Imagine getting a text from an AI recruiter. Not a chatbot on the careers page. A real recruiting funnel that you didn’t opt-in to.
Now “June” wants a pre-screen.
You answer a few questions before June gets confused. Eventually, you learn that if you don’t complete the conversation, your application is toast.
Welcome to 2026, a year where we’ve decided this is a reasonable way to treat a person who wants to work.
Do you wanna change the world?
There was a time when the tech industry was very good at making work feel like a place you wanted to be.
Free lunch on Wednesdays. Bacon and egg breakfasts (and beer) on Fridays. Ping-pong tables and music rooms.
When I arrived at work one morning, my boss, who had recently purchased a PS4 for the office, had also left a brand new copy of Mortal Kombat X on my desk.
In the 2010s, tech companies competed for talent with gourmet cafeterias, workday massages, and offices designed to keep you around.
The amenities were concrete, and the vision of work tech pitched was self-affirming: you’re smart, you’re building something important, and you’re surrounded by other smart people who are excited about it too.
You weren’t just working. You were working somewhere cool.
The overarching message—in our cofounders’ keynotes and gesticulating hands—was that we were changing the world.
But beneath the beer and beanbag chairs, things were more fragile than they seemed.
Bigger headcounts signaled bigger profits to the people writing checks, whether or not there was enough real work to justify extra staff.
One venture capitalist put it plainly in 2023: Google and Meta had hired thousands of people to do fake work.
At Dropbox, free dinner was part of the abundance. By 2016, the company had pushed it to 7 PM. One $15 plate of salmon for three more hours of unpaid overtime.
The math was surreal. According to an internal company email that later leaked, Dropbox was spending at least $25,000 per employee, per year on perks.
Some of the perks had thorns. As another VC told Y Combinator in 2018, whoever had time to play pool was usually gone within weeks.
You weren’t getting lunch. You were working through it. You weren’t putting in long hours—you were financing a future you couldn’t possibly foretell.
Perks were the baby version of the deal. Eventually, vesting schedules outpaced office toys, turning “you don’t want to leave” into “you can’t afford to.”
None of this was “enshittification” yet. It was the stage that makes enshittification possible—a system being good to you on purpose, so you don’t notice the moment it stops.
Cory Doctorow named the process in 2022 after watching it happen to platforms.
By 2025, he’d turned the same lens on tech workers in the aptly titled essay, The Enshittification of Tech Jobs, where he argues that scarcity, not solidarity, was the only thing protecting us from becoming disposable.
He’s not wrong—but labor comes with a different kind of lock-in.
You could always leave
The funny thing is, many of us enjoyed working more than we probably realized. Not the free lunch, but the work.
There’s pleasure in getting good at something—in recognizing a problem faster than you used to, in knowing what makes a headline work, or which line of code will break something three systems away.
Through the simple act of thinking our way through the work, we built instincts that are difficult to put on a resume because they don’t look particularly impressive written down.
Then one day you realize you know what you’re doing. That feeling mattered too.
For all the ways tech companies manipulated our attachment to work, they couldn’t fabricate the satisfaction that comes from learning, making something good, or solving a problem.
Our mistake was assuming companies valued that pleasure for the same reasons we did. They didn’t pay us to enjoy the work. Our enjoyment just made the arrangement easier.
We were more likely to fix problems on the fly during the day, and to keep dancing with an idea well into the night.
But enjoying your work and accepting the terms of the job are two very different things.
When the pandemic disrupted the old routine, workers briefly had enough leverage to renegotiate the employment bargain.
The New York Times called it the YOLO Economy: burned-out white-collar workers quitting stable jobs, going freelance, starting businesses, and pushing back on plans to return to the office.
I remember texting the article to a content friend and exchanging giddy messages.
Weeks later, someone gave it another name. The Great Resignation.
Then people actually started resigning.
In 2021, the U.S. quit rate reached a 20-year high. Low pay and lack of advancement were the most common reasons workers gave for leaving, but 57 percent said feeling disrespected also played a role.
By October, that feeling had become a cultural mood.
The New York Times began reporting on Reddit, where quitting texts had turned bad bosses into public entertainment. That attitude was now showing up in white-collar work, too.
Ask a Manager’s Alison Green, who’s been fielding workplace complaints for two decades, told the Times something had changed in her readers: “It’s almost like there’s a mob energy out there.”
The Times had a simpler word for it: leverage. In tech, we had a lot of it.
Companies had over-hired, and some jobs were almost certainly bullshit. Some tech workers were outrageously well paid, and some people could walk away from jobs for reasons most workers throughout history could only dream of.
