There is a phrase that has lived rent-free in the creative world for over a decade. In his 2012 manifesto Steal Like an Artist, Austin Kleon popularised an idea that thinkers from T.S. Eliot to Pablo Picasso had long understood: nothing is entirely original. All creative work builds on what came before. The artists who matter most are not those who conjure ideas from nowhere, but those who absorb their influences deeply enough to transcend them. The job of a creator, as Kleon framed it, is to collect good ideas, understand the logic beneath them, and synthesise something genuinely new from the collision.
It was never a license to copy. It was a call to do the hard, transformative work of internalisation.
Now, in 2026, in the age of generative AI, something is quietly happening to that philosophy — and it should concern anyone who builds things for a living.
We can no longer easily distinguish a photon captured by a camera lens from a pixel generated by a model.
The Craft Behind the Concept
To understand what is at stake, it helps to remember what "stealing like an artist" actually demanded. Kleon was drawing on a long tradition that treats creative lineage as a form of apprenticeship. You study your heroes. You find out who they studied. You trace the genealogy of ideas until you have built a constellation of influences rich enough to produce something that is genuinely, recognisably yours. The synthesis is the work. The synthesis is the originality.
This process is effortful almost by design. It requires a creator to internalise not just what a piece of work looks like, but why it works — the underlying logic, the deliberate decisions, the painstaking choices that make a methodology or a body of knowledge coherent, trustworthy, and alive. It is slow, iterative, and deeply human.
That slowness is not a bug. It is the mechanism by which understanding is transferred.
When Synthesis Becomes Extraction
Generative AI has introduced a new kind of borrowing into creative and knowledge work — one that operates at a structural level most industries have not fully reckoned with yet.
Consider two scenarios.
In the first, a designer studies the portfolios of fifty visual artists, absorbs their principles of composition, and develops a personal vocabulary that draws on all of them without reproducing any. The influences are real, but they have been metabolised. This is the kind of creative debt that has always powered human culture.
In the second, a practitioner takes a painstakingly built proprietary framework — a verified dataset, a meticulously structured methodology, a knowledge base that took months or years of human precision to construct — and feeds it directly into a generative tool. The output scales rapidly. It looks impressively fluent. But the underlying logic that made the source material trustworthy has not been understood. It has been ingested.
This distinction matters enormously, and the research is beginning to catch up with the intuition. A landmark 2024 audit published in Nature Machine Intelligence by MIT and Harvard's Data Provenance Initiative examined more than 1,800 AI training datasets and found licence omission rates above 70% and error rates above 50% in how datasets are categorised and attributed on major hosting platforms. The authors described what they found as a "crisis in misattribution." Research from the same team has since expanded to nearly 4,000 multimodal datasets covering 659 organisations across 67 countries — finding that almost all content sourced from major platforms carries implicit non-commercial constraints that most AI deployments simply ignore.
The infrastructure for acknowledging where knowledge comes from does not yet exist at the scale at which AI is consuming it.
The Architect, the Artist, and the Operator
To understand who is being harmed and who is being celebrated, it helps to name the roles more precisely. They overlap — but they are not the same.
The Architect builds systems. They do the foundational structural work: mapping a domain, testing edge cases, verifying claims, building frameworks that are internally consistent and reliable. A senior developer who not only writes the front-end interface but understands the back-end logic that makes it perform — the database queries, the API calls, the security model, the state management decisions — is an Architect. When something breaks, they know where to look. More importantly, they know why it broke.
The Artist builds meaning. Not separately from technical work, but through it — through craft, voice, accumulated judgement, and the slow, irreducible development of taste. A journalist who has spent a decade cultivating sources is an Artist. So is a composer who understands why a chord resolves the way it does, a filmmaker who grasps how an edit manipulates emotional time, a UX designer who can feel when a user flow is wrong before they can fully articulate why. This knowledge is embodied. It lives in the person. It is not reducible to a methodology.
