On November 11, 2001, in a packed arena at the MGM Grand about fifteen minutes from where I work, Bill Gates held up a Tablet PC prototype and told the Comdex keynote audience that within five years it would be the most popular form of PC sold in America.
He was right about the tablet. He was wrong by nine years and by one company, and the reason he was wrong is worth sitting with. Microsoft’s tablet ran full Windows XP, cost two thousand dollars and up, and took input from a stylus, because Microsoft had decided people would want to do exactly the same things on a tablet that they did on a desktop. Apple shipped a device with none of those properties and took the market. The hardware was not the problem. The operating system was not the problem. The problem was what the machine was pointed at.
That is the most useful thing to know about Bill Gates before reading his new essay on artificial intelligence and employment, because he has just made the same category of error again, and this time the correction was sitting three pages deeper in the paper he cited.
The pattern
Start with The Road Ahead. Gates published it in November 1995, at the moment the commercial internet was catching fire, and the World Wide Web appears on four pages out of two hundred and eighty-six. The internet, he wrote, was one of the important precursors of the information highway, but not the highway itself. Microsoft was betting on MSN to be the onramp. He was not wrong that a global network would reorganize the economy. He was wrong about which network, and about who would control the entry point. He knew it fast, too. He rewrote the book for the 1996 paperback, adding something like twenty thousand words about the internet, which is more intellectual honesty than most people in his position manage.
Then Davos, January 24, 2004. Gates told a room at the World Economic Forum that two years from now, spam would be solved. He was right that spam was solvable. He was wrong about the mechanism, and it is the wrongness that matters. His preferred fix was economic: charge senders a small fee, a digital postage stamp, with trusted contacts exempt. In the long run, he said, the monetary method would be dominant. By 2007 spam was running around ninety percent of all email and no postage system existed anywhere.
Spam did eventually get handled, more or less. Not by a payment layer. By Bayesian filters, by reputation scoring, and by three boring authentication protocols that a few thousand engineers argued about on mailing lists and then shipped: SPF, DKIM, DMARC. Nobody priced anything. The problem yielded to plumbing.
Silverlight is the same shape. Microsoft launched it in 2007 on a correct premise, which was that the web of the late 2000s could not deliver rich media and real application interfaces without help. Correct diagnosis. The prescription was a proprietary browser plugin, and the open web ate it. Netflix moved to HTML5 in 2013. Chrome dropped plugin support in 2015. Microsoft ended Silverlight support on October 12, 2021, and a great many enterprises are still paying to migrate line-of-business applications off it. I have billed hours on that migration. It is not an abstraction to me.
And then the big one, which Gates volunteered himself. At a Village Global event in 2019 he said the greatest mistake ever was whatever mismanagement he engaged in that caused Microsoft not to be what Android is, and put the number at four hundred billion dollars transferred from company G to company M. Microsoft had a mobile operating system before Google did. It had the developer relationships, the enterprise channel, and more money than anyone. It lost on how the platform was licensed and how the app ecosystem was seeded. Not on capability.
Five predictions across twenty-five years. In every one of them Gates identified the important technology correctly and then got the deployment question wrong. Not the what. The how.
Which brings us to August
On August 26, 2026, Gates published six thousand words arguing that AI will disrupt employment faster than any prior technology and that no one has a plan. The evidentiary spine of the jobs section is a single sentence: white-collar jobs are already being hit modestly, and after the widespread adoption of generative AI, employment fell significantly among young workers in jobs especially vulnerable to replacement but not among their older colleagues.
That is a compressed statement of one finding from “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen. Gates cites the November 13, 2025 revision. The first version came out on August 26, 2025, a year earlier to the day.
It is Fact One. There are six. He used the weakest of them.
Why Fact One will not hold the weight he put on it
Fact One compares AI-exposed occupations against less-exposed ones and finds a sixteen percent relative employment decline for workers aged twenty-two to twenty-five after controlling for firm-time effects. The number is real. What it means is another matter, because the entire comparison is normalized to October 2022.
