AI Is Eating the World

WHY NOW

People have been predicting AI's arrival since the 1950s. For most of that time, the predictions were embarrassingly wrong. The gap between what AI could do in a research lab and what it could do in the real world was enormous, and the people who confidently announced that machines would be translating languages, driving cars, and writing prose by next Tuesday kept being wrong by a decade, then two, then three.

So the skepticism that still exists today is understandable. It is also, I think, a category error. The relevant question is not whether AI will eventually transform everything — that debate is essentially settled among people who have used the current generation of tools seriously. The relevant question is whether we are in that period right now, in 2026, or whether the transformation is still five or ten years away.

I am going to argue that it is happening now. Not metaphorically. Not "the seeds are being planted." Now.

The reason this moment is different from all the previous moments when AI was going to change everything is the same reason Andreessen's 2011 prediction was different from the dozens of previous "the internet will change everything" predictions that came before it. The technology finally works at scale. Not perfectly. Not without significant limitations. But well enough, and cheaply enough, and accessibly enough, that the economic logic of adoption has fundamentally shifted.

Six decades into the computing revolution, four decades into the personal computer era, and roughly fifteen years into the modern deep learning era, the technology required to automate large categories of cognitive work has matured to the point where the question is no longer "can AI do this?" The question is "how quickly will AI do this, and what does that mean for everyone currently doing it?"

The answer to the first part of that question is: faster than almost anyone predicted, even the optimists.

THE NUMBERS ARE NOT ABSTRACT

Let me ground this in something concrete, because this is the kind of argument that can stay frustratingly vague if you let it.

As of early 2026, AI tools are generating roughly 41 to 46 percent of all code written by active software developers. Nearly half. That is not a rounding error or a measurement artifact — it is a structural shift in how software gets made. GitHub Copilot alone has crossed 20 million users, growing 400 percent year-over-year between early 2024 and early 2025. Ninety percent of Fortune 100 companies have deployed it. Developers using it complete tasks 55 percent faster in controlled studies. Pull request review time has dropped from 9.6 days to 2.4 days.

At Google, 25 percent of code is now AI-assisted, and the company's engineering velocity has increased roughly 10 percent as a result. Microsoft's CEO Satya Nadella announced that GitHub Copilot has already grown into a larger business than all of GitHub was at the time of its $7.5 billion acquisition in 2018.

Pause on that for a moment. A product that did not exist three years ago is now worth more than a platform that took twenty years to build. That is not a normal technology adoption curve.

The AI coding tools market hit $7.37 billion in 2025 and is projected to reach $30 billion by 2032. But the revenue figures miss the point. The point is that software, which itself was eating the world, is now being written at a pace and a cost that was structurally impossible before. When you can produce code at half the cost and twice the speed, the economics of what gets built — and what does not — change completely. Software that could not previously justify the engineering investment to build it becomes viable. Automation that was too expensive to implement becomes cheap. The barrier to shipping a product drops from "can we afford a team of engineers" to "can we afford a laptop and a subscription."

This is not just about developer productivity. It is about what becomes possible when the cost of creating software approaches zero.

THE NICHES ARE FALLING, ONE BY ONE

Andreessen catalogued the industries software was eating in 2011: books, music, retail, telecommunications, financial services. He was right about all of them, though the timelines stretched longer than he implied.

The AI invasion is broader and faster. It is not moving industry by industry, politely waiting for each sector to notice and respond. It is moving through every knowledge-intensive field simultaneously, targeting the most profitable and most automatable tasks first.

Legal work. The lower rungs of legal practice — document review, contract analysis, legal research, due diligence — are the most obviously exposed. These tasks are expensive, time-consuming, human-intensive, and almost perfectly suited to what large language models do well: reading enormous amounts of text, extracting relevant patterns, and summarizing clearly. Junior associates at major law firms are already seeing their billable hours shrink. Paralegal job postings have declined meaningfully. The entry-level positions that used to serve as the first step on the legal career ladder are contracting. Firms that have adopted AI for document review are reporting cost reductions of 60 to 80 percent on those specific tasks. The partners at the top of those firms are not complaining.

Financial services. Goldman Sachs — which, it bears noting, has an obvious interest in both the technology and in being honest about its effects — has been straightforwardly public about automating legal, compliance, and trading operations. Investment banking tasks that once required a team of junior analysts grinding through financial models are being compressed. Basic pitchbook generation, financial modeling, and market research synthesis are increasingly AI-assisted. The first-year analyst class at bulge-bracket banks is getting smaller. The work is not disappearing; it is being absorbed higher up the organizational chart, by more senior people with AI tools, rather than being distributed down to a class of juniors.

