not general intelligence. It is calibration, and calibration is built by making explicit calls, recording them, and finding out what happened. Most professionals accumulate experience without ever compounding it, because they never wrote down what they expected. Twenty years of unrecorded decisions produces confidence, not accuracy.
There is one bias worth naming specifically, because AI amplifies it more than any technology before it. Confirmation bias is the tendency to seek and accept evidence that supports what you already think. A system that will argue any position with equal fluency, on demand, instantly, is the most efficient confirmation bias engine ever built. It will never tell you that you asked a self serving question.
How to train judgment
- Answer before you ask. Write your own conclusion and a confidence number from 0 to 100 before you prompt. Then compare. The gap is your training signal, and it only exists if you write it down first.
- Keep a decision log. One line per real decision: what you decided, what you expect to happen, how confident you are, and the date you will check. Review monthly. This single habit does more for judgment than any course.
- Set a verification budget by consequence. Ask what it costs if this is wrong, then spend checking time in proportion. A social post and a regulatory filing do not deserve the same scrutiny, and treating them alike wastes effort in one direction and invites disaster in the other.
- Never verify a claim with the system that produced it. Check against a primary source, a named human, or a system with different failure modes.
- Ask for the strongest case against, then check that too. A counterargument you did not verify is just a second unverified answer.
SKILL 02
Problem framing: knowing what deserves to be asked
Getting cheaper
Getting an answer to the question you asked.
Getting expensive
Discovering that it was the wrong question.
Why it stays scarceA model inherits your framing and optimises it beautifully, including when the framing is wrong. Nothing in the output tells you that.
For most of human history, getting answers was the expensive part. Libraries, experts, consultants, research teams: all of it existed because answers were costly to obtain. That cost is falling towards zero. The bottleneck has moved upstream, to a skill that was never scarce enough to be trained deliberately.
There is an unusually clean piece of evidence for this in Anthropic’s usage data. Its researchers score how much decision making authority a user hands over. The lowest autonomy outputs are translations, calculations, and direct questions. The highest are apps, websites, games, and presentations. The pattern is not about difficulty. It is that in the first group, the question fully determines the answer, and in the second, it does not. The moment a task is underspecified, someone has to supply the missing judgment. Either the model supplies it silently, or you supply it deliberately.
The same report contains a detail worth sitting with. The median chat conversation that produced a blog post involved thirteen rounds of back and forth. The median coding agent session that produced one ran on a single human prompt. Same output category, completely different amount of human specification. Neither is wrong. But if you never notice which of the two you are doing, you have stopped choosing.
A weak operator asks for another campaign. A strong one notices the problem is not advertising at all.
Consider four versions of the same failure:
- The marketing team asks AI for a new campaign. Acquisition is not the problem. Sixty percent of customers leave in month three, which means every new customer is being poured into a bucket with a hole in it. More campaigns make the bleeding faster.
- The engineer asks AI to build the requested feature. It ships perfectly. Four users touch it. Nobody asked whether the request came from the loudest customer or the typical one.
- The founder generates a hundred business ideas in an afternoon. The scarce thing was never ideas. It was finding the one problem a specific group of people already pays badly to solve.
- The student asks for an essay on a topic. The examiner was assessing whether the student could identify what was actually contested in that topic. The essay answers the wrong exam.
Psychologists call the underlying trap Einstellung, the effect where knowing a method makes you reach for it before you have understood the situation. Cheap methods make the trap cheaper to fall into. When generating a solution costs nothing, the temptation to skip diagnosis is enormous, because diagnosis is the only part that still feels slow.
Doctors are trained against exactly this, which is why the medical framing is useful even if you never treat a patient. The valuable move is not producing treatments. It is separating the symptom from the mechanism. Falling revenue is a symptom. Pricing, positioning, churn, distribution and sales capacity are mechanisms, and they require completely different responses. AI will happily write you a plan for whichever one you name.
How to train problem framing
- Write five versions of the problem before writing any solution. The first two will be the obvious framings. The fifth is usually where the useful one lives, because you had to work past the obvious ones to reach it.
- Name the decision the answer will change. If no decision changes regardless of the answer, you are researching for entertainment. Stop and find the real question.