The Operator takes verified, structured input and uses generative tools to scale it — producing variations, translating formats, generating outputs at volume. Given good source material, the Operator can produce impressive results quickly. Their contribution is real. But it is downstream.
The tension in 2026 is that organisations are disproportionately rewarding the Operator while overlooking — and increasingly eliminating — the Architects and Artists who made the source material worth anything in the first place.
Web Design: Concepts Before Code
Web development and design sit at a peculiar intersection of the artistic and the architectural — and they illustrate the conceptual foundations problem with unusual clarity.
The visible surface of a website is not the work. It is the consequence of the work. Behind every functional, scalable interface is a set of decisions about information hierarchy, data flow, component relationships, and the structural logic that connects what the user sees to what the server does. These decisions are not arbitrary. They emerge from concepts — accumulated, tested, refined through genuine trial and error over years of building things that failed instructively.
Trial and error is how you discover that something works. Concepts are how you understand why — and therefore how you predict when it will stop working, under what conditions it will break, and what it will cost to change it later.
What is happening now in web development is that the trial-and-error phase is being bypassed entirely. Generative tools produce interfaces that look resolved. Templates supply components that function in isolation. AI fills the gaps between them with code that is syntactically coherent and conceptually hollow. A practitioner can ship something that appears complete without ever having acquired the mental models that would allow them to understand what they built — the relationship between front-end behaviour and back-end architecture, why certain design patterns exist, how data flows through a system, what load actually does to structural assumptions that were never stress-tested.
The result is not just fragile. It is unauditable. When something breaks — and it will — there is no conceptual map to navigate by. The person who assembled it cannot diagnose it because they never understood it. The decisions were not made deliberately enough to be documented. Nobody chose this architecture. It accreted — layer on layer of borrowed solutions to problems whose origins nobody traced, each addition making the next harder to reason about.
This is precisely why patchwork does not scale. Not because the individual pieces are poor, but because scalability is a conceptual property, not a technical one. A system scales when it was designed with scale in mind — when someone understood, before building, how each decision would behave under pressure, growth, and change. Without that conceptual foundation, every new requirement becomes a workaround, every workaround becomes a dependency, and every dependency becomes a liability that compounds silently until the system can no longer absorb what is asked of it.
The concepts cannot be prompted into existence. They are earned — through the kind of deep, iterative, often frustrating practice that AI is now allowing people to skip. What gets skipped with them is not just knowledge. It is the capacity to build anything that lasts.
Vibe Coding and the Security Crisis Nobody Planned For
The phenomenon that became Collins Dictionary's Word of the Year for 2025 — vibe coding — is not the villain of this story. It is the revealer.
The term, coined by AI researcher Andrej Karpathy, describes a mode of software development in which you direct AI in natural language, accepting and iterating on generated code fluidly rather than writing every line from scratch. In skilled hands it is genuinely transformative. Senior developers who can evaluate what the AI produces report productivity gains of up to 81%. The ability to describe intent and have working code appear is, for someone who understands architecture, an extraordinary force multiplier — provided the framing is precise, the iteration is disciplined, and the output is interrogated rather than blindly trusted.
The problem is not vibe coding. The problem is vibe coding without concepts.
When the person directing the AI does not understand the system they are asking it to build — when the prompts are vague because the mental model is vague — the results are predictably fragile. A December 2025 analysis of 470 open-source GitHub pull requests found that AI co-authored code contained 1.7 times more major issues than human-written code. Misconfigurations were 75% more common. Security vulnerabilities appeared at 2.74 times the rate of human-written code. Hardcoded secrets in public repositories rose 34% year-on-year, with AI-associated services accounting for over 1.2 million leaked keys. A controlled study built fifteen identical web applications using five major AI coding tools and found sixty-nine vulnerabilities across them, six of them critical.
By March 2026, at least thirty-five newly disclosed security vulnerabilities per month were being directly attributed to AI-generated code. The head of the UK's National Cyber Security Centre told the 2026 RSA Conference that the industry needed to develop vibe coding safeguards as an urgent priority.