Healthcare. Medical coding is already substantially automated. Radiology screening — the analysis of images for tumors, fractures, and anomalies — is one of the areas where AI has demonstrated performance competitive with or exceeding specialist radiologists on specific narrow tasks. Administrative roles throughout the healthcare system are contracting as AI handles scheduling, documentation, and triage. Medical transcriptionists, whose employment the Bureau of Labor Statistics projects to decline 4.7 percent through 2033, are the canary in this particular coal mine: a profession that consists almost entirely of converting spoken clinical language into structured text, which is almost precisely what AI does well.

Customer service. The call center is being hollowed out. Not in a slow, polite way — in the way that digital photography hollowed out film processing, which is to say quickly once the economics tipped. Customer service representative employment is projected to decline 5 percent through 2033. That BLS projection was almost certainly made with conservative assumptions about the pace of AI deployment; the actual number is probably larger and faster.

Content creation. Marketing copy, product descriptions, first-draft reports, data journalism, social media content — this is where AI has been most visible to ordinary people, and where the disruption is most mature. Junior copywriters and content producers are facing a market where a single experienced writer with AI tools can produce the volume that previously required a team. This is not a distant threat. It is already the hiring reality at companies that have been honest with themselves about it.

Software development itself. There is a particular irony in the fact that the most technical profession — the one that created all the tools doing the disrupting — is itself one of the most exposed. Junior software developers are seeing the same hiring headwinds that junior lawyers and junior analysts are seeing. Goldman Sachs research has documented that unemployment among young tech workers has risen roughly 3 percentage points since early 2025. Entry-level software engineer roles, which used to be the most reliable on-ramp into the professional class for technically capable young people, are under real pressure.

The pattern across all of these sectors is the same. AI is eating from the bottom up. It is targeting the work that is most structured, most repetitive, most document-heavy, and most expensive relative to the value it creates. Junior roles — the ones that have historically served as the training ground for building senior judgment — are the first to contract. The ladder that previous generations used to climb from entry-level to expert is losing its lower rungs.

This is one of the more underappreciated consequences of the current moment, and I will return to it.

THE SOFTWARE AMPLIFICATION EFFECT

Here is something that should concern anyone who thinks the AI disruption will be limited to specific industries or specific role types.

When the cost of building software drops by half or more — when a developer with AI tools is doing the work that previously required two or three developers — the economic logic of what gets automated changes entirely. Projects that did not previously clear the bar for return on investment now do. Industries that could not afford custom software solutions now can. Automation that required a six-figure engineering budget now requires a four-figure subscription.

This is the amplification effect. AI does not just do what software used to do more cheaply. It expands the frontier of what software can economically do at all.

Consider a small regional law firm that could never afford to build custom document analysis software. With current AI tools, a paralegal or a solo practitioner can build a workflow that reads and summarizes contracts, flags unusual clauses, and extracts key terms — in an afternoon, for almost nothing. Consider a regional hospital system that could not afford a team of data scientists to analyze patient outcomes. With AI tools, a physician who can write plain English descriptions of what they want to understand can get meaningful analysis from their own data without a data science team.

This is what Andreessen was pointing at in 2011 when he talked about software democratizing access to capabilities that had previously required large capital investments. But AI is doing it faster, more broadly, and with less technical skill required from the person deploying it.

The implication is that the economic disruption is not limited to the industries that can afford to hire AI engineers. The disruption extends to everyone who can afford a subscription — which, at current AI pricing, is almost anyone.

THE GREAT MISCONCEPTION

There is a narrative circulating in certain professional circles that goes roughly like this: AI is a tool that makes workers more productive, not a technology that replaces them. AI does the routine stuff so that humans can focus on the high-value, judgment-intensive work. AI is a partner, not a competitor.

This narrative is comforting. It is also partially true. And the part that is true makes the part that is false more dangerous, because it lulls people into a false sense of security.

The data does not support the full version of the partnership narrative. When AI makes workers 55 percent more productive, that does not automatically mean companies hire 55 percent more workers to do the same things at a larger scale. Sometimes that happens. More often, it means companies do roughly the same amount of work with fewer people, or significantly more work with the same number of people. The Andreessen prediction that software would eat certain industries did not mean that those industries would hire more people to do software-adjacent things. It meant the industries shrank and the software companies grew.