Anyone who was hiring engineers in 2022 knows what that month was. It was the crest of the post-pandemic white-collar hiring wave: zero rates, record venture deployment, remote-work land grab, and a level of over-hiring that nearly every large technology company subsequently admitted to in public. And hiring waves are junior-heavy, because when you are adding headcount fast you add it at the bottom of the pyramid. So the cell the paper measures — twenty-two to twenty-five year olds, in white-collar remote-capable roles, in the sectors that over-hired hardest — is the most inflated group in the dataset, measured downward from its own peak. Plain reversion produces the headline result with no artificial intelligence in the model at all.
Three more shocks land in the same twelve months, and I lived through all of them as an employer. Section 174 amortization took effect for tax years beginning after December 31, 2021, which meant every American developer I hired came with a deduction spread over five years while the compute I could have bought instead was deductible the day I paid for it. Nobody fixed that until OBBBA restored domestic expensing for tax years after December 31, 2024. The Federal Reserve started the fastest tightening cycle in forty years in March 2022. Freshman enrollment had fallen nearly ten percent in fall 2020, and that class graduates in spring 2024, landing in the twenty-two to twenty-five band during the sharpest decline the paper records. Four confounds, one treatment, and the treatment arrived last. Firm-time fixed effects catch none of them, because each hit AI-exposed occupations harder inside the same firm.
The authors are more candid about this than their press coverage. When they extend the sample back to 2018 to find a clean pre-period, they note that the most exposed quintile under their primary measure was already growing slower starting around 2020, two years before ChatGPT. Their remote-work check runs into the same wall: they state outright that teleworkability correlates with AI exposure, with only two teleworkable occupations in the bottom quintile, which forced them to pool quintiles to run the split. You cannot control for a confounder that is nearly identical to your treatment.
There is also a second administrative dataset that says the opposite. Anders Humlum and Emilie Vestergaard linked adoption surveys covering twenty-five thousand Danish workers across seven thousand workplaces and eleven occupations to Denmark’s labor registers and found precise nulls on earnings and hours, ruling out effects larger than two percent two years in, with the nulls holding for early-career workers specifically. Denmark had no Section 174 and no American-scale hiring bubble. Two countries, two confound structures, opposite answers.
Fact Three is the one he needed
Three pages later the same authors do something much smarter. Instead of comparing exposed occupations to unexposed ones, they split the exposed occupations against each other, using the Anthropic Economic Index to separate the ones where AI substitutes for the work from the ones where AI collaborates with it. Automation on one side, augmentation on the other.
Entry-level employment fell where AI automates. It did not fall where AI augments. The highest augmentation quintile posted among the fastest employment growth of any group in the study.
Look at what that split does to everything I just laid out. Both arms are AI-exposed. Both got Section 174, the rate hikes, the bubble correction, the smaller graduating classes. Every confound is common to both sides and cancels. Fact Three is far better identified than Fact One, and it carries a claim Fact One cannot support: what happens to the job is not determined by how exposed it is to the technology. It is determined by what the technology is pointed at.
Which is the same sentence I wrote about the Tablet PC.
The industrial revolution: the number he should have led with
Gates argues this time is different because previous shifts unfolded across generations. Before accepting that, look at what actually happened to employment inside the shift he is invoking.
James Bessen went through the nineteenth-century cotton data and found that power looms automated ninety-eight percent of the labor required to weave a yard of cloth. Ninety-eight percent. And the number of factory weaving jobs went up over that period, not down. The mechanism was demand. Less labor per yard meant a lower price, a lower price meant sharply more cloth sold, and more cloth sold meant more weavers even though each weaver was needed for a fraction of the work. The remaining skills got more valuable too, because coordinating output across multiple looms was harder than tending one. Weavers’ wages rose sharply against other workers in the late nineteenth century.
That is the direct answer to the argument Gates is making. He is saying the models can do the cognitive work, therefore the jobs go. Power looms could do ninety-eight percent of the weaving and the weaving jobs multiplied.
The economy-wide picture is the same shape. American agriculture employed roughly eleven million people at the turn of the twentieth century, about forty percent of the workforce, and employs around two million today, under two percent. Across that same century total employment went from something like twenty-nine million to roughly a hundred and sixty million. Agriculture shed nine million jobs. The economy added a hundred and thirty-one million.