The engineer who "didn't write it and couldn't explain it if asked" is not a careless person. They are a human being operating in a system that rewarded speed and punished the time it takes to understand. Vibe coding did not create that system. It exposed it.
The consequences of measuring usage over understanding arrived in March 2026. Amazon had set 80% AI coding adoption as a corporate OKR. When the target was hit, so was the retail site — AI-assisted code changes triggered a series of outages that caused a reported 99% drop in orders across North American marketplaces, an estimated 6.3 million lost. Amazon engineers reported the pressure had "resulted in worse quality code, but also just more work for everyone." Amazon responded with a 90-day code safety reset and mandatory senior engineer sign-off on all AI-assisted code. The target had been met. The architecture of the answer had not been understood.
Practitioners have named what happens next: the Vibe Coding Doom Loop. A bug appears. The AI agent reports it fixed. It is not fixed. Another prompt. Another claimed fix. Still broken. The agent has no model of the system it built — it can only pattern-match against the immediate context of the conversation. Without an Architect who understands the structure, there is no way out of the loop except to start again. Industry analysts now project $1.5 trillion in technical debt will accumulate by 2027 from AI-generated code. Vibe-coded projects accumulate that debt roughly three times faster than traditionally written ones — not because the code looks wrong, but because it lacks the architectural coherence that comes from a human who actually thought through the system design. The "three-month black box" has become a recognised phenomenon: projects that seemed to work at launch become entirely unmanageable once real users arrive and do unexpected things, because no-one on the team understands the AI-generated structure well enough to modify it safely.
The senior engineers who are thriving in this environment are those who have developed something AI cannot replicate: taste. The judgment to know when to trust generated output and when to interrogate it. The ability to write a specification — requirements, architecture, data models, edge cases, test strategy — that constrains what the AI can do before it does it, so that what it produces can be evaluated against a human-defined standard. That is still Architect work. The AI generates within the container. Building the container requires understanding.
The Same Story, Across Every Field
The extractive pattern is not confined to software. It plays out wherever the value of a craft lives in accumulated, embodied judgement — and the specifics matter, because they reveal how consistent the mechanism is.
Journalism. A reporter's value is not the words they type. It is the decade of source relationships, the instinct for what a document is hiding, the legal and ethical accountability that makes published information trustworthy. You cannot prompt that into existence. By March 2026, the journalism industry had already absorbed more layoffs than it recorded in the entirety of 2024 — the Washington Post proposing cuts of roughly one-third of its organisation, along with the Atlanta Journal-Constitution, Politico, Nexstar, Vox Media and others, all in the first two months of the year. AI-generated content is being deployed to fill the gaps. The Washington Post launched a feature for personalised AI-generated podcasts relaying its stories, which immediately produced factual errors and editorialised on developing news. AI is being adopted most heavily in precisely the places least equipped to vet its outputs: resource-strapped local newsrooms. The United States lost on average two local newspapers a week in 2025. Here is the provenance problem in its sharpest form — AI models can only remain factually grounded if they ingest fresh, verified, human-reported information. The industry is systematically eliminating the people who produce that information while deploying systems that depend on it.
Music. A composer understands not just what notes to write but why a harmonic progression resolves the way it does — the emotional logic beneath the technique, the relationship between tension and release that makes a listener feel something without knowing why. A producer knows how a mix breathes, where space creates anticipation, why a particular snare sound belongs to a specific moment in culture. These are not preferences. They are understandings built across thousands of hours of attentive listening and failure. AI can generate music that is superficially plausible. It cannot tell you why a generated piece feels hollow — or fix it when it does. That diagnosis requires the Artist, and the Artist is increasingly being told the generation is enough.
Film and video. An editor does not just select shots. They manipulate the viewer's sense of time, interiority, and causality. A two-frame difference in a cut changes whether a scene feels like memory or dream. A colourist is reading emotional temperature, not just matching values. A sound designer understands that what the audience hears they also feel — physically, in their chest and their attention. These crafts exist at the intersection of technical precision and embodied intuition. AI tools can now produce rough cuts, generate colour grades, synthesise sound beds. What they cannot do is understand why the result doesn't work — because understanding why requires having felt, over years, what working looks like.