The World Economic Forum's 2025 Future of Jobs Report is instructive here. It projects 170 million new jobs created globally by 2030 and 92 million displaced — a net positive. The economics profession's consensus is similar: AI will create more jobs than it destroys, in aggregate, over the medium term. This is probably true, in the same sense that it was true that the automotive industry created more jobs than it destroyed — eventually, in aggregate, for the economy as a whole. But that aggregate truth was cold comfort to the individual blacksmith or carriage maker whose specific skill became economically obsolete during the transition.

The net positive number also masks the transition cost. The 92 million displaced jobs represent real people with real mortgages and real kids and real communities built around industries that are contracting. The 170 million created jobs are in fields that may require years of retraining to enter, and they may be geographically, educationally, or economically inaccessible to the people who need them most.

Dario Amodei, the CEO of Anthropic — a company that is directly building the technology in question — said in 2025 that AI could eliminate roughly 50 percent of white-collar entry-level positions within five years. That is a striking prediction from someone with a financial interest in AI's success and a professional interest in being taken seriously. He did not say it to scare people. He said it because the evidence supports it.

The workers who are most exposed to this transition are not who most people expect. They are not, primarily, the workers doing physically repetitive blue-collar tasks. They are highly educated, experienced professionals in knowledge-intensive fields. The Anthropic labor market analysis published in March 2026 found that workers earning above-average wages, with advanced degrees, in fields requiring substantial cognitive effort, are the ones most exposed to current AI capabilities. This inverts the conventional assumption that automation hits the bottom of the labor market first and works its way up.

It inverts it because the current generation of AI is not, primarily, an automation tool for physical tasks. It is an automation tool for cognitive tasks — specifically for the kind of structured cognitive work that involves reading, writing, analyzing, summarizing, classifying, and reasoning over text and data. That is the cognitive work that highly educated professionals do.

THE LADDER PROBLEM

Let me dwell on one consequence that I think is getting insufficient attention.

The way that expertise develops in most knowledge-intensive fields is through a hierarchical apprenticeship model. You start at the bottom doing the least glamorous, most structured work. You make mistakes, you get corrected, you build intuition. Over years, you take on more complexity and judgment. Eventually you develop the kind of hard-won expertise that makes you genuinely irreplaceable.

This model is breaking. Not everywhere, not all at once. But the early evidence is clear.

Junior software engineers are being hired less because AI can write basic code. Junior lawyers are facing a squeezed market because AI can review documents and research cases. Junior financial analysts are seeing their workload compress because AI can build models and summarize research. Junior medical coders are being replaced by automated systems. Junior copywriters are finding that the entry-level content work that used to sustain a career while you developed taste and judgment is mostly gone.

If the rungs at the bottom of the ladder disappear, the people who would have used those rungs to develop expertise never develop it. In ten years, we may find ourselves with a shortage of senior talent in fields where the pipeline of junior trainees dried up in the mid-2020s. The companies and institutions that are cutting junior headcount today may find, a decade from now, that they have no one to promote — because the generation of workers who would have developed into senior experts by 2035 never got the foundational experience.

This is not hypothetical. It is the predictable result of optimizing for near-term efficiency at the expense of long-term talent development. And most organizations are not thinking about it.

THE NEW GEOGRAPHY OF VALUE

If AI is doing the routine cognitive work, what is left for humans?

The honest answer is: more than people fear, but less than the most optimistic narratives suggest, and the transition will be genuinely hard for many people.

The work that remains distinctly human is characterized by a few qualities. It requires real-world judgment in ambiguous, high-stakes situations where the cost of error is severe. It requires trust-based relationships that cannot be mediated by a machine. It requires physical presence. It requires creativity that draws on lived experience in ways that current AI systems cannot replicate. It requires the integration of information across extremely long time horizons in ways that current context windows do not support.

A surgeon performs a procedure that requires fine motor control, real-time adaptation to unexpected anatomy, and accountability that can only be borne by a licensed human professional. A therapist builds a relationship over years with a patient, reading emotional subtext and navigating crises in ways that require genuine human connection. A skilled trades worker troubleshoots a mechanical failure on-site in conditions that no training dataset has specifically anticipated. A corporate litigator makes judgment calls in a deposition based on decades of reading witnesses and juries. A founder builds a company by making hundreds of calls per week that require synthesizing interpersonal, market, and technical information in real time.