- Identify the binding constraint. Time, money, trust, attention, distribution, or permission. Most plans optimise a constraint that was never binding.
- Ask what would have to be true for this to be the wrong problem. Then check whether any of those things are true. This is a five minute habit that has saved entire quarters.
- Write the brief you would give a competent new hire before you delegate to a model. If you cannot write it in a paragraph, you do not yet understand the problem, and the model is about to hide that from you.
SKILL 03
Taste: choosing well when everything is available
Getting cheaper
Producing a hundred competent options.
Getting expensive
Knowing which one to keep and why.
Why it stays scarceWhen everyone starts from similar suggestions, the average becomes visible and the exception becomes valuable. Selection, not production, is now the bottleneck.
The weak version of this argument says AI cannot be creative. That is not true and has not been true for some time. Models produce output that people rate as novel, well made and enjoyable. Pretending otherwise makes the rest of the argument easy to dismiss.
The strong version is more interesting, and it has direct experimental support.
What happens to creative output at scale
Anil Doshi and Oliver Hauser ran an experiment published in Science Advances in which 293 writers produced short stories, some with access to AI generated story ideas, and 600 evaluators assessed the results. Stories written with AI assistance were rated more creative, better written and more enjoyable, and the gain was largest for the least creative writers. The stories were also measurably more similar to each other than the human only stories were.
Individual quality rose. Collective variety fell. The authors describe it as a social dilemma: each writer is better off using it, and the pool of work becomes narrower.
Translate that into market terms and you get the defining condition of creative work for the next decade. The floor rises fast. The ceiling barely moves. The distribution compresses.
Competent stops being a differentiator when competent is free. What becomes scarce is output that is both different and good, which is a much harder pair of conditions than either one alone. Different and bad is easy. Same and good is now automatic.
This is where taste earns its keep, and it is worth being precise about what taste is, because the word gets used as a synonym for personal preference. It is not. Taste is the ability to predict how something will land on a specific audience in a specific moment, combined with the willingness to discard things that are perfectly fine.
Most of it is subtraction. Editors, art directors, film editors and the good product managers are paid overwhelmingly for what they remove. That skill has no obvious output, which is exactly why it has always been underpriced and is about to stop being.
Where does AI already do part of this? Generation, variation, style range, first passes, and increasingly critique. Used well it is a genuinely useful second opinion. Where it is weakest is knowing what has already saturated your specific market this year, what your particular audience has grown tired of, and what would feel cheap coming from you specifically. Those are all context problems, not capability problems, and context is the thing you own.
Taste develops through exposure with attention, and the mechanism is articulation. Vague admiration transfers nothing. A stated reason becomes a rule you can apply tomorrow. This is why studying under a demanding editor works and why scrolling through beautiful work does not.
How to train taste
- Ten and ten, with reasons. Collect ten excellent and ten mediocre examples in your field. For each, write one sentence naming the specific mechanism that makes it work or fail. Not “it flows well”. Something like “the first line makes a claim the reader wants to argue with”.
- Predict before you measure. Choose which of three options will perform best, write it down, then run them. You are calibrating a forecast, not expressing a preference.
- Generate twenty, kill nineteen, log the kills. The reasons for rejection are the actual asset. After thirty rounds you will have a personal standard nobody else has.
- Cut ten percent last. Apply it to every draft, deck, design and video. The version that survives the cut is almost always the better one.
- Once a month, make something the model would not have suggested. Then check whether it lands. This is how you learn the difference between originality and self indulgence, which is a distinction nobody can teach you in the abstract.
SKILL 04
Emotional intelligence: being believed, not just sounding warm
Getting cheaper
Expressing empathy in words that land well.
Getting expensive
Being the person whose empathy is believed.
Why it stays scarceEmpathy works partly as a costly signal. It carries information because producing it takes something from the giver. Automate the signal and it stops predicting anything.
This is the section where sentimental writing usually takes over, so let us start with the finding that makes sentimentality impossible.
AI already beats humans at expressed empathy, under controlled conditions, on average.