Cybersecurity. A security professional is not just running scans. They are modelling the thinking of an adversary — asking not what the system does but what it assumes, where the gaps between intent and implementation live, what a motivated attacker would notice that a developer was too close to see. Threat modelling is an act of imagination grounded in deep technical knowledge. Vibe-coded systems are now producing attack surfaces that no-one on the development team can fully map, because no-one fully understands what was built. The demand for people who can think adversarially about systems they didn't build — who can read generated code with the eyes of someone trying to break it — is not decreasing. It is becoming urgent.
Marketing and brand strategy. A brand strategist's value is not the campaign slogan. It is the exhaustive market research, the deep understanding of consumer friction, and the deliberate construction of a distinct voice that resonates culturally over time. In 2026, marketing departments are flooded with Operators scaling content at near-zero marginal cost using generative tools. The output is massive. It is also conceptually unmoored — mimicking the surface vocabulary of a demographic without understanding the structural nuances of human desire or brand equity. When a high-volume campaign alienates a core audience, the Operator cannot diagnose the failure because they skipped the foundational research and psychological architecture that would have predicted it. They are optimising for algorithmic metrics, entirely decoupled from the human audience those metrics are supposed to represent.
Legal and compliance. A solicitor or compliance architect does not simply locate relevant precedents. Their expertise lives in the precision of risk mitigation — the meticulous interrogation of a contract's hidden liabilities, the deep understanding of regulatory intent, the strategic foresight required to protect an organisation from future litigation rather than just current exposure. Automated document review now produces beautifully formatted compliance matrices in seconds. But these systems identify linguistic patterns, not anomalous systemic risk. When a novel regulatory shift occurs or a hidden contractual liability is exploited, the Operator who deployed the automated system cannot defend the position — because they never built the legal logic from first principles, and they cannot open the box they're holding.
Strategic and innovation roles. The person who identifies a problem before it becomes a crisis, maps the dependencies nobody else has noticed, builds the framework that makes the next five decisions faster — this work is nearly invisible by design. It happens before the deliverable. By the time something is solved, the problem is gone, and what remains is only the solution. This is the role most vulnerable to credit extraction, because the diagnosis is the hardest part and the least documented. When someone takes that solution, presents it as their own insight, and is rewarded for it, the incentive structure inverts: the more valuable your contribution, the less you are seen to have contributed.
What connects all of these is the same structural dynamic. The work that matters most is the work that is hardest to see — the understanding that precedes the output, the judgment that shapes what the tool is asked to do, the taste that recognises when the result is wrong. AI can accelerate the output. It cannot replace the understanding. But organisations in 2026 are systematically removing the understanding while keeping the output pipeline running — and calling it efficiency.
The Illusion of the Prompt
This asymmetry is driven by a profound misunderstanding of what prompt engineering actually is. The prevailing narrative suggests that directing a generative system is a standalone skill — that if a practitioner learns the correct syntax, templates, or phrasing, they can command an AI to execute any complex task reliably.
This is a fallacy. A model reflects the quality of what it is given. A shallow directive yields a statistically plausible, shallow execution. Prompt engineering is not a replacement for expertise. It is a multiplier of it.
To ask a generative tool the right question — the kind that forces a precise, secure, and structurally sound output — requires already knowing the architecture of the answer. The difference is not subtle. Consider how a structural vulnerability gets approached:
The Operator's prompt: "Create a compliance framework and build an automated assessment matrix for this workflow." Shallow, trusting, and prone to confident hallucination where the system's gaps are widest.
The Architect's prompt: "Audit this workflow against Section 4 of the regulatory framework. Interrogate the data ingestion layer for single points of failure, specifically mapping how edge cases are validated, and flag any hidden liabilities."
The Operator can mirror the terminology of an audit once the objective has been defined. They cannot construct the investigative framework from first principles — because they have not spent years learning where systems fracture before they reach crisis point. That knowledge cannot be acquired by reading the output of someone who has it. It is built through the kind of deep, frustrating, iterative practice that AI is now allowing people to skip.
