Three years into the generative AI wave, the most sophisticated measurement efforts by the companies building the technology find something remarkable: almost nothing definitive. This document reads the evidence from every direction, and finds a story no single stakeholder is telling.
It synthesizes evidence from AI labs (Anthropic, OpenAI), platforms (Google Workspace, Upwork), capital allocators (Madrona, J.P. Morgan, Deutsche Bank), enterprise management (McKinsey, Deloitte, Conference Board), boards (IBM CEO Study), employees (PwC, Mercer, JFF), and independent academic research (Acemoglu, Brynjolfsson, Yale Budget Lab, Penn Wharton, Brookings). It privileges empirical findings over opinion, notes where findings conflict, and generates a thesis from the convergence.
What the evidence actually shows.
No displacement signal in aggregate data
The single most important empirical finding across all sources is the absence of detectable job displacement at scale. Anthropic’s March 2026 labor market research introduces “observed exposure,” a metric combining theoretical LLM capability with actual usage data. The finding: no systematic increase in unemployment for workers in the most AI-exposed occupations since ChatGPT’s release.
Yale Budget Lab (2025). Tracked occupational mix changes using the Current Population Survey and found them “unremarkable.”
Danish administrative records (2025). Linked ChatGPT usage to eleven exposed occupations. Zero effects on earnings or hours through 2024.
Hartley et al. (2026). 35.9% of U.S. workers using generative AI by December 2025. Small positive wage effects. No measurable decline in job openings.
ICLE literature review (2026). The emerging pattern is adjustment at the margin, through task reallocation and changes in career ladders, rather than broad displacement.
Penn Wharton projects AI will increase productivity and GDP by roughly 1.5% by 2035, the boost peaking in the early 2030s and then fading. Acemoglu’s own estimate is more conservative still, a 0.5% productivity gain over the next decade.
Young workers at the margin
The one consistent signal across multiple datasets is pressure on young workers, ages 22 to 25, entering AI-exposed occupations.
The mechanism is a hiring freeze at the entry level, not displacement of incumbents. Companies are not firing people because of AI. They are not hiring new people into roles where AI might eventually reduce need. The effect is invisible in unemployment data because workers who are never hired never become “unemployed.” They return to school, take different roles, or exit the labor force entirely.
Theoretical capability against actual adoption
OpenAI’s Signals data adds texture: a significant share of usage is asking, seeking information and advice, rather than doing. People are consulting AI, not delegating to it. The gap is not a technology problem waiting to be solved. It persists because of institutional friction, organizational complexity, tacit knowledge, and the economic rationality of human labor.
Adoption without transformation
The Conference Board’s 2026 C-Suite Outlook finds AI ranked the number one societal and technological factor CEOs believe could negatively affect their businesses, above political polarization. McKinsey documents the perception gap: executives estimate only 4% of employees use generative AI for a third or more of daily work. The actual number is closer to 13%.
The dominant management challenge is not that AI is replacing workers. It is this: we are spending on AI, cannot measure the return, and our workforce is adopting unevenly while we lack a coherent strategy.
A paradox of optimism and anxiety
PwC (50,000 respondents). Optimism about AI’s potential “greatly outweighs anxiety.” But daily usage remains low.
Google Workspace (ages 22 to 39). 92% want AI personalization. 91% say AI helped them contribute above their role level. 77% describe themselves as “active designers” of AI workflows.
Deloitte early career. 79% are excited about AI opportunities. One interviewee called AI “that first person you ask before going to a manager.”
Mercer 2026 (4,500+ U.S. employees). 63% enthusiastic about efficiency potential, yet 53% believe technology will affect their job security. Only a quarter regularly experiment with AI tools.
JFF (March 2026). A sentiment reversal: workers now more likely to call AI net-negative than net-positive. Employer-provided AI training dropped ten points. 38% of workers of color plan to change career pathway due to AI, against 23% overall.
4 Corner Resources (2026). 88% satisfied with their employers, yet 69% anxious about the job market. The two coexist.
The split is not young against old, or tech against non-tech. It is between workers who have the agency to shape their relationship with AI and those who do not.
Freelance markets show real displacement
Unlike aggregate labor markets, freelance platforms show clear, measurable displacement. An early-warning system for what may eventually register in full-time employment.
But at the top, the opposite. Adapted freelancers earn 40 to 60% more than before AI arrived. The floor collapsed. The ceiling rose. The pattern is not mass displacement. It is a hollowing out of the middle while the top is elevated.
AI-washing is doing more damage than AI
55,000 U.S. layoffs in 2025 cited AI. Twelve times more than two years prior. But of 1.2 million total job cuts, AI was cited in only 4.5%.
Forrester (January 2026). Many companies announcing AI-related layoffs do not have mature AI applications ready to fill those roles.
Deutsche Bank. Coined “AI redundancy washing”: companies citing AI because it sends investor-friendly signals, masking weak demand or overhiring corrections.