Researchers at the University of Toronto ran four preregistered experiments published in Communications Psychology. Third party evaluators rated AI generated responses as more compassionate and more responsive than human ones, including responses from trained crisis responders. The preference held when authorship was disclosed. A separate study in PNAS found that AI written messages made recipients feel more heard than human written ones, partly because the model had more discipline: it acknowledged the feeling instead of rushing to fix the problem.
Anyone whose argument for human value rests on machines being unable to sound caring should update now.
Then comes the twist, and it is one of the most useful findings in the whole literature on AI and work.
The empathy attribution gap
A team led by Anat Perry at the Hebrew University of Jerusalem, with collaborators at Harvard and the University of Texas, ran nine studies with 6,282 participants, published in Nature Human Behaviour. Every response was AI generated. Half were labelled as coming from a human, half from an AI. The identical words were rated as more empathic, more supportive and emotionally more satisfying when people believed a human wrote them. The effect was strongest for responses expressing shared feeling and care. Participants’ own unprompted suspicion that AI had helped write a “human” response reduced perceived empathy and support.
A 2026 paper in Communications Psychology names the resulting behaviour the AI empathy choice paradox: given the choice, people ask for human empathy while rating the AI version higher. Some will wait days for the human reply rather than take the instant one.
Why would people behave that way? The best available explanation, argued in Trends in Cognitive Sciences in 2026, is that empathy has always functioned as a costly signal. When someone attends to your situation, it tells you something predictive about their future behaviour: that you have their attention, that they will probably show up next time, that the relationship carries weight. Remove the cost and the signal stops predicting anything. This is not sentimentality. It is information theory applied to relationships.
Sounding warm has become a commodity. Being believed has not.
That reframes the practical skill entirely. Producing compassionate language is no longer the differentiator, because your competitor can produce it too, instantly, at scale, in nine languages. What remains scarce is the substrate underneath it: sustained attention, memory of what this person told you three months ago, willingness to take a social risk on their behalf, accurate reading of what is being avoided, and the ability to stay regulated while someone is angry at you.
The gaps show up in ordinary situations. A manager receives a flawless AI drafted performance review and delivers it without noticing the person stopped listening in the second sentence, because they heard one word and started defending themselves. A salesperson with perfect product knowledge loses to someone who understood that the buyer’s real fear was looking foolish in front of their own boss. A teacher covers exactly the right material and misses that a student went quiet out of embarrassment rather than boredom. A founder reads a hundred survey responses and never learns why people actually cancelled, because nobody writes “it made me feel incompetent” in a text box.
The most experienced workers seem to know this. In Anthropic’s June 2026 survey, when respondents were asked what AI would never be able to do, those with fifteen or more years of experience disproportionately named the relational parts of their jobs: building trust and managing people. That is a belief rather than a measurement. But it is a belief held by the people standing closest to the work.
Be careful with the conclusion though. None of this proves machines can never take the relational ground. Attribution effects can shift, and a generation raised with these systems may not devalue machine empathy the way current participants do. The defensible claim is narrower and more useful: today, attribution matters enormously, and attribution attaches to people.
How to train emotional intelligence
- Summarise before you respond. In any difficult conversation, state the other person’s position until they agree you have it right. Perspective taking is a trainable skill and this is the drill. It also makes it impossible to argue with a version of them you invented.
- Ask what they need before offering what you have. Most failed conversations are two people solving different problems politely.
- Watch the gap between the stated objection and the real one. The stated objection is usually price or timing. The real one is usually risk, status, or the fear of looking stupid in front of someone who matters.
- Add the detail only you could know. If you use AI to draft something emotionally loaded, rewrite it in your own voice and include one specific thing you remember. That detail is the entire signal.
- Regulate first. If your pulse is up, you are not reading the room, you are defending yourself. Delay by an hour. Almost nothing emotionally important gets worse from an hour.
SKILL 05
Trust: the scarce asset in a synthetic world
Getting cheaper
Producing persuasive words, at any volume, for anyone.
Getting expensive
Being believed by people who have other options.
Why it stays scarceVerification costs rose for everybody at once. When any message might be synthetic, the value migrates from the message to the sender.
Words got cheap. Belief did not.