The structural irony is precise: the act of professional scepticism — possessing the diagnostic intuition to recognise that a system is fragile, and knowing exactly how to stress-test it — is itself the moment of genuine innovation. But because this diagnostic work occurs in the abstract phase, before any visible output exists, it is uniquely vulnerable to extraction. A management layer focused on execution can mistake the vocabulary of an audit for the capacity to conduct one. They can ingest the diagnostic concept, feed it into an automated tool, generate a superficial checklist, and misattribute the volume of machine output as strategic insight.
Amazon made this dynamic concrete. Following the deployment of its internal agentic AI platform, MeshClaw, the company introduced targets mandating that over 80 percent of its developers interact with generative tools weekly, tracking consumption via internal team leaderboards. Workers responded with a behaviour that rapidly entered the tech lexicon: tokenmaxxing. To satisfy the mandate and appease monitoring managers, engineers began automating trivial, non-essential tasks — generating meaningless AI interactions and looping automated agents purely to inflate their data consumption metrics.
The metric replaced the work. Volume was measured; value was not. It is a textbook manifestation of Goodhart's Law: the moment a metric becomes a corporate target, it ceases to be a reliable measure of productivity.
This operational breakdown awaits any organisation that conflates fluency with comprehension. When the automated report or codebase inevitably returns filled with complex anomalies, the system is left in the hands of people holding a statistical black box they cannot open — because they only learned how to repeat the question, not comprehend the architecture of the answer.
The Overwork Trap
There is a specific mechanism by which highly skilled people are being consumed rather than just displaced, and it deserves to be named plainly.
It works like this. An organisation identifies someone who can solve problems others cannot. They load that person with an expanding scope of critical work — because the person is capable, and because capable people tend to absorb problems rather than escalate them. The organisation becomes structurally dependent on this person's output without ever formally recognising the dependency. The person is not promoted into the scope they are carrying. They are simply given more of it.
At some point — after the frameworks are built, the automations are running, the institutional knowledge is documented or demonstrated — the organisation makes a calculation. The systems now exist. The problems are mapped. The person who built them is expensive, visible, and harder to manage than a tool. The tools, after all, don't ask questions or notice when credit is misattributed.
So the person is exited. Through redundancy, through restructuring, through the slow withdrawal of support that makes a role untenable. The organisation retains the output of years of expertise while eliminating the cost of the expertise itself. For a quarter or two, it looks like a clean efficiency gain. Then the edge cases arrive. The novel problems. The situations the frameworks were not designed for. And there is nobody in the room who knows why the system was built the way it was.
This is not a new phenomenon. But AI has accelerated it sharply, because it has made the outputs of expertise far easier to capture, codify, and deploy without the person who created them. The consultant who builds the methodology. The engineer who writes the automation. The strategist who maps the problem space. Each of these people has, in some sense, made themselves replaceable in the short term by doing their job well. The tragedy is that what they have actually created is irreplaceable — because what they built next would have been better, faster, informed by what they learned building this. That future output is what gets eliminated with them.
The pattern is not accidental. It is the logical outcome of a culture that rewards the Operator and extracts the Architect.
The Employment Reckoning
These are not abstract observations. They are being settled — badly — in real time.
The global tech sector eliminated nearly 60,000 jobs in the first quarter of 2026 alone, at companies posting record revenues while citing AI as the justification. Amazon cut approximately 16,000 corporate roles while reporting its fastest cloud growth in thirteen quarters. Meta announced 8,000 cuts — 10% of its global workforce. Oracle eliminated up to 30,000 positions. Meanwhile, Google, Amazon, Microsoft, and Meta collectively plan $725 billion in AI capital expenditure in 2026, up 77% year-on-year. Across the industry, 275,000 AI-specialist roles sit open while the workers being laid off cannot cross the skills gap to fill them.