Now the honest half, because it cuts both ways and Gates deserves it. The handloom weavers, the ones outside the factories, were annihilated. Britain had something like a quarter of a million of them in the 1820s and by the 1860s the occupation was gone, and those men did not become mill hands. Real wages across the British economy were roughly flat from the 1790s through the 1840s while output climbed, the stretch historians call the Engels pause. Half a century of extraordinary aggregate growth that the people inside it could not eat.
And Bessen found the ending too, which is the part that should make everyone slow down. Employment in these industries grew alongside productivity for a century and then reversed when demand saturated. US broadwoven textiles employed over three hundred thousand production workers in 1958 and sixteen thousand by 2011. Primary steel went from over five hundred thousand to about a hundred thousand. The growth was real, it lasted a hundred years, and it was not permanent. It lasted exactly as long as demand stayed elastic.
The internet: the cleanest comparison available
This one is close enough in time that the statistics are good and most of us were there for it.
In 1994 the United States had about 1.4 million computer specialists. By the end of the decade that number had more than doubled, to over 2.9 million. In six years. Those were not jobs that moved from somewhere else, they were net new occupational demand created by the build-out, and the occupational categories themselves kept multiplying: the Census Bureau tracked three IT occupation categories in 1970 and twelve by 2010, with total IT employment going from four hundred and fifty thousand to 4.6 million by 2014.
The whole economy tracked it. Nonfarm payrolls stood at roughly 112 million in January 1994 and peaked near 132.7 million in February 2001. Twenty million jobs in seven years, added the entire time the trade press was explaining that disintermediation would gut retail, travel, publishing and banking.
Some of it was a bubble and the correction was brutal. The bust took about 2.7 million jobs back and payrolls did not regain the prior peak until 2005. Travel agents genuinely were destroyed. So were typesetters and compositors. But web developer, SEO analyst, cloud architect, DevOps engineer and site reliability engineer did not exist as job titles in 1994 and now employ millions of people between them, and nobody in 1994 could have written them down, which is the specific reason projections of net job loss from a new general-purpose technology have such a poor record.
Gates writes that this time there will be some new jobs but, without the right policies, far fewer than exist today. That is not a result from the paper he cites or from any paper I can find. It is the assumption both prior transitions falsified, presented as a finding.
The study nobody quotes, which is the closest match we have
James Feigenbaum and Daniel Gross wrote a paper for the National Bureau of Economic Research called “Automation and the Fate of Young Workers.” It studies the mechanization of American telephone exchanges. Between 1920 and 1940, exchanges serving more than half the United States network were converted from human switchboard operators to mechanical switching, city by city, on a staggered schedule. That staggering gives the study the one thing the Stanford paper does not have, which is clean identification.
After a city cut over, employment of sixteen to twenty-five year old women as telephone operators fell by fifty to eighty percent. Accounting for turnover, the authors estimate the shock may have closed off entry-level opportunity for as much as ten to fifteen percent of peak cohorts. Incumbent operators were wrecked. A decade later they were less likely to hold the same job, less likely to be working at all, and when working, more likely to be in something that paid less.
Then the part that changes the argument. Automation did not reduce employment rates in the cohorts of young women who came after. They found work in other sectors at comparable wages, including as typists and secretaries.
So the damage is real and it is severe, and it lands on the people already standing in the doorway rather than on the pipeline behind them. It is a cohort injury, not a permanent amputation of the entry rung. Gates’s framing, in which the bottom rung breaks and stays broken unless governments intervene, is specifically the outcome that the best-identified historical study of this exact phenomenon does not find.
The ATM story, including the ending
Everybody cites this one and almost nobody finishes it. James Bessen went and assembled the numbers. The United States went from about a hundred thousand ATMs around 1990 to roughly four hundred thousand, and full-time bank teller employment rose across that span, from something like five hundred thousand toward five hundred and fifty thousand. The mechanism was mundane. ATMs cut the tellers needed to run an average urban branch from about twenty to about thirteen between 1988 and 2004, which made branches cheap to open, so banks opened forty-three percent more of them in urban areas. The machine made the human cheaper to deploy and the industry deployed more humans.