A chief executive can generate a flawless strategy document before lunch. Getting five hundred people to change what they actually do on Monday morning is exactly as hard as it was in 2019. Microsoft’s 2026 research puts a number on where the constraint sits: organisational factors such as culture, manager support and talent practices explained roughly twice as much of AI’s measured impact as individual mindset and behaviour did. The bottleneck on AI value inside companies is not the model. It is coordination, and coordination runs on trust.
Meanwhile the ground under trust is moving.
What synthetic media has already cost
For the first time in roughly twenty five years of reporting, the FBI’s Internet Crime Complaint Center broke out AI enabled fraud as its own category in its 2025 report: 22,364 complaints and about 893 million dollars in adjusted losses. The World Economic Forum’s Global Risks Report 2026 again ranked misinformation and disinformation as the most severe short term global risk. In the best documented corporate case, the engineering firm Arup lost roughly 25.6 million dollars after an employee joined a video call in which every other participant, including the apparent chief financial officer, was synthetic.
Treat the wilder deepfake statistics circulating online with suspicion. Many come from vendors selling detection products, share no baseline, and cannot be reconciled with each other. The government figures above are among the few that have been independently audited.
The most instructive story in this area is not a loss. In 2024, an executive at Ferrari received messages and then a call that appeared to come from the chief executive, complete with a convincing voice and the right regional accent. Something felt slightly wrong. So the executive asked a question the impersonator could not have answered: the title of a book the real chief executive had recommended a few days earlier. The call ended immediately.
That is the future of verification compressed into one move. Shared history cannot be synthesised. Everything else about a person now can be.
In a market flooded with fluent strangers, the premium moves from the message to the sender.
The consequences run through every kind of professional work. Cold outreach depreciates, because it is now indistinguishable from automated outreach at scale. Warm introductions appreciate. Anonymous credibility depreciates. Named, accountable credibility appreciates. Claims that cannot be checked depreciate, because generated claims are also unfalsifiable. Specific, dated, checkable claims appreciate.
Which means the communication skills that matter are not the ones that produce more words. They are the ones that build a record: saying the unpopular thing early enough to get credit for it, being clear under scrutiny, teaching well enough that people learn, negotiating in a way that leaves the other side willing to deal with you again, and putting predictions in public where they can be checked against reality.
One warning worth stating plainly. Clear writing and clear thinking are close to the same activity. Outsourcing the writing can quietly outsource the thinking, and the loss is invisible because the output looks fine. That is not an argument against using these tools. It is an argument for being deliberate about which parts of the thinking you are willing to stop doing.
How to train communication and trust
- One sentence, one paragraph, five minutes. Practise the same idea at three lengths. If you cannot do the sentence, you do not have the idea yet, you have a topic.
- Write the objection before the pitch. Then answer it in the pitch. Audiences relax when you have clearly already thought of the thing they were about to say.
- Publish predictions with dates. Being right in public, on the record, over years, is one of the few remaining assets that cannot be generated.
- Build a second channel habit and tell people about it. For anything that moves money, access or sensitive data, verify through a separate channel you established in advance. A visible verification ritual is now a trust product, not a nuisance.
- Say the number. Vague claims read as generated. Specific, checkable ones do not.
SKILL 06
Adaptability: becoming a beginner on purpose
Getting cheaper
Mastery of any specific tool, interface or workflow.
Getting expensive
The ability to be visibly incompetent again, on schedule.
Why it stays scarceTool specific skill now depreciates faster than a career lasts, and most people defend their professional identity instead of updating it.
Telling people to keep learning is not advice, it is a slogan. The useful part is the mechanism.
The half life of a specific interface is now measured in months. The half life of the underlying capability is measured in decades. Someone who becomes extremely good at one tool and cannot transfer will be obsoleted repeatedly, each time by something that took a week to learn. The WEF’s employer survey shows what this looks like from the hiring side: 39 percent of core skills expected to change by 2030, with curiosity and lifelong learning among the fastest rising requirements. That is what a survey looks like when nobody, including the employers, knows what the tools will be.
There is a second finding that changes how you should think about time. Anthropic’s March 2026 report found that experienced users get materially more out of these systems than newcomers do. Skill with AI is a learning curve, not a switch you flip. The compounding happens in the person, not the tool, which means the cost of starting late is higher than it looks and the benefit of starting badly is higher than it feels.
