There’s a new acronym reshaping how workers think about their careers: FOBO — the Fear of Becoming Obsolete. Unlike traditional job insecurity, FOBO isn’t about getting fired. It’s about becoming irrelevant. Four in 10 workers now name AI-driven job loss as one of their primary fears — a share that has nearly doubled in a single year, according to KPMG. Sixty-three percent say AI will make the workplace feel less human. Skill demands in AI-exposed roles are shifting 66% faster than they did just one year ago. In 2026, FOBO became the defining psychological condition of the American workplace.
After Dario Amodei, CEO of Anthropic, claimed last year that AI could eliminate 50% of entry-level white-collar positions within five years, he was joined within months by Microsoft AI CEO Mustafa Suleyman, who offered a similar outlook. More recently, Senator Mark Warner (D-VA) said that AI leaders themselves have been surprised and alarmed at the pace of disruption, and they are “literally consciously pulling back on their predictions because of the short-term economic disruption.” Warner put the new college grad unemployment at 35% within two years.
These are the predictions feeding FOBO — and they’re landing. A massive new study from MIT wants to pump the brakes. Not on the fear — FOBO, it turns out, is pointing in roughly the right direction — but on the timeline. And the timeline, it turns out, changes everything.
Researchers at MIT FutureTech published findings this week showing that AI’s march through the labor market looks far less like a sudden catastrophe and far more like a slow, rising flood — serious and accelerating, but not the overnight apocalypse that has dominated headlines and executive anxiety for the past two years.
“Rather than arriving in crashing waves that transform a certain set of tasks at a time,” the researchers write, “progress typically resembles a rising tide, with widespread gains across many tasks simultaneously.”
The study, titled “Crashing Waves vs. Rising Tides,” is one of the most comprehensive empirical examinations of AI’s real-world task performance to date. The team of nine researchers led by Matthias Mertens and Neil Thompson collected more than 17,000 evaluations of LLM outputs from domain-expert workers across more than 3,000 labor market tasks drawn from the U.S. Department of Labor’s O*NET classification system. Those tasks spanned everything from legal analysis to food preparation, management to computer science. More than 40 AI models were tested, ranging from GPT-3.5 Turbo to GPT-5, Claude Opus 4.1, Gemini 2.5 Pro, and DeepSeek R1.
For anyone gripped by FOBO, the core question the researchers asked is also the most unsettling one: Can AI complete these tasks well enough that a manager would accept the output without any edits? The answer is already yes — frequently.
Across all models and job categories tested, AI successfully completed roughly 50% to 75% of text-based labor market tasks at a minimally acceptable quality level. That’s not a future projection. That’s today. More specifically, the study found that by the third quarter of 2024, frontier AI models were already hitting a 50% success rate on tasks that take humans about a full workday to complete.
The improvement trajectory is steep. Between the second quarter of 2024 and the third quarter of 2025, frontier models went from clearing a 50% success threshold on 3- to 4-hour tasks to clearing the same bar on tasks that take humans an entire week. Failure rates are halving roughly every two to three years across the board, which translates to annual gains of 15 to 16 percentage points in success rates.
Extrapolating those trends — and the researchers are careful to note this represents an optimistic, upper-bound scenario — AI systems could complete most text-based tasks with 80% to 95% success rates by 2029 at a minimally sufficient quality level. For the majority of survey tasks, which take a few hours for a human to complete, the projected 2029 success rate approaches 90%.
MIT doesn’t use the phrase but this is FOBO, calibrated. The fear isn’t irrational — it’s premature. The water is rising. But the MIT data suggests the floorboards won’t be underwater by next Tuesday. The researchers’ most consequential line for anxious workers: “Workers are likely to have some visibility into these changes, rather than facing discontinuous jumps in AI-driven automation.” The rising tide gives you time to move. The question is whether you’re moving.
FOBO at the institutional level
Here’s the irony: even as MIT documents AI’s sweeping capability gains, most companies have yet to deploy the tools at all. FOBO isn’t just a personal condition, then — it’s an organizational one. According to Goldman Sachs economists Sarah Dong and Joseph Briggs, citing Census Bureau data in their March 2026 AI Adoption Tracker, fewer than 19% of U.S. establishments have adopted AI. Goldman projects that adoption will reach only 22.3% over the next six months.
Compounding that paralysis: only about one-third of workers say their employer is providing adequate AI training, guidance, or reskilling opportunities — down nearly 10 percentage points from 2024, according to research from workforce nonprofit JFF. Most companies are leaving workers to manage FOBO alone, without the infrastructure that would actually resolve it.
That gap has a measurable cost. Enterprise workers who do use AI are recapturing 40 to 60 minutes per day, according to OpenAI enterprise data from December 2025, and 75% say they can now complete tasks they previously couldn’t do at all.
“We continue to observe large impacts on labor productivity in the limited areas where generative AI has been deployed,” Goldman’s economists wrote. “Academic studies imply a 23% average uplift to productivity, while company anecdotes imply slightly larger efficiency gains of around 33%.”
Put simply: the companies using AI are pulling ahead. And the math is unforgiving. Across a team of 50, that 40-to-60-minute daily time saving translates to 33 to 50 hours of recovered productivity every single day. The race is on, then, but many companies are still strapping on their running shoes and waiting for the whistle to blow.