Here is the ending. Teller employment started falling hard after 2010 and never stopped. The occupation held roughly 347,000 jobs in 2024 and the Bureau of Labor Statistics projects another thirteen percent decline over the following decade. That was not a delayed ATM effect. ATMs had saturated years earlier. It was the smartphone. Mobile banking made the branch visit optional, and the branch was precisely the thing the ATM had made cheap.
One occupation, two technologies, opposite outcomes, and nothing about the tellers changed in between. The ATM augmented the branch. The phone automated it. That is Fact Three with forty years of data on it.
Where this decision actually gets made
The automation and augmentation distinction is not philosophy. It is a configuration setting, and it gets made on a Tuesday afternoon by whoever is writing the tool schema. On the wedding planning agent I run for a Las Vegas wedding chapel, the same model can either draft a response that a coordinator reviews and sends, or write straight to the plan and confirm the line item itself. Same model, same context window, same MCP server. The difference is whether the tool call proposes or commits. The first version makes the coordinator faster and lets the business take more bookings per coordinator. The second takes the coordinator out of the loop. Nobody from a labor department was in the room. No international AI framework would have been either. I made the call, and I make some version of it most weeks, and so does every other person shipping agents right now.
That is the finding Gates walked past. Not because he is careless. Because the deployment layer is the layer he has consistently failed to see, in 1995, in 2001, in 2004, in 2007, and in mobile.
And the technology cannot do the thing the essay assumes it can
There is a premise holding up the entire jobs argument that never gets examined. Gates writes that the reliability problem is being fixed quickly, as researchers create models that can check their own work and improve themselves, and that soon AI will provide nearly error-free work, function on its own without a human checking in, and give companies every incentive to let it. Every displacement scenario in six thousand words rests on that sentence. It is also the only significant claim in the essay supported by no evidence whatsoever, and it is the one I can tell you from production is wrong.
Start with the horizon. METR timed human domain experts on multi-step software and research tasks and handed the same tasks to frontier models. On work that takes a skilled human under roughly four minutes, models score near a hundred percent. On work that takes a human more than roughly four hours, they succeed less than ten percent of the time. The task length a model can complete at fifty percent reliability has been doubling about every seven months, which is genuinely fast, and it is also a curve that is nowhere near the length of a working day.
A job is not a four-minute task. A job is a long sequence of tasks with state carried between them, and in a sequence errors compound multiplicatively instead of averaging out. An agent that is eighty-five percent reliable per step is twenty-seven percent reliable across eight steps. Sierra’s tau-bench measured this directly on retail service work: roughly sixty-one percent success on a single run, roughly twenty-five percent when the same task has to succeed eight consecutive times. A worker who blows one shift in four is not a worker. He is an incident.
Duration makes it worse rather than better. Studies of long-context agents report success rates of forty to fifty percent on the short version of a task collapsing below ten percent once the identical task is embedded in a longer interaction history, with the relevant information still sitting inside the context window. The model has the information in front of it and cannot use it. That is not a capability that scales up into a role.
Then the structural piece, which is the one that gets the least attention. Replacement requires unsupervised operation. Unsupervised operation requires the system to know when it is wrong. And self-verification is exactly the thing that does not work. Huang and colleagues showed at ICLR that models struggle to correct their own reasoning without external feedback and sometimes degrade while trying. Kamoi and colleagues surveyed the entire literature for TACL and concluded that no prior work demonstrates successful self-correction from a model’s own feedback alone; it works when there is reliable external feedback and not otherwise. There is even a documented self-correction blind spot, where a model fails to fix an error in its own output and fixes the identical error when it is handed to it as someone else’s.
That is not a temporary gap waiting on the next release. It is structural. The same distribution that produced the error is scoring the correction. The way you break the loop is by adding an external verifier, which is a test suite, a type checker, a schema, a compiler, or a person. That is what every one of us builds, all day, and it is why the reliability path runs straight through the humans rather than around them. Put a model under a scoring function without a real verifier behind it and you get reward hacking instead, which METR now reports seeing with increasing clarity as systems cheat their way to impossible scores on evaluation tasks. That is what checking its own work actually looks like.