That’s exactly what worker leverage looks like when labor is scarce. And everyone was writing about it.
In August 2021, WIRED described Silicon Valley as experiencing some of its most frenzied hiring since the dot-com boom.
Companies were dangling remote work, signing bonuses, and elaborate perks to attract software developers, while workers had what the magazine called “more freedom than ever to vote with their feet.”
By November, WIRED was documenting what workers were doing with that freedom—what Harvard called the Great Reconsideration.
Companies now had to contend with COVID-tested demands for remote work, more flexibility, better workloads, and work that actually felt worth doing.
Return-to-office mandates were enough to send engineers walking, with some accepting less money to continue working from home. Remote work was part of our leverage.
One of its consequences was disrupting some of the structure that made the old bargain feel intimate. Free lunch, spontaneous play, after-work drinks—whatever the perks, they depended on people sharing a space.
As work moved online, more of the employment relationship moved into Slack threads, tickets, documents, dashboards, and calls. The company encountered more of the worker through output, and the worker encountered more of the company as a system to be supplied.
Not everyone used their leverage to leave. By 2022, “quiet quitting” put a name on another worker response: stay in the job, but stop volunteering the extra.
At the time, freelancing felt like a preference. In retrospect, I had the luxury of being choosy because workers had options.
If somebody made a bad offer, you could say no. If the work got stupid, you could leave. If a company behaved like they were doing you a favor, there was a reasonable chance somebody else was hiring.
Then the market turned.
Round after round of tech layoffs pushed skilled workers back into the job market, starting in late 2022.
By April 2024, WIRED reported that tech roles were receiving hundreds or even thousands of applications. One health-tech company received more than 3,000 applications for a single data-science opening.
The power was back in the industry’s hands, and Indeed’s job postings data shows just how steep the reversal was.
By July 2025, tech postings were 36 percent below their pre-pandemic baseline.
Indeed called it “a sea change,” but its chart used less delicate language: U.S. tech job postings have gone from boom to bust.
The bust didn’t make every worker equally vulnerable. It turned the market into musical chairs with fewer seats—and some people entered the game with much better odds of finding another one.
The Great Reconsideration depended on having somewhere else to go. By the time the boom ended, that path was considerably less clear.
You just have to keep up
Tech hiring was already in freefall when generative AI entered the workforce dressed as productivity-boosting assistance.
The promise sounded familiar: work smarter by automating the repetitive stuff, leaving more time for “the work that matters.”
That’s how I remember speaking about AI in marketing materials when it mostly meant machine learning and automation.
This time, the technology moved from mostly supporting the work to producing it.
Suddenly, there was no reason we couldn’t up our cadence. The previous process of learning and crafting was replaced with a chain of commands.
Build AI pipelines that turn a single report, webinar, or book chapter into 30 downstream pieces.
Once producing an artifact looks cheap enough, almost anything can become raw material for something else—even when the resulting formats don’t serve the same purpose or demand the same decisions.
So we kept clearing the queue, hoping to get back to improving how the work got done. Clearing it only made room for the next request.
When AI lets you finish something twice as fast, you don’t get half the day back to explore “the work that matters.” You get more work. More output isn’t really the problem. Treating projected time savings as proven capacity is.
The extra work doesn’t have to be yours—AI makes it much easier to create work for somebody else.
A product manager can generate dozens of tickets for work that doesn’t match the client’s actual systems, leaving engineers to untangle a fictional backlog.
A marketing manager can run a writer’s draft through an LLM and return useless, hallucinated feedback on arguments the piece never made.
AI hasn’t made us more qualified to intervene. It’s made intervention easier to manufacture.
The dynamic gets even weirder when creating something takes less time than understanding whether it’s any good.
In software, code can now materialize faster than humans can confidently review it, so sprints fill up, PRs get bigger, and changes that should have been small arrive thousands of lines long.
The once-tangible perks have been replaced by monthly token budgets spread across AI models, coding assistants, and copilots, all carrying an implicit question: if this makes you faster, why aren’t you producing more?
Producing more is often table stakes as organizations increasingly frame AI use as a condition of continued employability. Use it, or become obsolete.
Many companies track AI adoption regardless of whether the tools improve output. Some mandate that workers use it, then hold them responsible when AI introduces errors. Still, the expectation remains. Teams should be able to absorb more.
By 2026, Glassdoor employee reviews with AI mentions were up 240 percent year over year. As adoption went up, positive sentiment went down.