But the real obstacle is not learning. It is identity.
“I am a writer” makes every improvement in machine writing an attack. “I solve communication problems” makes the same improvement a windfall.
Professional identity built around an output puts you in direct competition with the thing getting cheaper. Identity built around a problem puts you in charge of a newly cheap input. Same person, same skills, opposite emotional response, and therefore opposite behaviour.
“I am a programmer” versus “I build systems that work, and code is one way I do it.”
“I am a designer” versus “I make things people can understand and want to use.”
“I am an accountant” versus “I make sure the numbers are right and the decisions based on them are sound.”
This is not positive thinking. Loss aversion is well documented: people fight roughly twice as hard to protect what they have as to acquire something equivalent. If your identity is the thing being automated, every capability release registers as a loss, and losses trigger defence rather than curiosity. Reframing what you own is the cheapest intervention available, and it takes about ten minutes.
One honest complication. The traditional way people built judgment was by doing the simple work first: the junior tasks, the first drafts, the routine analysis. That is exactly the layer being automated, and the Stanford payroll data suggests those roles are thinning through reduced hiring. If the apprenticeship is disappearing, learning has to be constructed on purpose rather than absorbed by proximity. Anthropic’s survey found most respondents believe they are learning more with AI, not less, but the report notes that self assessments cannot rule out skill erosion. Believing you are learning and learning are two different things, which is the theme of this entire article.
How to train adaptability
- Rewrite your job description as a problem, not an output. Say the new version out loud to someone who knows your work and see whether it survives contact.
- Once a quarter, spend ten hours being visibly bad at something. The subject barely matters. The capacity being trained is tolerance for the beginner phase, which is what most people actually lack.
- Do one delegated task manually every month. It is the only reliable way to find out whether your own capability is still there or has quietly left.
- Keep a “what I was wrong about” list. One entry a month. Belief updating is trainable and almost nobody practises it, which is why almost everybody is bad at it.
- Delete one workflow per quarter. Obsolete methods survive because nobody is ever assigned to kill them. Assign yourself.
SKILL 07
Agency: owning what the machine produced
Getting cheaper
Execution. Doing the steps. Producing the deliverable.
Getting expensive
Deciding what should exist, and answering for it.
Why it stays scarceAccountability has no machine readable format. Institutions assign it to people, and they change slowly and for reasons of their own.
Execution is not agency, and confusing the two is the most expensive mistake available right now. Leverage without agency does not produce better outcomes. It produces more output, faster, in a direction nobody examined.
Someone still has to decide what you are trying to achieve, what the machine is allowed to do unsupervised, what stays human, which trade offs are acceptable, when to override the recommendation, and what happens when it goes wrong. None of those are execution tasks. All of them are getting more valuable as execution gets cheaper.
The direction of travel shows in usage data. Anthropic’s enterprise API traffic is overwhelmingly automated, with the vast majority of interactions being single direction task delegation. Consumer chat is roughly an even split between delegation and collaboration. Work migrates to the automated end once the specification is stable, which tells you something precise about your own position: the durable role is not performing the task, it is producing the specification and owning the result.
Microsoft’s framework describes the same movement as a progression from author to editor to director to orchestrator. It comes from a vendor with an obvious interest in the conclusion, which is worth remembering, but the shape is corroborated by independent usage data and by the payroll evidence about which jobs are holding up.
Then there is the part nobody can delegate, and the courts have made it unusually explicit. Across hundreds of rulings, judges have converged on the position that responsibility for a filing sits with the person who signed it, regardless of what produced the text. “The model said so” has failed as a defence every single time it has been tried, in every jurisdiction.
That is not a claim about machine capability. It is a claim about how human institutions assign blame, and institutions are conservative by design. Anywhere consequences are real, someone’s name goes at the bottom of the page, and the value of being that someone is rising in step with how much the machine underneath them produces.
The unit of work is becoming the brief, not the task. Learn to write briefs.
There is a harder half of agency that gets less attention. When the cost of doing more collapses, the discipline of doing less becomes the main lever. Every organisation now has the capacity to produce ten times the content, twenty times the analysis and fifty times the features. Almost none of them have ten times the strategic clarity. The constraint moved from capacity to choice, and choice is a responsibility that cannot be automated without also handing over the goals.
