94 Studies Measured What AI Did to the Job Market. Nobody Likes the Median.
The first PRISMA systematic review of observed AI labor displacement synthesized 94 empirical studies covering 2020 to 2025 and found a median 23 percent decline in entry-level software development and content creation job postings across high-income economies. Cross-referencing with Bureau of Labor Statistics employment data, that percentage translates to roughly 218,000 fewer US job openings. Meanwhile, Upwork just laid off 24 percent of its staff and Fiverr's stock is down 35 percent this year. The displacement everyone predicted is here. It looks nothing like anyone expected.
Twenty-three percent. That is the median decline in entry-level and mid-level software development job postings across high-income economies between 2022 and 2024, measured not by a think tank's forecast model or an AI company's self-serving projection but by a systematic review of 94 empirical studies that tracked what actually happened after large language models entered the workforce. The number appears in a paper published May 7 in Frontiers in Human Dynamics by Nassim Dehouche at Mahidol University, and it is the first time anyone has assembled the full empirical record of AI-driven labor displacement into a single, peer-reviewed synthesis following PRISMA 2020 guidelines, the same gold-standard methodology used in medical meta-analyses.
Individual studies span 14 to 41 percent, depending on geography, occupation, and measurement window. A 14 percent decline might be cyclical, perhaps a post-pandemic correction or a hiring slowdown in response to rising interest rates that compressed tech spending across the board. But 41 percent is not cyclical. And the median of 23 percent, calculated across 42 studies with quantitative extraction, sits in territory that no plausible macroeconomic confound fully explains.
What 94 Studies Actually Show
Dehouche's review began with 1,847 records from six academic databases: Scopus, Web of Science, EconLit, SSRN, IEEE Xplore, and Google Scholar. After de-duplication and screening, 94 studies met inclusion criteria for qualitative synthesis and 42 for quantitative data extraction. Dehouche set a strict eligibility bar: studies had to present original empirical findings on labor market outcomes (not predictions), cover the period from January 2020 onward, and either use quasi-experimental methods or document systematic trends in AI-exposed occupations with explicit comparators. Studies that relied on task-content analysis to estimate future vulnerability, the approach that powered Frey and Osborne's famous 2017 estimate that 47 percent of US jobs were at high risk, were excluded. This review wanted receipts, not forecasts.
Four patterns emerged from the synthesized evidence. First, the 14 to 41 percent reduction in entry-level and mid-level job postings for software development and content creation roles, concentrated in high-income economies between 2022 and 2024. Second, a 15 to 22 percent wage premium for workers who demonstrated AI-augmentation capabilities, meaning the ability to use AI tools productively rather than competing against them. Third, a 2 to 21 percent reduction in posting volumes for automatable creative tasks on online labor platforms after ChatGPT's November 2022 release. Fourth, heterogeneous sectoral effects: infrastructure, security, and quality-assurance roles expanded during the same period that developer and content-creator roles contracted.
Wage polarization accompanied the posting declines. Workers who adapted to AI tools earned more. Workers who competed against them earned less. Or stopped earning altogether. Credential erosion intensified: the traditional signaling value of a computer science degree weakened as employers increasingly valued demonstrated AI fluency over formal education. Geographic unevenness was severe, with developing economies reliant on cognitive-services outsourcing facing what the paper calls "disproportionate disruption through both direct exposure and indirect demand-erosion channels."
Translating Percentages to People
Percentages obscure human scale, so consider the math. The Bureau of Labor Statistics reported 1,897,100 software developer, quality assurance analyst, and tester jobs in the United States as of 2023, with a median annual wage of $133,080. Conservative estimates put entry-level and mid-level positions at roughly half that total, approximately 949,000 roles. Apply the median 23 percent posting decline to that pool: 218,000 job openings that existed in 2022 no longer exist in 2024. At the high end of the range (41 percent), the figure rises to 389,000.
These are not layoffs in the traditional sense. Most of these positions were never eliminated from a payroll; they were never posted. Companies that previously hired three junior developers now hire one senior developer who uses Copilot, Cursor, or Claude to do the work of three, and jobs vanish before they materialize, which makes them invisible to layoff trackers and unemployment statistics but perfectly visible in exactly the kind of job-posting databases that the systematic review examined, where every unfilled opening that once would have been listed now simply never appears at all.