New York State WARN data. Since employers were given the option to cite “technological innovation” in legally required layoff notices, zero of 160 companies, including Amazon and Goldman Sachs, checked the box.
The most visible case study is Klarna. After replacing 700 customer-service agents with AI and publicly celebrating it, the company reversed course in mid-2025 and is rehiring human agents. Orgvue and Forrester found that 55% of companies that rushed to replace workers with AI now regret it. The fear-confidence gap, between what employees believe is happening and what is empirically happening, is now itself an economic variable.
Pouring money in while signaling doubt
Madrona’s Karan Mehandru writes that AI companies now sell outcomes, not tools, collapsing the boundary between product and services. His key insight: in AI, services are often how the system learns, and the system learning is the moat. The human-in-the-loop phase is where defensibility is born, not a cost to be eliminated. J.P. Morgan draws a direct parallel to the late-1990s telecom fibre buildout.
Seven propositions about what is actually happening.
AI is reorganizing work, not eliminating it.
Three years in, there is no detectable displacement at scale. What is happening is task reallocation. Some tasks move to AI, some change shape, and new tasks are created: verification, prompt engineering, workflow design, quality assurance. This is the automation and reinstatement cycle Acemoglu and Restrepo described. The critical variable is whether new task creation keeps pace with automation. So far, it has.
The gap between capability and adoption is structural, not temporary.
The 61-point gap is not a diffusion lag. It persists because of institutional friction, organizational complexity, tacit knowledge, and economic rationality. Human labor often remains cheaper than the full cost of automation once implementation, monitoring, error correction, and organisational change are counted.
The real displacement is invisible: entry-level workers who were never hired.
Young workers are not being fired. They are not being hired. The long-term consequence is not unemployment but a broken career ladder. If entry-level roles are how workers acquire tacit knowledge, institutional understanding, and professional networks, removing those rungs creates a mid-career talent crisis in five to ten years. No one is tracking this.
The dominant enterprise experience is adoption without transformation.
Two-thirds report productivity gains. One-fifth report revenue growth. One-third are deeply transforming. The bottleneck is not technology. It is organizational capacity for change. Insufficient worker skills as the biggest barrier is the enterprise equivalent of the capability and adoption gap.
Employee sentiment is bifurcating along lines of agency, not demographics.
Workers with the education, support, and initiative to shape their AI relationship experience it as empowering. Workers without those resources experience it as threatening. The 92% optimism and the net-negative sentiment are both true, about different populations.
The narrative layer is as economically significant as the technology layer.
AI-washing inflates public fear beyond what the data supports, undermines worker trust, and creates a credibility gap that will show up in earnings. The narrative that AI is taking jobs is doing more immediate economic damage, through anxiety, reduced engagement, and misallocated policy attention, than AI itself is doing to employment.
The winners will not be the fastest adopters but the most honest reorganisers.
Organisations reporting significant returns were twice as likely to have redesigned workflows before selecting AI tools. The technology works well enough. The organisations do not know how to reorganise around it. That is a human problem requiring human expertise.
What this means for the world of work.
2026 to 2028
Messy coexistence
Broad adoption continues while returns stay hard to measure. The entry-level hiring freeze deepens. AI-washing persists until investor patience runs thin. The freelance middle market keeps hollowing out. Anxiety rises faster than displacement. Organisations that communicate honestly and invest in reskilling retain talent. The rest bleed it.
2028 to 2032
The career-ladder crisis arrives
The cohort not hired into entry-level knowledge work in 2024 to 2028 is conspicuously absent from mid-career talent pools. The capability and adoption gap narrows but does not close. The most exposed occupations restructure fundamentally: fewer workers doing qualitatively different work, rather than mass elimination.
2032 onward
Distribution depends on choices made now
The historical pattern suggests AI, like every prior general-purpose technology, creates more economic value than it destroys. But the distribution depends entirely on choices made in the next three to five years, by those who invest in workforce transition, redesign career ladders, and build the capacity for continuous reorganisation.
The uncomfortable conclusions.
Your AI strategy is your workforce strategy. They are not separate line items. The 66% productivity against 20% revenue gap says AI makes existing work slightly faster but does not, by itself, generate new value. New value requires reorganised work, which requires reorganised people.
The biggest governance risk is not that AI displaces your workforce too fast. It is that your organisation adopts AI too superficially, spending on tools without redesigning work, and faces a compounding deficit in organisational capability. The 55% regret rate should be your risk metric.
The data says your job is probably safe near-term. But the nature of your job will change, and your employer may or may not invest in helping you navigate it. Your agency depends on willingness to learn the tools, and insistence that your employer provide clarity, training, and redesigned career paths.
The entry-level hiring freeze is the most important finding in the entire evidence base, and the least likely to generate a policy response, because it is invisible in unemployment statistics. By the time it shows up in aggregate data, the damage to career ladders will be structural.