The Klarna story has become the case study in what happens when this logic is applied without understanding the value chain. Between 2022 and 2024, Klarna cut approximately 700 jobs and replaced them with AI-powered customer service systems. The CEO publicly claimed the AI performed at human-equivalent quality. By early 2025, customer satisfaction had deteriorated on complex interactions, internal reviews revealed the AI lacked the nuanced problem-solving the work actually required, and the company began rehiring the human staff it had let go. The rehiring costs — recruiting, onboarding, retraining — exceeded the original savings projections. The true cost of removing the Artists and Architects, it turns out, includes the cost of discovering you cannot do without them.
"This represents a fundamental structural shift rather than a temporary market correction," said one executive coach and former AI industry professional quoted in CNBC's coverage of the 2026 wave. "We're witnessing the beginning of a permanent transformation in how work gets organised and executed."
That transformation is real. Whether it is wise is a different question.
Three Risks Nobody Is Pricing In
When the glory is systematically decoupled from the actual work, three compounding risks follow.
Fragility. A generative system trained on high-quality source data can convincingly mimic a process. It cannot troubleshoot when conditions change, because mimicry is not understanding. The moment the underlying domain shifts — new regulations, new evidence, new edge cases — the system that seemed fluent reveals itself as brittle. The Architect or Artist who understood the original framework's limits is, by then, often no longer in the room. The vibe-coded codebase that nobody can debug, the AI-generated news podcast riddled with errors, the customer service system that couldn't handle complexity — these are not failures of technology. They are failures of provenance.
Erosion of incentive. Serious builders are paying attention to how their work is being used. If the pattern they observe is that precision gets extracted and credit goes elsewhere, the rational response is to stop building openly. MIT Media Lab researcher Robert Mahari, co-lead of the Data Provenance Initiative, put it directly: "There's an ethical dimension to our work — we want to give proper attribution to folks who contribute to AI training models." The open-collaboration norms that powered the best of the digital knowledge economy depend on basic reciprocity. That reciprocity is currently failing at scale.
Misalignment of merit. When career advancement focuses on deployment rather than formulation, organisations optimise for execution over strategic problem-solving. They get faster at producing outputs but slower at producing good questions. Microsoft's 2025 New Future of Work report flagged this directly: automation risks reducing the autonomy, recognition, and connection that workers receive — all of which matter for engagement, performance, and long-term capability. A workforce that has had its Architects and Artists replaced has also had its institutional memory and diagnostic capacity replaced. The system keeps running until it doesn't. Then nobody knows how to fix it.
Where Is This Going?
The honest answer is that nobody knows with certainty — but the data points to something important that the loudest headlines obscure.
The displacement is hitting down the skills ladder more than up. The new jobs being created by AI adoption are disproportionately those requiring seniority and judgment — AI orchestrators, security reviewers, system architects — not the mass-market entry-level roles that are disappearing. PwC's analysis found that wages are rising twice as quickly in industries most exposed to AI compared to those least exposed, for workers who can operate at the level where AI still struggles: ambiguity, ethics, cross-domain synthesis, accountability.
The entry-level pipeline is, however, in genuine crisis. The traditional apprenticeship model — where junior workers did rote tasks in exchange for mentorship and the slow accumulation of judgment — is being automated away. This is a serious structural problem, because those junior workers were not just doing useful work. They were becoming the senior workers of the future. The learning curve is being automated, leaving early-career professionals stranded between AI agents and senior incumbents, with no obvious path from one to the other. The pipeline for future Architects and Artists is narrowing precisely as the demand for their capabilities grows.
The deeper risk is the one the Kleon argument implies: when you remove the people who understand the foundations, you do not just lose their current output. You lose the capacity to build the next generation of foundations. AI trained on AI-generated content degrades. Systems maintained by people who don't understand them become fragile at exactly the moment they face novel conditions. The compounding effect of removing understanding from the value chain is not visible immediately. It takes time to become a crisis. By which point the people who could have prevented it are gone.