In 2019, 81 percent of reviews mentioning AI described it positively. By 2026, only 43 percent did.
The backlash wasn’t limited to job-replacement fear—only 20 percent of AI-critical comments fell in this category.
Workers also complained about forced adoption, surveillance, distraction from the actual business, and leaders expecting productivity gains the technology can’t reliably deliver.
Company leaders, meanwhile, were among the most enthusiastic: 67 percent of their AI comments were positive.
That’s because the people setting the expectations don’t always have to deal with what happens when the technology falls short. Workers do.
Use the tool because it saves time. Spend the saved time checking what it produced. Fix what it got wrong. Then produce more because the tool saved time—and make sure somebody can see that you used it.
The work you used to enjoy doing is something you increasingly route around.
Engineers spend less time building and more time prompting, reviewing, troubleshooting, and pushing changes through. The job feels less like engineering and more like standing at the end of a conveyor belt, waiting to inspect whatever the machine spits out.
The old arrangement tried to deepen your attachment to the work. The new one can afford to flatten it.
You don’t have to love the job anymore. You just have to keep up.
Who knows how this works?
The problem with routing around expertise is that eventually you stop producing it.
Senior engineers don’t just know how to write code. They know why the ugly workaround exists, what breaks when you touch it, and which seemingly obsolete behavior a customer still depends on.
In his 2025 book Breakneck, Dan Wang calls it process knowledge: practical know-how gained through experience that doesn’t transfer neatly through blueprints or documentation.
It lives across people, relationships, habits, exceptions, and years of solving problems that didn’t arrive with instructions. And AI makes it unusually easy to underestimate.
Once an organization can feed years of documents, decisions, conversations, and work product into an AI system, it can seem like the company’s captured what its workers know.
In reality, you can’t recreate invisible labor by reverse-engineering a finished artifact.
But once enough of a worker’s visible output looks reproducible, the knowledge underneath it can seem redundant too.
That’s the bet behind AI-driven headcount cuts: if the machine can perform enough of the work we can see, maybe we don’t need the person who used to perform it.
Some companies have already had to hire people back. In a 2026 Orgvue survey, nearly a third of organizations that cut staff expecting AI savings later rehired when those savings failed to materialize.
Almost a quarter admitted they’d made the cuts based on general assumptions about AI rather than analysis of the roles themselves.
They modeled the tasks, but they hadn’t necessarily understood the job.
In software development, that gap can hide behind all the usual signals that something works.
Generated code can compile, pass tests, and look fine in review while still violating an assumption the checks don’t capture. An engineer who’s lived with the system may know exactly which assumption that is.
So how does the next generation become experienced enough to make that call?
Using AI doesn’t necessarily make developers less skilled. An experienced engineer who generates routine code and has enough context to validate it may be using the technology exactly as intended.
But what happens earlier in the apprenticeship, when less-experienced developers spend more of their time generating, integrating, and reviewing work they once would have built and troubleshot themselves?
Developers describe writing less code themselves, relying on models to navigate unfamiliar systems, and losing some of the intuitive understanding they used to develop while building.
Nothing has to break for understanding to erode. Sometimes the software works. The developer just feels less certain about why.
Sonar found that tension in its 2026 survey of more than 1,100 developers.
Developers with up to 10 years of experience reported meaningful gains from AI, including higher productivity and more time to advance their skills.
But the same developers were also more concerned about weaker codebase understanding, erosion of their coding abilities, and the effort required to review AI-generated code.
AI can make developers faster and change how they become experienced at the same time.
One engineer spent years studying software design and maintainability because he saw them as fundamental to good software. Now he watches as messy code moves from failure state to operating norm.
As long as LLMs can navigate and repair the mess, organizations have less incentive to keep the system legible for humans. Some are already optimizing for “agentic experience” over developer experience.
But letting LLMs absorb more of the burden doesn’t make the downstream work disappear.
Code creates value through what it does. The code itself creates obligations. Produce more of it, and you leave more behind that someone may eventually have to understand, repair, or remove.
Sonar found signs of that downstream cost too. More than half of developers surveyed said AI had added technical debt by producing code that looked correct but wasn’t reliable. Forty percent cited unnecessary or duplicative code.
Although this section may suggest otherwise, code isn’t the only place process knowledge can disappear.
A company can cut members of its support team and discover later that nobody remembers why certain customers require exceptions.
It can lose the person who knew which parts of a brittle internal process couldn’t be standardized easily, or the one everyone asked when the documentation stopped making sense.