How to train agency
- Write four lines before you delegate anything, to a person or a model: the outcome you want, the constraints that cannot be violated, what good looks like specifically, and what to do when stuck.
- Define the stop condition. What result would make you cancel this project? Decide before you are emotionally invested, because afterwards you will not.
- Name the owner in writing. If it is you, say so. Diffuse ownership is how bad outputs reach customers.
- Run a five minute pre mortem. It is six months later and this failed badly. Write the three most likely reasons. Then remove one of them now.
- Keep the last mile human wherever consequences are irreversible. Money, safety, health, legal exposure, reputation, anything involving another person’s data. Automate the ninety percent that is reversible and guard the ten percent that is not.
The other half of the equation
So which AI skills should you actually learn?
Nothing above is an argument for staying away from these tools. It is the opposite. The formula is not human skills or AI skills. It is human judgment multiplied by AI leverage, and multiplication is unforgiving about zeros.
Excellent judgment plus refusal to use the tools loses to good judgment plus fluent use. Fluent use plus no judgment, no curiosity and no ability to work with people also hits a ceiling, and hits it fast, because that person’s entire contribution is the part getting cheaper.
Most people do not need to become machine learning engineers. What they need is operator literacy, which is a much smaller and more learnable thing. Six components cover almost all of it:
- Specification. Writing a brief a model can execute: outcome, constraints, audience, format, what to avoid, what to ask about rather than assume.
- Verification. Knowing the characteristic failure modes, which are fabricated sources, confident retrieval errors, silent staleness and quietly dropped constraints, and having a cheap check ready for each.
- Workflow design. Deciding which steps to automate, which to keep human, and where the review gate goes. The gate placement matters more than the tool choice.
- Data hygiene. Knowing what must never be pasted into a third party system: client data, health information, unreleased financials, other people’s private details, anything under contract.
- Tool range. Enough breadth to know which class of tool fits a problem, with no loyalty to any single product, because the leaderboard changes every few months.
- Economic sense. Knowing what the work is worth, so you do not spend ninety minutes automating a task that takes two.
The evidence for the multiplication effect is reasonably good. Anthropic’s data shows experienced users succeed at delegation at materially higher rates than newcomers. Microsoft identifies a group it calls Frontier Professionals, only about 16 percent of AI users, who orchestrate multi step workflows and redesign how their work happens: 80 percent of them say they are producing work they could not have produced a year ago, against 58 percent of AI users overall. Both figures are self reported and correlational, so read them as a signal about direction rather than a measured return. The direction is consistent across every dataset in this article.
THE OBJECTION
What if AI gets good at all seven?
This is the question that should be asked of every article like this one, and most of them dodge it.
Take it seriously, because in places it has already happened. AI outperforms trained crisis responders on rated compassion. It produces creative work that readers prefer. It is closing quickly on tasks that looked like pure judgment three years ago. In Anthropic’s June 2026 survey, more than a third of respondents expected AI to handle most or nearly all of their work tasks within twelve months, and those expectations were remarkably uniform across occupations. A software engineer and a construction manager anticipated roughly the same rate of progress in their own fields.
So the claim in this article is not that these seven capabilities are permanently, metaphysically human. Drawing a line around supposedly uniquely human skills has been a losing bet for a decade straight, and there is no particular reason this round is different.
The claim is narrower and it is economic. As some capabilities become abundant, value moves towards whatever remains scarce and complementary. Today, and for the visible future, these seven sit on the scarce side. That could change, and an honest version of this argument says so out loud.
Three things are more durable than capability, though none of them are guarantees.
Legitimacy. Who is permitted to decide is a social question, not a technical one. A system can produce a better sentencing recommendation, a better loan decision or a better diagnosis and still not be allowed to make it alone, because permission is granted by institutions rather than earned by accuracy.
Accountability. Someone has to be exposed to the consequence. Arrangements where no one is responsible tend to get regulated until someone is. This is a structural feature of how societies handle risk, and it has survived every previous automation wave.
