In 2026, the loudest voices on AI and employment fall into two camps, and both are wrong. The doomers point to mass layoffs in customer support, junior copywriting, and entry-level coding, and predict half of all white-collar work evaporates within five years. The boosters point to record-low unemployment, a $297 billion funding boom, and Anthropic’s own data showing that 4 times as many workers use Claude alongside their jobs as use it to replace one, then declare the labor scare a moral panic. Neither reading survives the actual 2026 data.<\/p>\n\n
The peer-reviewed research on this question is finally starting to land \u2014 Eloundou, Manning, Mishkin, and Rock (2023)<\/a> at the OpenAI-MIT Economics team, the Stanford HAI 2026 AI Index<\/a> released in April 2026, Anthropic’s Economic Index<\/a> last refreshed June 26, 2026, and a substantial body of follow-on research \u2014 and a more careful answer than either headline has emerged. This piece is the synthesis that the news cycle keeps skipping over.<\/p>\n\n If you read yesterday’s coverage of the 2026 AI jobs landscape from a hiring angle<\/a>, you already know what the AI-skilled labor market looks like. Here is the other half: every job not on that list.<\/p> The predictable shape of the public debate is this. Someone picks a productivity win \u2014 Klarna replacing 700 customer-service reps with an AI, IBM cutting back its back-office hiring plans citing AI automation, AT&T’s 2025 internal memo flagged by Goldman Sachs research analysts<\/a> \u2014 and the doomers call it the start of the 21st-century white-collar collapse. Someone else picks a hiring stat \u2014 the U.S. Bureau of Labor Statistics’s July 2026 jobs report showing 4.1% unemployment, the OECD’s continuing record-low cross-country rate \u2014 and the boosters pronounce the scare over.<\/p>\n\n Both readings mistake layer. The Klarna layoffs, the IBM freeze, the resume-builder 2024 survey showing 1-in-4 employers planning AI substitution, the Slack survey showing 38% of desk workers are already using AI in their job: these are about task-level<\/em> adoption within occupations. The OECD and BLS unemployment rates are about headcount-of-workers<\/em> across the labor market. The two metrics answer different questions. Tasks get substituted without reducing headcounts when workers shift the time they saved to higher-value work or when companies have a hiring pipeline already frozen.<\/p>\n\n This is also exactly the resolution the Anthropic Economic Index<\/a> lands on, and it has the data advantage: it measures Claude usage by occupation, not by company announcement. The split matters.<\/p> The headline number in this space is from the Eloundou et al. “GPTs are GPTs” study, first posted to arXiv in March 2023<\/a> and revised in August 2023: 80% of the U.S. workforce could have at least 10% of their work tasks affected by LLMs, and 19% could see at least 50% of their tasks affected.<\/strong> It is the largest of the peer-reviewed exposure estimates and still the most-cited.<\/p>\n\n But “affected” is doing a lot of work. The Eloundou paper evaluates exposure<\/em>, which means the share of tasks for which an LLM (or LLM-augmented software) could plausibly reduce the time to complete the task by at least 50%. That is a ceiling. The actual share of workers in 2026 with an AI workflow replacing part of their work depends on whether their employer bought a subscription, whether the worker has been trained on it, and whether the use case is auditable. Most workers in most occupations are still in the “have heard of it” stage.<\/p>\n\n The Anthropic Economic Index, which has the rare advantage of measuring observed production usage instead of exposure estimates, puts the picture more conservatively. As of the June 26, 2026 update<\/a>, roughly 36% of occupations show AI usage in at least 25% of their tasks<\/strong> in the Claude API logs the index draws from. That is a much smaller number than the Eloundou 80%, and the gap is the difference between theoretical exposure and observed production use. For the Anthropic Economic Index occupation-level data<\/a>, both numbers shape the conversation.<\/p> Read the Eloundou number the other way and the doomer narrative gets uncomfortable fast: 81% of workers could see less than half of their tasks changed<\/strong>. The universal-exposure talking point obscures this. Roughly 8 of 10 workers, even on a maximalist reading of LLM capability, would still spend the majority of their day on tasks that aren’t automated. That ratio \u2014 81\/19, with the 19 exposed at risk \u2014 is not the labor-market-of-the-Age-of-AI shape.<\/p>\n\n And the more recent follow-on work bears this out. O*NET’s occupational taxonomy<\/a>, cross-walked against Anthropic’s occupation-level usage data, shows that the high-exposure occupations cluster around five task profiles: text drafting, summarization, classification, code generation, and structured data extraction. Outside of those five profiles, real production adoption is still small. The Klarna customer-service case is dramatic in part because customer-service is uniquely concentrated in those five profiles, not because customer-service is representative of the labor market.<\/p>\n\n The doomer narrative tends to point at any headline automation story \u2014 Klarna, IBM, Duolingo’s 2023 contractor cuts \u2014 and generalize from it. our agent landscape analysis<\/a> That framing fits the data, but only on a 19% slice.