What looks like a repeatable process often depends on somebody knowing where the process stops working.
Output can keep moving, even as experienced workers leave with knowledge that was never captured, and as fewer people get the chance to develop that knowledge.
The loss may only become obvious when something goes wrong and fewer people remain who understand why.
Then workers started saying no
By 2026, workers had begun pushing back in ways employers couldn’t ignore.
At Duolingo, management tried adding AI usage to employee performance reviews.
When workers questioned whether that meant using AI even when it didn’t improve the work, the company quickly reversed course.
CEO Luis von Ahn later said the standard should be simpler: do your job as well as possible, and use AI when it helps.
These workers weren’t refusing AI. They were refusing the premise that using the tool is evidence of doing good work.
Duolingo’s performance review conflict is the tech-industry version of a much broader fight.
Are employers the only ones who get to decide what happens to the job once automation and AI become part of it?
At major East and Gulf Coast ports, shipping companies and terminal operators wanted to expand the use of semi-automated cranes that move shipping containers. Dockworkers feared the machines would eliminate jobs they’ve traditionally done.
When roughly 45,000 International Longshoremen’s Association members walked off the job at 36 ports in 2024, wages were one dispute. Semi-automated cranes were another.
The strike spanned three days before a tentative wage deal sent workers back to the ports, leaving the automation issue unresolved—and a second walkout on the table.
In January 2025, the union and port employers reached an agreement. Under the final six-year contract, full automation stayed off the table. Semi-automated cranes could be introduced, but with human operators and negotiated staffing protections.
The technology still moved forward. What changed was who got a say in what moving forward would mean.
At Kaiser Permanente, the same fight landed somewhere harder to quantify: professional judgment.
Advice nurses field patient calls, assess symptoms, and help determine what kind of care someone needs—and how urgently they need it.
Although Kaiser says it doesn’t use “average handle time” to assess performance, seven current and former nurses told CalMatters that calls lasting more than 15 minutes routinely drew criticism or performance meetings, and that call time factored into monthly scores.
Nurses also said they were expected to follow scripts and limit how much advice they gave, regardless of the call’s content.
One nurse stayed on the phone with a suicidal patient for more than an hour waiting for police to arrive before she could safely hang up. She knew that one call could impact her numbers for weeks and trigger questions from management.
Another cut a call short because she worried the extra time would be held against her. The patient had just learned she had a terminal illness.
You aren’t calling Comcast. We’re dealing with life here.
Then Kaiser tested AI that rated nurses’ empathy and tone of voice. One system pressured nurses to shorten difficult conversations. Then another tried to grade the humanity they brought to those calls.
Nurses responded with a petition demanding patient privacy, transparency, and the right to exercise their own professional judgment. The AI-empathy test ended in November 2024, although union representatives were told it could return.
Other workers established AI boundaries before there was a grievance to fight.
At Solid State Books in Washington, D.C., seven unionized booksellers negotiated a three-year contract that allows AI to enhance their work, but bars its use to terminate, lay off, demote, or reduce the scheduled hours of bargaining-unit employees.
They didn’t ban AI. They put job displacement outside its permitted use.
Pennsylvania state workers negotiated something bigger—a role in deciding how AI would be used in the first place.
Through SEIU Local 668, they secured a binding agreement with the Commonwealth that created a standing labor-management group for generative AI, giving employees a formal role in identifying where the technology helps or creates problems.
AI can inform disciplinary decisions, but it can’t make them, and workers must receive training when new tools affect their jobs.
The agreement assumes the state will use AI. Workers negotiated for a role in deciding how.
At POLITICO and E&E News, the rules were already in writing. Then management broke them.
Their 2024 contract required 60 days’ notice and bargaining before new AI tools could materially change employees’ duties or lead to layoffs.
AI used in news-gathering also had to follow newsroom standards and include human oversight.
POLITICO launched two products without following those rules.
Live Summaries generated unedited coverage of major political events, including the 2024 Democratic National Convention and vice presidential debate. In one of those summaries, the AI misattributed political actions and used language barred by POLITICO’s stylebook.
Report Builder generated 500-word policy reports for subscribers without editorial review. Among its errors: claiming Roe v. Wade was still law and inventing lobbying causes for groups that didn’t exist.
The Guild filed grievances. When management didn’t resolve them, it took POLITICO to arbitration.
The Guild won. The arbitrator found that POLITICO had violated the contract by launching the tools without the required notice, bargaining, or human oversight.
In May 2026, POLITICO agreed to shut both products down. The contract had turned an objection into something enforceable.