Relationship. Value that depends on being a particular person, known to another particular person, over time. The Ferrari executive did not defeat a deepfake with better technology. He used a shared memory.
Now the part that is genuinely uncomfortable, and that most optimistic writing on this subject leaves out.
The gains are not being shared evenly, and the cost is landing on a specific group. The clearest labour market signal available says young workers in exposed occupations are being asked to demonstrate judgment before anyone will pay them to develop it. The rungs of the ladder that used to build judgment are exactly the rungs being automated. No amount of individual skill building fixes that at population scale. It needs employers who deliberately rebuild apprenticeship, and policy that treats this as a transition rather than a rounding error.
If you came here for a guarantee, there is not one. There is a defensible position, which is a different and more useful thing.
Key takeaways
- Exposure is not displacement. The ILO finds one in four workers globally in an exposed occupation, but concludes transformation is the likelier outcome for most of them.
- The dividing line is substitution versus complementarity. Where AI substitutes for tasks, entry level employment is falling. Where it complements, employment is flat or growing.
- Evaluation is the new bottleneck. Production got cheap in one step. Verification, framing, selection and accountability did not move at all.
- Signals lose value when they become free. This applies to polish, to fluency, to expressed empathy and to persuasive writing. It is why trust, track record and shared history are appreciating.
- The formula is multiplication. Human judgment multiplied by AI leverage. Either factor at zero produces zero.
What is actually left
For two centuries, the reliable way to be valuable was to know things other people did not know, and to make things other people could not make. Education was built around the first. Careers were built around the second.
Both advantages are being priced down at the same time, in the same decade, for nearly everyone at once. That has not happened before, and the people telling you it will be painless are guessing, as are the people telling you it will be catastrophic.
What is left is smaller, harder and considerably more interesting than what is going away. Deciding what is worth knowing. Deciding what is worth making. Working out what can be trusted and what a specific person actually needs. Choosing among a hundred available options and defending the choice. Standing behind the result when it fails.
That is not a consolation prize. It is a job description, and almost nobody has been trained for it, which is precisely why it pays.
Nobody wins a race against a machine at being a machine. The opening is on the other side of the work, where the questions get chosen, the options get rejected, and someone’s name goes on the outcome. That part was never really production. It was always responsibility, and responsibility has not become abundant.
Sources
- Becker, J., Rush, N., Barnes, E., Rein, D. (2025, updated 2026). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR. metr.org
- Gmyrek, P. et al. (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140, ILO and NASK. ilo.org
- Brynjolfsson, E., Chandar, B., Chen, R. (2025, revised August 2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. digitaleconomy.stanford.edu
- Massenkoff, M. et al. (2026). Anthropic Economic Index report: Cadences. Anthropic, June 2026. anthropic.com
- Anthropic (2026). Anthropic Economic Index report: Learning curves. March 2026. anthropic.com
- Microsoft (2026). 2026 Work Trend Index Annual Report: Agents, human agency, and the opportunity for every organization. microsoft.com/worklab
- World Economic Forum (2025). The Future of Jobs Report 2025. weforum.org
- Charlotin, D. (ongoing). AI Hallucination Cases Database. HEC Paris Smart Law Hub. damiencharlotin.com
- Doshi, A. R., Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10(28). science.org
- Ovsyannikova, D., Oldemburgo de Mello, V., Inzlicht, M. (2025). Third-party evaluators perceive AI as more compassionate than expert humans. Communications Psychology. nature.com
- Rubin, M., Perry, A., Goldenberg, A., Ong, D. C. et al. (2025). Comparing the value of perceived human versus AI-generated empathy. Nature Human Behaviour. nature.com
- Communications Psychology (2026). People choose to receive human empathy despite rating AI empathy higher. nature.com
- Yin, Y., Jia, N., Wakslak, C. (2024). AI can help people feel heard, but an AI label diminishes this impact. PNAS. pnas.org
- Dell’Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School Working Paper 24-013.
- Federal Bureau of Investigation, Internet Crime Complaint Center (2026). 2025 Internet Crime Report.
- World Economic Forum (2026). Global Risks Report 2026.
Figures were current as of publication in August 2026. Several of these datasets are updated continuously and may have moved since.