<\/p> The most consequential real-world productivity data on this question comes from Brynjolfsson, Li, and Raymond’s 2023 paper, “Generative AI at Work”<\/a>, published in the Quarterly Journal of Economics<\/em> in 2025. The study looked at 5,172 customer-support agents at a Fortune 500 software firm and measured what happened when AI assistance was rolled out incrementally.<\/p>\n\n Average productivity rose 15%, but the gain was highly bimodal.<\/strong> The lowest-experience, lowest-skilled agents saw the largest improvement \u2014 they got faster and<\/em> better. The highest-experience, highest-skilled agents saw small gains in resolution speed and small declines in quality. Translation: AI assistance pulls the floor up faster than it raises the ceiling. New hires who used to ramp in three to six months started performing like six-month veterans within weeks.<\/p>\n\n This is the empirical result that best explains the labor-market paradox of 2026. When productivity jumps the most for new hires, firms’ wage curves and hiring scales adjust in non-obvious ways. In mature teams, the marginal benefit of AI is small, so headcount stays steady. In growth phases \u2014 and 2025 and 2026 have been growth phases for most knowledge-work firms \u2014 firms can expand output without expanding hires by giving new employees AI. That difference looks like AI-helped productivity on the BLS productivity stats and looks like AI-replacing-jobs on the firm-level announcements. Same data, different slicing. our Claude-vs-GPT-5 code review benchmark<\/a><\/p> Anthropic’s economic data is the cleanest cut here. In its August 2025 release and confirmed in the June 26, 2026 update, the Anthropic Economic Index classified Claude usage by intent. They distinguish between augmentation<\/em> (using AI in the loop of a human task \u2014 drafting, summarizing, code autocomplete) and automation<\/em> (delegating a whole task to the model \u2014 auto-replying to a ticket, generating a script).<\/p>\n\n The split is roughly 4 to 1 in favor of augmentation as of the June 2026 refresh: for every 100 turns of Claude usage, about 80 are collaborative with a human in the loop and about 20 are run end-to-end. This is a usage pattern, not a projection of where the market will land in 2030, but it’s the best empirical read on what AI actually does<\/em> in 2026 inside the firm context.<\/p>\n\n Two caveats. One, the Anthropic dataset only covers Claude users \u2014 a self-selected group of firms and roles that chose Claude for some reason. Two, the augmentation\/automation ratio has been trending toward more automation with each release cycle, but the ratio is still 4:1 in 2026. The argument that “all AI is automation in disguise” is rhetorically convenient but currently wrong on observed production data. The argument that “AI is going to stay augmentation forever” is also wrong \u2014 automation is rising, just not as fast as the augmentation narrative predicts. enterprise automation deployments<\/a><\/p> For the people whose jobs did disappear, the clustering is sharp. Look at the four-year trailing Bureau of Labor Statistics occupational projections plus the data-drop from Challenger, Gray & Christmas’s monthly layoff tracker, plus the 1-in-4 employer survey from ResumeBuilder in early 2024:<\/p>\n\n Outside of those four clusters, the BLS occupational data do not show a meaningful AI-driven headcount contraction. Lawyer-headcount is up. Accountant-headcount is steady. Healthcare job growth is the leading source of net new employment. Trades and services are labor-constrained rather than AI-substitutable.<\/p> The argument that AI will create more jobs than it destroys has been the default optimistic position for three years. The argument that AI will destroy more jobs than it creates has been the default pessimistic position for two years. Both are projections, and projections on five-year horizons in 2026 are essentially prophecy.<\/p>\n\n What is<\/em> observable now is whether AI is showing up as net-positive or net-negative at the macroeconomic level. The OECD’s 2024-2026 employment outlook, the U.S. BLS payroll data for 2024-2026, and Stanford HAI’s 2026 AI Index<\/a> on the Economy chapter all land on the same answer: in aggregate, AI deployment correlates with low or falling unemployment in 2026<\/strong>, and with productivity gains that have not yet been priced into wage data. This is the cleanest single empirical fact in the debate.<\/p>\n\n That said, the aggregate could mask distributional disaster. AI’s productivity gains may be concentrated in the firms and sectors that adopt earliest, while the labor-market costs fall on workers in industries where adoption is slower. The 2026 data do not resolve this at the firm-vs-worker level. the AI safety research consensus<\/a> The honest position is “aggregate employment is healthy; distributional effects are unmeasured.”