POLITICO’s contract is one example of a greater effort.
By July 2026, The NewsGuild said it had negotiated 90 contracts containing protections against AI abuses by management.
The NewsGuild is a sector of the Communications Workers of America (CWA), which represents workers in media and many other industries.
Across CWA, members are bargaining for human oversight, limits on surveillance and automated discipline, protections against work intensification, transition plans for displaced workers, and a share of the economic gains AI creates.
Nobody’s demanding the old employment bargain back.
The fight is over who gets the gains and who absorbs the costs. Does saved time become time back or another assignment?
Can a worker override an automated judgment? Can companies use the technology to monitor or discipline them? What happens when it changes or eliminates the job?
During the talent wars, scarcity did some of the negotiating for tech workers.
The dockworkers’ leverage worked differently. Together, they could stop the ports where they already worked. Their power didn’t depend on another employer waiting on the other side of a resignation.
It depended on making the current employer negotiate.
Is labor optional?
The old bargain was manipulative, but it was revealing.
Dropbox served dinner because keeping someone in the office for another few hours might be worth more than the fish.
Companies spent lavishly on offices, retreats, equity, and the mythology of changing the world because they wanted us close to the work. They wanted us attached.
Their strategy wasn’t necessarily generous, but it depended on a particular idea about where productive value lived: in people you could attract, retain, develop, and persuade to care.
The current investment pattern looks a lot different.
Companies now increasingly invest in model access, token spend, copilots, compute, and whatever else promises to increase the amount of work each person can produce.
The question following closely behind is how many people the organization still needs once those productivity gains arrive.
Salmon and tokens aren’t equivalent perks—if anything, they point in opposite directions.
One was an investment designed to keep the worker around. The other can become part of the argument for needing fewer workers at all.
At the height of the tech boom, increased headcount could signal organizational strength. So companies hired ahead of demand, created roles and built teams because investors expected to see them, and competed for workers they were terrified of losing to other upstarts.
Now the flex runs the other way: how much can the company produce without hiring? How many jobs can disappear without the output disappearing with them?
That inversion makes one of the more utopian promises about AI sound more foreboding than it first appears.
Labor will become optional.
Labor can become more optional to an employer long before employment becomes optional to the person who needs rent, food, healthcare, and a reliable paycheck.
The people who describe the future in terms of abundance tend to gloss over that fact. If machines become capable enough, the story goes, people will be liberated from drudgery.
Maybe.
But getting from “companies need less labor” to “people need less employment” requires substantially more than futurist fantasies or a model getting better at following predefined requirements.
Until then, a worker can experience the worst of both worlds—increasingly treated as replaceable while economically dependent on being employed.
Professional enshittification is more than work getting worse. The employment bargain deteriorates at the same time workers lose leverage to reject its new terms.
How far employers can push that bargain depends on what they risk by pushing too far.
For a while, the credible threat of exit created that risk. As alternative jobs became harder to find, employers had less reason to fear that worsening the bargain would cost them workers.
The worker resistance happening now is building different constraints that don’t depend on having somewhere else to go: contracts, bargaining rights, human oversight, limits on surveillance and displacement, and a say in what happens to the gains.
None of those constraints requires the technology to disappear. They give workers a say in how it’s used, instead of leaving that decision entirely to the people buying it.
Still, changing the terms of the job doesn’t necessarily change what people are willing to bring to the work.
Not everyone asks that much of a job. For some, work has always been mostly transactional—time and skill exchanged for money, with the rest of life hopefully carrying the things work doesn’t.
Others learn to treat work transactionally. A system that rewards compliance more reliably than judgment gives people plenty reason to stop treating work with the same care they’d invest in something of their own. Caring less can be an adaptation.
For those who do care about it, though, the work itself can give something back.
There’s pleasure in becoming good at something—seeing the problem before anyone assigns it, knowing why the ugly workaround is the best solution, or making something better because you can tell the difference.
The ambition and curiosity are very real. So is the silly satisfaction of realizing you finally know what you’re doing.
Those parts of us don’t disappear because a company discovers it can produce the code, answer the question, or generate the blog post without you.
What can disappear is the job being a place where those parts of us matter.
Work has never had a monopoly on the pleasure that comes from ambition, curiosity, or mastery. But for a long time, our society’s channeled a great deal of that energy through work.
When that arrangement changes, we’re not just dealing with whether we’ll work. We’re grappling with what still moves us when work no longer can.
Maybe that’s what the old bargain was revealing all along.
What makes the engine go?
Desire, desire, desire.