FOBO with a corner office
The MIT data lands at a moment when corporate leaders are scrambling to get their arms around a technology that, as one senior executive put it, is “outpacing the ability for humans and businesses to adopt it.” Joe Depa, the global chief innovation officer at EY, told Fortune in a recent interview that “the technology is in many ways ready, but it’s taking some time for us to … take advantage of it.”
Depa, who oversees AI strategy for one of the world’s largest professional services firms, described the pressure he sees across industries as relentless. “Every day there’s a new headline, every day there’s a new, you know, something that we have to get ready for. Every day, I get an email from my boss asking about some new event that happened somewhere in the world that’s raising the stakes of how fast things are moving within AI.”
That pressure is sharpened by a stark internal reality at many companies: 83% of executives — drawn from a survey of 500 business leaders — say they lack the right data infrastructure to fully leverage AI.
EY’s clients, based on 4,500 surveys, say they still lack the right data infrastructure to fully leverage AI. In other words, the technology is racing ahead while the organizational plumbing needed to actually use it lags far behind.
FOBO’s cruelest irony
That’s where the “rising tide” framing offers some reassurance to the many companies grappling with this dynamic. The MIT findings directly challenge research from METR, a prominent AI safety organization, which has argued that AI capabilities surge abruptly for specific sets of tasks — a “crashing waves” model that implies workers could suddenly find themselves obsolete with very little warning. “We find little evidence of crashing waves,” they wrote, “but substantial evidence that rising tides are the primary form of AI automation.”
The MIT data, drawn from realistic and representative job tasks rather than stylized benchmarks, consistently shows a flatter performance curve. AI doesn’t suddenly master a narrow set of tasks and leave everything else untouched. Instead, it gets broadly, incrementally better across nearly all task types and durations simultaneously.
“Workers are likely to have some visibility into these changes,” the researchers write, “rather than facing discontinuous jumps in AI-driven automation.” More broadly, the projection of AI improvement to a near-perfect automation level through the next three years, not the next 18 months of doomsday scenarios, provides what the researchers call “a window for worker adjustment, particularly in tasks with low tolerance for errors.” Furthermore, their estimates assume AI progress continues at the pace seen over the last two years, meaning it’s an upper-bound or particularly fast scenario. AI just may not keep evolving and advancing as fast as it has recently.
That matters for how companies plan and how workers prepare. A crashing-wave model demands emergency triage; a rising-tide model demands strategic adaptation. The MIT researchers argue the latter is the more accurate frame — though they’re emphatic that “gradualism is not inherently protective.”
There are meaningful differences by profession. Legal work had the lowest AI success rate among the domains tested, at just 47%. Installation, maintenance, and repair work — for text-based tasks specifically — topped the chart at 73%. Management tasks came in around 53%; healthcare practitioners at 66%; business and financial operations at 57%. In other words, no white-collar sector is immune, but some are considerably closer to the inflection point than others.
Depa said he sees this sorting happening in real time inside EY’s own workforce, and humans are acting unpredictably, even strangely at the prospect of this strange new work partner. The firm is the third-largest Microsoft Copilot user in the world, he shared, and the adoption data tells a generational story: junior employees are all in; senior leaders are lagging. “When I look at the breakdown,” he said, “two of my junior levels — high adoption, right out of the gate … and then when you get to the more senior levels, that’s where the adoption starts to drop off.”
He described a particularly worrying cohort: skilled, experienced workers who are simply refusing to use AI tools. “We’ve got some software engineers that are 10x, 20x more productive than last year using AI, like, they’re just killing it.” He said he’s seen workers go from “mediocre” to really “at the top of their game” once they master these new tools. At the same time, you have others “that used to be really, really strong software developers that are somewhat resistant to using AI,” he said. They have an attitude that they can do it better, so they don’t need the tool. “And they’ve gone from being top of their class to now bottom of the peer group, right. And those are the ones I worry about the most.”
The fear of becoming obsolete, in other words, is accelerating the very outcome that workers dread most. Left untreated, a serious case of FOBO becomes self-fulfilling.
These AI resisters, with tremendous functional skills and experience that are super critical, but productivity lagging their peer group at 10x or even 20x, “at some point, those individuals would have to find a different role,” Depa said. “And I think those are the ones that we’re trying to figure out.”
What’s still missing from the AI-at-work story
The MIT team is careful not to oversell its own findings. High task-level success rates, they note, don’t automatically translate into job displacement. The “last-mile costs” of integrating AI into actual workflows — organizational friction, liability concerns, the economics of deployment at smaller firms — remain significant barriers that are poorly captured by any benchmark.
Near-perfect AI performance on most tasks also remains years beyond 2029. The flat logistic curve that makes the rising tide gradual also means the final climb toward 99%-plus reliability is a long one, a meaningful buffer for error-intolerant professions in law, medicine, and engineering.
“While progress is significant,” the researchers write, “widespread automation, particularly in domains with low tolerance for errors, may still be some distance away.”
The bottom line is more complicated than either the doomers or the dismissers want to admit. AI is already capable, improving fast, and headed for most of your inbox in the next three to five years. But the transformation is likely to arrive as a steady, visible tide rather than a sudden drowning, which means the window to adapt is real, if not infinite. If you want to adapt, that is.
FOBO is rational. The MIT data confirms it. But the antidote isn’t denial or paralysis — it’s exactly what the workers thriving inside EY are already doing: treating AI as a tool, not a verdict. The window is open. The question is whether you’ll walk through it.