<\/p> The strangest moment in the 2026 labor data is that AI-related hiring is on a tear while AI-related displacement is also on a tear. The Goldman Sachs’s 2023 forecast<\/a> of AI investment approaching $200B globally has been passed \u2014 the actual 2025 number is closer to $250B and Q1 2026 alone hit $110B in global private AI investment. That spending has to go somewhere, and a lot of it goes to engineers, PMs, sales teams, and MLOps staff building, deploying, and integrating these models. the $297B Q1 2026 funding boom<\/a><\/p>\n\n That accounts for the visible “AI labor market.” It does not account for the layoffs it funds. The Anthropic Economic Index, the Stanford 2026 chapter, and our own analysis of the BLS occupational data all point to the same takeaway: the jobs AI is creating are tech\/engineering, sales, and ops adjacent to model deployment; the jobs AI is displacing are the high-volume customer-service, writing, and tier-1 white-collar roles that have historically been the on-ramp from education to professional work for non-CS graduates. The transition is widening the gap. That is what the macro labor statistics hide.<\/p> Two practical takeaways for readers in 2026, in plain language.<\/p>\n\n If you are a hiring manager or business owner:<\/strong> the AI skills gap is upstream, not downstream. You will hire fewer customer-service reps and tier-1 engineers than you would have absent LLMs, but you will pay higher salaries for the ones you do hire. The trained-out-of-AI junior candidate is a real category now and they are scarce. The market is paying for orchestration skills over task-doing skills, which is what the Brynjolfsson 2023 study<\/a> showed on the small scale. Plan around that, not around the headline layoff count.<\/p>\n\n If you are a worker in a high-exposure role:<\/strong> the floor is rising faster than the ceiling. Waiters are hard to replace. Junior copy is. Tier-1 software engineering still beats a frontier model for non-trivial PRs. Augment yourself<\/em> \u2014 learn the workflow that uses the model from end-to-end of a unit of work, not the workflow that does part of your job and complains about the rest. The pricing math, per our AI inference cost analysis<\/a>, still favors high-context human + AI over AI alone for an expanding class of problems.<\/p>\n\n The doomer \/ booster debate was never the right frame. The right frame is “task substitution is happening at speed, aggregate employment is roughly stable, the distributional costs are real and concentrated.” The 2026 research lands in exactly that place.<\/p> No, not in the form the doomers describe. The peer-reviewed exposure studies (Eloundou 2023, Anthropic Economic Index 2026) put the share of workers whose tasks are at least half-automatable at under 25%, and the share of occupations showing meaningful AI use is closer to 36% than to the 95% range the headline writers claim.<\/p>\n Customer-support agents, junior copywriters, entry-level transcriptionists, basic data-entry clerks, and tier-1 software coding roles. These are the early observable clusters \u2014 sectors where tasks are repeatable, language-mediated, and don’t require a license or physical presence.<\/p>\n No. The Goldman Sachs and McKinsey baseline scenarios project net employment growth through 2030, but with very different sectoral composition. The displacement narrative is true at the occupation level (some roles shrink); the unemployment narrative is false at the aggregate level (the labor market is near full employment in most OECD countries in 2026).<\/p>\n Move from task-doing to task-orchestrating. The Brynjolfsson 2023 customer-support study and the Anthropic Economic Index both find that the productivity premium from using AI shifts to workers who supervise AI outputs, not to workers who compete with them. Forklift certified drivers spend less time driving. The skill gap is moving upstream, not disappearing.<\/p>\n Partly. Anthropic’s 2025 release showed augmentation at roughly 4x automation in Claude usage patterns, but the same dataset shows automation rising steadily as workflow integrations mature. The composition shifts; the augmentation-as-dominant framing is data-supported for now but not eternal.<\/p>\n\n\n The empirical claims in this post are drawn from peer-reviewed working papers (Eloundou et al. 2023; Brynjolfsson et al. 2023), three years of running industry data (Anthropic Economic Index Aug 2025 \/ Jun 2026), and macro labor-market reports (Stanford HAI AI Index 2026, OECD Employment Outlook 2024-2026, U.S. BLS payroll data, Goldman Sachs Global Investment Research 2023 forecast update). Forecast claims are labeled as such. Worker-level distributional effects are flagged as unmeasured where the data do not yet support measurement. This synthesis was last updated August 22, 2026.<\/p>\n\n\n\nThe two camps are both wrong<\/h2>\n\n
How much of the economy AI can already touch<\/h2>\n\n
The 80\/19 split that broke the doomer narrative<\/h2>\n\n
Why the productivity gain looks bimodal<\/h2>\n\n
Augmentation vs automation: what workers actually do<\/h2>\n\n
Where the jobs disappeared first<\/h2>\n\n
\n
The “AI kills jobs \/ AI creates jobs” debate has been settled by data<\/h2>\n\n
The hiring-vs-displacement paradox<\/h2>\n\n
What this means for you, depending on which side you are on<\/h2>\n\n
Will AI replace most jobs by the end of the decade?<\/h3>\n
Which jobs are actually being displaced right now?<\/h3>\n
Is the AI economist consensus that overall employment will fall?<\/h3>\n
What should I do if my job is at high exposure?<\/h3>\n
Is “augmentation” just a cope for automation that hasn’t happened yet?<\/h3>\n
Methodology and citations<\/h2>\n