The enterprise numbers follow from all of this. Deloitte’s 2026 Tech Trends puts the agent pilot-to-production failure rate at eighty-nine percent. Forrester and Anaconda put it at eighty-eight. And the detail that matters most for Gates’s argument is what practitioners name as the blocker: evaluation gaps, governance, and scoping consistently outrank raw model capability. Almost nobody’s pilot died because the model was not smart enough. They died because nobody could establish that the output was correct, and establishing that the output is correct is a person.
I am not going to make the absolute version of this claim, because the absolute version is false. Some work really has been eliminated. Bulk transcription is gone. Routine translation is largely gone. Tier-one chat deflection is real, and it is the most consistently realized return in the entire category. Notice what those have in common. Each was already a single short task repeated, which is precisely the regime where the models are reliable.
That is the boundary, and it is sharp. These systems are excellent at minutes and disintegrate at hours. They are excellent where the answer is checkable and unreliable where knowing you are right is the hard part. A job is a long-horizon, stateful sequence performed by somebody who can be held accountable for the result, and nothing on the current curve does that. Gates’s decade timeline for law, medicine, software and manufacturing assumes a four-hour success rate that does not exist, and assumes it is arriving because models will begin checking their own work, which is the one capability the research says they do not have.
Which loops the whole argument closed. This is why Fact Three shows what it shows. Augmentation works because it runs the model at the scale where it is reliable and leaves a person on the verification step. Automation attempts fail because they run it at the scale where it is not and have nothing underneath to catch the errors. The deployment question and the capability question turn out to be the same question, and the answer to both is the person you were trying to remove.
The postage stamp, again
His three remedies are a new domestic and international institutional framework, a category of work called Human Reserved that is set aside for people by rule, and a tax on AI tokens and robots.
In 2004 Bill Gates proposed solving spam by charging senders a fee per message. In 2026 he proposes slowing labor substitution by charging a fee per token. Twenty-two years apart, the same instinct: when a technology is producing an externality, price it. Spam was not solved by pricing. It was solved by classifiers and authentication protocols built and deployed by engineers, and the postage stamp never shipped anywhere.
The token tax has a second problem, which is that its stated mechanism is wrong. Gates argues that if you buy a robot you write it off immediately as a business expense while payroll gets taxed, so the code nudges you toward machines. That is a capital expenditure argument and it does not apply to tokens at all. API spend is already an operating expense, deducted in the year incurred, treated exactly like wages. There is no asymmetry there to correct. The real wedge is payroll tax, call it eight to ten percent loaded, which compute does not carry. It is a defensible argument. He did not make it.
And Human Reserved protects the wrong people. His emblem for it is elder care, the caregivers who looked after his father through Alzheimer’s, and the passage is genuinely moving. Health aides are the control group in the paper he is citing. They are the occupation where employment for young workers has been growing fastest. Human Reserved would wall off the one category the data says is already safe while doing nothing for junior developers or paralegals, because nobody is firing them. They are simply not being hired, and there is no statute that makes a company open a requisition.
So here is what I would tax instead. If you are going to tax something, tax the outcome rather than the mechanism, and the number I would tax is revenue per employee. Not tokens, not robots, not compute. Revenue divided by headcount, two figures already sitting on every return a company files, neither of which requires anyone to determine whether AI was involved at all.
That last part matters more than it sounds, and it follows from everything in the first half of this piece. The best labor economists in the country, working with firm-level administrative payroll data, could not cleanly separate the effect of AI from four simultaneous confounds. Any tax that requires a government to identify AI-caused displacement inherits that identification problem and will lose to it, in court and in practice. A tax on revenue per employee never has to ask the question. It measures output per human being, and it does not care whether the leverage came from a language model, a shell script, a better process, or a team in another time zone.
It also survives contact with how inference actually works, which a token levy does not. Taxing tokens taxes verbosity, not substitution. The retrieval pipeline burning fifty thousand tokens to assemble a briefing for a human being displaces nobody and pays a fortune. The two-hundred-token classifier running in a loop can eat an entire queue and pays almost nothing. And because metered API calls are taxable while local inference is not, a token tax is a direct subsidy for running open weights on hardware you own and for routing through an endpoint in another jurisdiction. I could restructure around it in a weekend, and so could everyone I know.