Now apply the wage premium: if AI-augmented workers earn 15 to 22 percent more than their non-augmented peers at the median salary, that translates to $19,962 to $29,278 per year per worker. Over a career, the gap compounds: a 35-year-old developer who learns to use AI tools effectively will earn roughly $600,000 to $900,000 more over the next 30 years than a peer with identical credentials who does not, assuming the premium holds at current rates and the rest of the compensation trajectory remains constant. This is a back-of-the-envelope calculation with obvious limitations, but the directional implication is clear: the labor market is bifurcating, and the dividing line is not education, not experience, and not intelligence but adaptation speed.
The Canaries Stopped Singing
If you want to see where labor displacement arrives first, watch freelance platforms. Freelancers have no employment contracts to buffer transitions, no severance packages, no HR departments managing gradual headcount reductions. When demand for a task category evaporates, it evaporates overnight in the marketplace data.
Upwork's Q1 2026 earnings tell the story. Revenue grew 1 percent year-over-year to $195.5 million. Gross services volume was flat at $987.1 million. Net income fell 17 percent to $31.5 million. Upwork laid off 24 percent of its workforce and announced $16 to $23 million in restructuring charges, and its own AI-related contracts, the very category that was supposed to save the platform, are declining rather than growing.
Fiverr's stock price fell more than 35 percent since January 2026. Independent analyses of platform spending data show that freelance spending as a share of total enterprise software spending dropped from 0.66 percent in 2021 to 0.14 percent in 2025, while AI tool spending surged from zero to 3 percent over the same period. Over half of active freelancers on major platforms in 2022 had stopped using them by 2025. What the freelance marketplace is experiencing is not a contraction driven by recession or a credit cycle but a structural replacement event, because the buyers found a cheaper, faster, always-available substitute that does not have a profile page on Upwork.
The Prediction Gap
In 2017, Carl Benedikt Frey and Michael Osborne at Oxford published the number that defined a decade of AI anxiety: 47 percent of US jobs were at "high risk" of automation. Policymakers cited it, newspapers headlined it, and labor economists spent years arguing about whether it was too high or too low.
Nine years later, the first comprehensive measurement of what actually happened tells a different story. Displacement is real. It is measurable. And it is concentrated in exactly the cognitive-task categories that Frey and Osborne identified. But it is not 47 percent of all jobs. It is 14 to 41 percent of job postings in specific entry-level and mid-level cognitive roles, predominantly in software development and content creation. Displacement is more surgical than the apocalyptic predictions suggested and more concentrated than the "transformation, not elimination" counter-narrative allowed, landing almost exclusively on the entry-level cognitive workers whose jobs overlapped most directly with LLM capabilities.
Daron Acemoglu at MIT, who has spent a decade building the task-based framework that predicted this pattern, identified the mechanism in his 2024 working paper: automation creates a "displacement effect" that reduces demand for labor in automated tasks and a "productivity effect" that increases overall output and potentially creates new tasks. What makes LLMs different from spreadsheets, ATMs, and shipping containers is the speed of the displacement effect relative to the creation of new roles. Previous automation waves unfolded over decades, yet ChatGPT reached 100 million users in two months.
What the Data Cannot Tell Us
Honest limitations deserve their own section, and this analysis has several that matter. The systematic review covers 2020 to early 2025, meaning it captures the GPT-3 through early GPT-4o era. Model capabilities have advanced substantially since then, meaning the observed displacement rates likely understate the current situation; most included studies come from high-income Anglophone countries, and the developing-economy data that does exist, particularly regarding Indian IT outsourcing and Filipino content-creation services, is thinner and methodologically weaker than the US and European evidence.
Job posting volumes do not equal employment levels. Postings are a leading indicator, not a census. Companies may have stopped listing entry-level roles not because they eliminated them but because they changed sourcing strategies: internal upskilling programs, bootcamp partnerships, or AI-assisted candidate screening that reduces public-facing posting volume. Observation windows in the underlying studies are short. Dehouche's own methodology section acknowledges the difficulty of disentangling AI effects from pandemic disruption, monetary policy tightening, and the tech sector's post-2021 valuation correction, all of which compressed hiring independently of automation and make causal isolation genuinely difficult even with well-designed quasi-experimental methods.