These are not activities that AI is likely to automate soon. They require the kind of situated, embodied, high-accountability judgment that language models are not yet close to replicating reliably.

The workers who will do best in the near-term transition are those who understand both domains — who can direct AI effectively because they understand what they are asking for, and who provide the human judgment that validates and deploys the AI's output. A lawyer who can use AI to do in two hours the document review that used to take a team two weeks is not replaced by AI. She is ten times as productive. The same is true for the engineer who can specify and review AI-generated code. The financial analyst who can frame the right questions and interpret AI-generated models. The doctor who can use AI diagnostics as one input among many, integrated with clinical judgment.

PwC's 2025 AI Jobs Barometer found that workers with demonstrable AI skills earned an average of 25 percent more than peers without them in equivalent roles. That premium will probably compress over time as AI proficiency becomes baseline. But for now, it is real, and it is significant, and it rewards people who invested early in understanding the tools.

WHAT THE ENTREPRENEURS ARE SEEING

One of the most reliable leading indicators of a technology wave is what entrepreneurs are building. In 2010, if you paid attention to what the most capable founders were working on, you would have seen the software-eating-the-world thesis playing out before Andreessen wrote it.

Pay attention to what capable founders are building in 2026, and the picture is striking.

The most interesting companies are not building AI as a feature. They are building businesses that are only possible because AI exists — businesses where the product is fundamentally AI-powered in a way that would have been impossible or economically unviable two years ago. AI-native legal research platforms that let a single attorney do the research of a small team. AI-native financial analysis tools that let a solo portfolio manager do the work of a research department. AI-native customer service systems that handle 80 percent of inquiries without human involvement. AI-native content operations where a two-person team produces the output of a twenty-person editorial department.

These businesses are not incrementally better than their predecessors. They operate with fundamentally different unit economics. Their cost structures are structurally lower, their margins are higher, and their scalability is greater. And they are competing with companies that built their cost structures around human labor assumptions that no longer hold.

This is what "eating" looks like from the inside. Not a dramatic announcement. Not a moment of defeat. Just a slow erosion of competitive position as the new entrants build on better economics and the incumbents defend positions that are becoming less and less tenable.

The venture capital numbers reflect this. AI investment in the United States reached $285.9 billion in 2025. Microsoft's investments in AI are already returning an average of 3.5x, with some deployments hitting 8x. The capital is not chasing a bubble. It is chasing the same thing it always chases: a structural shift in what is economically possible.

THE LONG CONSEQUENCES

Let me be direct about the long-term consequences, because I think the discourse around AI tends to either catastrophize or wave away things that deserve serious attention.

The concentration problem. The software wave created enormous wealth, but it also concentrated it in ways that were not uniformly good. A handful of platforms captured network effects and became monopolies or near-monopolies. The returns accrued disproportionately to the people who owned equity in the winning companies, not to the people who used the products or the workers whose industries were disrupted.

AI has the potential to be far more concentrated than software. The frontier models — the ones actually capable of performing the tasks that matter — require billions of dollars of compute to train, and are controlled by a handful of companies. If those companies capture the value of AI broadly, the economic gains from AI will accrue to a very small number of shareholders. The people and places whose labor AI replaces will not automatically receive a share of those gains.

This is not a certainty. It is a risk that depends significantly on competitive dynamics, regulatory choices, and decisions made by the companies themselves. But it is a risk worth naming clearly.

The epistemic problem. When AI is writing nearly half of all code, and an increasing fraction of all written content, and an increasing share of legal documents, financial analysis, and scientific research summaries, it becomes genuinely difficult to know where human judgment ends and machine output begins. This is not a problem of malicious deception. It is a structural feature of a world in which AI is deeply embedded in knowledge production.

If AI systems have systematic biases — and they do — those biases propagate through the outputs at a scale that manual review cannot catch. If AI systems make systematic errors in domains where humans have stopped developing independent expertise — and that is the direction we are heading — those errors can go unchallenged because no one has the baseline competence to challenge them.

The epistemological consequences of a world in which most structured cognitive work is AI-mediated are genuinely hard to think through, and I do not think we have thought about them seriously enough.

The education problem. The skills that the AI economy rewards are not the skills that most educational institutions currently teach. Data literacy, AI proficiency, and the capacity to frame problems and evaluate outputs are going to be fundamental competencies across essentially every professional field — not just in technology. Educational systems that are not actively building these competencies into their curricula are producing graduates for a labor market that will have meaningfully changed by the time those graduates enter it.