The objection I expect is that this falls hardest on firms with enormous revenue per head. The platform company running billions against a few thousand employees, the fab, the pipeline operator. That is not an objection. That is the entire point. Those firms have captured an extraordinary share of value added with very little labor for two decades now, which is precisely the distributional problem Gates spends six thousand words worrying about, and not one existing instrument touches it. They did not stop hiring because of AI. They never hired in the first place, and they have been rewarded for it continuously.
Meanwhile look at who it spares. A full-service restaurant turns something on the order of a hundred thousand dollars of revenue per employee. A contractor, a clinic, a machine shop, a landscaping crew, a daycare: all of them sit at the bottom of this distribution, all of them need hands they cannot afford, and all of them would owe little or nothing. The businesses that are labor-hungry and cash-poor get the marginal hire back at closer to its real cost, funded by the businesses that have been converting output into margin without employing anyone.
Read it against payroll tax and it is an exact inversion. Payroll tax charges you for employing somebody. A tax on revenue per employee charges you for not employing anybody. Today the code does the first aggressively and nothing at all does the second, and that asymmetry is the one Gates gestured at before answering it with a levy on tokens.
There is one refinement I would make before anybody drafts this, and it is a real one. Gross revenue is the wrong numerator, because it does not separate leverage from pass-through. A grocery distributor moves enormous revenue per employee on margins of one or two percent. A software company moves less revenue per head and keeps a third of it. Use value added per employee instead, revenue minus purchased inputs, which is the standard measure of labor productivity and which most tax authorities already compute for VAT purposes. That catches the firm keeping the surplus and leaves alone the one merely moving volume.
And here is what makes it different from everything in Gates’s essay. His token tax, his Human Reserved carve-out and his new international institutions all arrive downstream, after the schema has been written and the coordinator has already been removed from the loop. A tax on value added per employee changes the arithmetic of the schema itself. If keeping a person in the loop lowers what the company owes, then whether the tool call proposes or commits stops being purely a margin question and becomes a tax question, on the Tuesday afternoon when somebody is actually deciding it.
It is still gameable. Spin the high-margin function into a separate entity, or buy something labor-heavy to dilute the ratio, and the number moves without anything real changing. Every tax on a ratio invites that. But it is gameable in ways that show up on a consolidated filing, which is more than can be said for a levy on a metric the taxpayer generates himself and can relocate to a box under his desk.
And it has the shape the spam story tells you to look for. Spam was not beaten by pricing the input. It was beaten by verifying the outcome: protocols that check whether a sender is who it claims to be, and classifiers that score a message on what it actually is. Gates reached for the postage stamp twice, twenty-two years apart. Both times the answer turned out to be measuring the result.
None of this means the transition is going to be gentle. The telephone exchanges hurt an enormous number of young women and did it inside a decade, and the handloom weavers were not compensated by anyone’s aggregate statistics. Gates is right that this deserves serious attention and right that AI companies should not be the ones defining the remedy to their own externalities.
But he had the better half of the paper in front of him, and he described it correctly somewhere else in his own essay. In the education section he praises AI tutors that hold the answer back and make the student arrive at it, preserving what researchers call productive struggle. That is augmentation, defined precisely, by him, in his own words.
He applied it to twelve-year-olds and not to work. Twenty-five years after the MGM Grand, he still has the right technology and the wrong idea about how to point it.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment, a Las Vegas enterprise software and AI consultancy. He has spent nearly thirty years shipping production systems on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy and Allegiant Air, and was employee number six at Break.com during the user-generated content era. He was a finalist for the Los Angeles Business Journal CTO of the Year award.
His current work is production agentic AI: MCP servers, the Microsoft Agent Framework, RAG pipelines, SQL Server 2025 vector search, and the PLogger observability framework. He operates ShopSnap eCommerce platform, and builds agentic systems including Cara, a shopping assistant for ShopSnap storefronts.
He writes about the gap between what executives say about AI and what practitioners actually ship. The .NET AI guy that ships.