Worker-level longitudinal data is almost entirely absent: we know that 218,000 fewer job openings exist, but we do not know what happened to the specific people who would have filled them. Did they pivot to adjacent roles, leave the labor force, or take lower-paying positions outside tech? Aggregated posting data cannot answer these questions, and that gap constitutes the most urgent research need in this field.
The Strongest Case Against Panic
The best counterargument to structural-displacement readings of this data comes not from AI optimists but from economic historians. In 2000, credible analysts projected that solar power would remain expensive for decades, and they were right about the physics but wrong about China's industrial policy, which drove production volumes so high that experience-curve effects crushed costs faster than any model predicted. Consider the parallel: in 2024, researchers measured job-posting declines and inferred permanent displacement, but what if we are measuring a transition dip, the same temporary disruption that accompanied every major automation wave from textile looms to spreadsheets to containerized shipping?
Bank tellers did not disappear when ATMs arrived. Their numbers actually increased, because cheaper branch operations made it economical to open more branches, and the tellers' job description shifted from cash handling to customer service and sales. Accounting clerks did not disappear when spreadsheets arrived. Cheaper analytical capacity created demand for more analysis, and clerks became analysts. In each case, temporary posting declines preceded eventual job creation in adjacent, higher-value roles that nobody had named yet.
This argument has genuine force, and dismissing it would be dishonest. But it also has a structural weakness: the transition dips in previous automation waves unfolded over 10 to 20 years, giving workers time to retrain, institutions time to adapt curricula, and markets time to discover new roles. LLM deployment compressed the initial displacement into 18 months. Whether the creation side can operate on the same compressed timeline is an open, genuinely uncertain question, and pretending to know the answer in either direction is not analysis but advocacy.
What You Can Do
If you are a software developer at any level, the 15 to 22 percent wage premium for AI-augmented workers is not an abstraction. Learn to use coding assistants (Copilot, Cursor, Claude Code) at a level where you can demonstrably produce more, not just faster but better, reviewed output. When your next performance review arrives, quantify the difference. Companies paying the premium need evidence that the premium is justified, and workers claiming the premium need to provide it.
If you manage a team that has quietly stopped posting junior roles, acknowledge what you are doing and build an alternative pipeline. Not widely advertised: the invisible job elimination documented in this review creates a training gap that will haunt companies in five years when they need senior engineers and discover they never developed any, because the junior pipeline through which seniors are made was quietly disconnected in 2024. Apprenticeship models, structured mentorship programs, and AI-assisted onboarding that lets one senior engineer effectively train three juniors simultaneously are all viable options that cost less than discovering your talent pipeline dried up in 2024.
If you are a policymaker, the developing-economy angle demands immediate attention. Countries whose GDP depends substantially on cognitive-services exports, India and the Philippines foremost among them, face a demand-erosion channel that no domestic policy can fully address, because the jobs are vanishing in the importing countries. International coordination on workforce transition support is not a luxury. If you work in education, the BLS still projects 17 percent growth in software development jobs through 2033, a projection made before the evidence in this review existed, so update your enrollment guidance accordingly.
The Bottom Line
For three years, the debate about AI and employment has oscillated between two poles: catastrophists who warned that millions of jobs would vanish overnight and optimists who insisted that AI would only transform work, not eliminate it. Dehouche's systematic review, the first to apply medical-grade evidence synthesis to observed labor market data, suggests that both sides were partially right and partially wrong. Displacement is real. Measurable and concentrated in entry-level cognitive roles. It is not the 47-percent apocalypse that Frey and Osborne warned about, but neither is it the painless "transformation" that AI boosters promised. The median finding across 94 empirical studies is a 23 percent reduction in job postings for the specific categories where LLMs are most capable, combined with a 15 to 22 percent wage premium for workers who learned to use the tools that displaced their peers. The labor market did not collapse. It bifurcated, cleanly and fast. And whether you land on the higher or lower branch depends on what you do in the next 12 months, not the next 12 years.