The Consequence
A lucrative sub-economy has rapidly emerged around this chaos. Platforms like VibeCodeFixers.com appeared within weeks of vibe coding entering the mainstream — experienced engineers commanding premium emergency rates to untangle brittle, hallucinated codebases that nobody on the internal team understood well enough to fix.
Whether that window stays open depends on how fast the tools improve. What it does not represent is a structural change in how organisations value the Architect. Leadership has not had an epiphany about the value of foundational expertise; they have simply paid for an Architect on invoice when things were actively on fire. The underlying corporate incentive — to skip the understanding, ship the output, and hire back the consultant to clean up the debris — remains entirely intact.
What "Stealing Well" Looks Like Now
None of this is an argument against generative tools. The engineers who are thriving in 2026 are those who have integrated AI deeply into their workflows — using it to execute inside containers they designed with understanding. Senior developers who can evaluate AI output are reporting productivity gains of up to 81%. The leverage is real. The question is whether the organisation has preserved the human judgment required to direct it.
The original lesson of Steal Like an Artist was always that good stealing required mastery. You had to understand the work deeply enough to transform it. The theft was in service of something new. The "thief" had done the apprenticeship.
The equivalent in 2026 is a culture of provenance — one in which organisations and practitioners consistently ask: where did this knowledge come from, who built it, what did it take to make it trustworthy, and how are we acknowledging that?
The creators most directly affected are already saying so. A survey by the Writers' Guild of Great Britain found that 81% of writers believe they should be paid when their work is used to train AI systems. 65% believe AI will reduce their income. 61% fear it will replace their roles entirely. The WGGB has since called for an independent AI regulator and welcomed a March 2026 UK government report on the economic impact of AI on copyright. These are not abstract policy positions. They are the documented concerns of the Artists who built the foundations on which generative AI depends.
The newsrooms laying off journalists while deploying AI to fill the gap are not just making an ethical error. They are degrading the inputs their AI systems depend on. The development teams shipping vibe-coded interfaces without architectural understanding are not just producing fragile code — they are removing from their organisations the diagnostic knowledge required to fix it when it fails. The companies replacing expert human judgment with automated outputs trained on that judgment are not innovating. They are liquidating a knowledge asset and calling it efficiency.
Klarna had to rehire. The Washington Post's AI podcast had to be pulled. The vibe-coded codebase will need a senior engineer to untangle it. The pattern is not unique to any one company or field. It is the predictable consequence of separating the output from the understanding that produced it.
The Work Is the Architecture
There is a version of the current moment that treats the ability to prompt a generative system as the primary creative skill of our era. In that version, the person who hits run is the innovator. The frameworks they ran on are just inputs. The journalist who built the source relationships, the engineer who understood the back end, the strategist who saw the problem before it existed, the composer who knew why the chord mattered — these people are overhead. Their outputs are assets. Their understanding is a cost.
That version is not just incomplete. It is self-defeating. And it is being tested right now, at scale, across every industry simultaneously.
The reporters are gone. The AI podcast has errors nobody can fix because nobody knows what accuracy looks like in that domain. The vibe-coded platform has vulnerabilities nobody can patch because nobody understood what was built. The customer service system couldn't handle complexity and the company had to rehire. The innovation frameworks are running on autopilot and the edge cases are beginning to arrive.
Kleon's "steal" was always about the hard work of apprenticeship and synthesis. It required the thief to master the craft. The theft was in service of something new — something the thief had earned the right to make by understanding what they were taking.
In the age of AI, the theft has been industrialised. The apprenticeship has been skipped. The synthesis has been replaced by ingestion. And the people who built the foundations — who spent years developing the judgment, the taste, the diagnostic capacity that makes any of this work — are being systematically overworked, extracted, and discarded before the consequences are visible.
The consequences are becoming visible.
Crediting the Architect and the Artist is not a moral nicety. It is how you keep the quality of the foundations high enough to be worth building on. It is how you ensure that when the system breaks — and it will break — there is someone in the room who understands why, and knows what to do.
And it is how you stop consuming the very people on whom the next generation of everything depends.