The entry-level problem. I mentioned this earlier, but it deserves to be stated plainly: if AI systematically eliminates entry-level work, and entry-level work is how expertise develops, then we are running a slow-motion experiment in whether it is possible to sustain senior-level capability without junior-level apprenticeship. The experiment will take ten to fifteen years to return results. The organizations cutting junior headcount today may not notice the consequences until it is too late to address them.

THE CHOICE THAT IS ACTUALLY BEING MADE

Here is the thing about technological waves: they do not ask for your permission. They do not wait for everyone to agree that they are real. They do not slow down because the disruption is inconvenient or the transition is painful.

Andreessen wrote in 2011 that incumbent companies faced a choice: transform or be disrupted. Most of them chose to believe that transformation was optional, or that they had more time than they did, or that the disruption would stop at the edge of their particular sector. Some of those companies no longer exist. Others spent the next decade scrambling to catch up.

The same choice is on the table today, and it is being made by individuals and organizations whether they think they are making it or not.

The worker who does not develop AI proficiency is choosing to compete at a structural disadvantage against workers who have. The company that does not embed AI into its core operations is choosing to compete against companies with fundamentally better unit economics. The educational institution that does not update its curriculum to reflect how knowledge work is actually done is choosing to produce graduates who are less prepared for the world they will enter. The policymaker who does not engage seriously with the labor market consequences of AI is choosing to leave those consequences unaddressed until they become crises.

None of these choices are inevitable. They are choices. But they are being made passively, by inaction and delay, and that form of decision-making tends to produce worse outcomes than conscious engagement.

WHAT ANDREESSEN GOT RIGHT, AND WHAT THE SEQUEL LOOKS LIKE

In 2011, Andreessen made a few predictions that turned out to be precisely right.

He predicted that software companies would invade non-software industries and that the incumbents would be slow to respond. He predicted that the disruption would be broader and faster than most people expected. He predicted that the companies doing the disrupting would look, at first, like overvalued novelties to people who were not paying close attention.

He also made one prediction that was partially wrong, or at least incomplete. He argued that the software wave would ultimately create more jobs and more economic growth than it destroyed. He was right at the aggregate level, over the long term. He underweighted how painful the distributional consequences would be — how the gains would concentrate and the losses would disperse.

The AI wave is going to be right in the same ways, and wrong in the same ways, but at greater speed and scale.

AI is going to eat the work that software made possible. It is going to eat law, finance, medicine, education, content, and research in the same way that software ate retail, music, and hospitality. It is going to create enormous value. It is going to create new industries that we cannot currently name or predict, the way that software created the gig economy, the creator economy, and the platform economy — categories that did not exist before the tools that enabled them.

And the gains are going to be concentrated in the hands of the people who own the models, who can direct the models, and who provide the human judgment that gives AI output its value. The losses are going to fall on the people whose labor was most directly replaceable by the current generation of tools — people who are, on average, younger, earlier in their careers, and often in fields that are not currently thinking of themselves as AI-exposed.

The people who see it clearly — who recognize both the opportunity and the disruption, who develop the skills to direct AI rather than compete against it, who build the companies and the careers and the institutions suited to the world that is being built rather than the world that existed — are going to have an extraordinary decade.

The people who look up in five years and discover that the world reorganized itself while they were busy defending the old one are going to have a very different experience.

In 2011, the right response to Andreessen's essay was not panic. It was not denial. It was clear-eyed attention to what was actually happening and an honest assessment of what it meant for your industry, your company, and your career.

The right response to AI eating the world is the same thing.

Pay attention. The meal is already underway.

Token Bureau covers the business and consequences of artificial intelligence. Published April 2026 · The Token Review · thetokenreview.ai

A NOTE ON SOURCES

The statistics in this piece draw on data from GitHub and Microsoft's public disclosures, Goldman Sachs Research, McKinsey Global Institute, the World Economic Forum's 2025 Future of Jobs Report, PwC's 2025 AI Jobs Barometer, the U.S. Bureau of Labor Statistics, and Anthropic's March 2026 Labor Market Impacts Report. Where ranges are cited, they reflect the spread across credible sources rather than a single point estimate. The labor market projections should be read as directional rather than precise — the honest truth is that the pace of this transition is genuinely uncertain, and anyone claiming precision is overstating their confidence.