Summary (TL;DR)
Anthropic’s 2026 study introduces a new metric — observed exposure — for assessing the impact of artificial intelligence on the labor market. The main conclusion: despite a theoretically high automation potential, no real increase in unemployment among vulnerable workers has been recorded. However, hiring of young people aged 22–25 in highly exposed occupations has slowed by about 14%.
Contents
- What observed exposure is
- Which professions are most vulnerable
- Profile of a vulnerable worker
- What is happening with unemployment
- Risks for young people
- Bureau of Labor Statistics forecasts
- What this means for you
- FAQ
What observed exposure is and why it matters
The impact of AI on the labor market is one of the main topics in economic discussions in recent years. Forecasts range from “AI will replace half of all jobs” to “there is no real threat.” In March 2026, Anthropic published a study that proposes a new, more accurate way to measure this risk.
The study’s authors — economists Maxim Massenkoff and Peter McCrory — introduce the metric “observed exposure” (observed exposure). It differs fundamentally from previous approaches because it combines two data sources:
- Theoretical automation potential — estimates from Eloundou et al. (2023), describing which tasks language AI could theoretically speed up by 2x
- Real-world usage — data on how people actually use Claude in professional contexts (Anthropic Economic Index)
Key finding: AI is far from realizing its theoretical potential. For example, in occupations in the “Computer and Mathematics” category, language models can theoretically perform 94% of tasks, but in practice Claude covers only 33%. This means there is a huge gap between what AI can do and what it actually does.
A profession’s observed exposure is higher if:
- Tasks are theoretically feasible with AI
- These tasks are actively carried out through Claude in work contexts
- A significant share of usage is automated (rather than assistive)
- AI-dependent tasks account for a large share of working time
Which professions are most vulnerable to AI
According to the study, the top 3 most exposed occupations are:
- Programmers — 75% task coverage. This is consistent with data showing that Claude is actively used for writing and reviewing code.
- Customer service specialists — a significant portion of their tasks is increasingly performed via API in automated workflows.
- Data entry operators — 67% coverage. Their main task — reading documents and transferring data — is being actively automated.
Among other highly exposed occupations are financial analysts, technical writers, accountants, and information processing specialists.
At the same time, about 30% of workers have zero exposure — their tasks appear in the data too rarely. This group includes cooks, mechanics, rescue workers, bartenders, and dishwashers. Physical labor and work that requires a live presence remain outside the risk zone for AI automation.
Profile of a vulnerable worker: who AI will affect most
One of the study’s unexpected findings is that the demographic profile of workers in the most vulnerable professions differs sharply from that of workers whose professions have zero risk. The comparison was made using data from before the launch of ChatGPT (August–October 2022):
- +16 percentage points — likelihood of being female
- +11 percentage points — likelihood of being white
- 2x more likely — Asian background
- +47% — average wage level
- 4x higher — share of people with an advanced degree (17.4% vs 4.5%)
In other words, AI threatens above all highly paid, educated professionals, not low-skilled workers, as is commonly believed. This radically changes the usual narrative about automation.
What is actually happening with unemployment
This is perhaps the most important question. And the researchers’ answer is unexpectedly reassuring: no systematic increase in unemployment has been recorded among the most exposed workers.
An analysis of data from the U.S. Current Population Survey shows that since the release of ChatGPT in late 2022, the unemployment rate in highly exposed occupations has changed little relative to less exposed occupations. The gap between the groups is statistically insignificant.
The researchers emphasize that their methodology can detect effects on the scale of the “Great Recession” — when unemployment doubled from 5% to 10%. If something like that were happening in highly exposed occupations, it would be noticeable. But that is not observed.
Important caveat: the authors measure differential changes — that is, they compare the vulnerable group with a control group. If unemployment rises for all workers at the same time (for example, due to an economic downturn), this will not be attributed to AI.
Warning sign: risks for young people
The only area where the study found a statistically significant (though preliminary) signal is hiring of young professionals aged 22–25.
The data show that starting in 2024, young workers have been less likely to find jobs in highly exposed occupations. If the monthly employment rate in less vulnerable occupations remains stable at 2%, then in the most vulnerable occupations it has fallen by about 0.5 percentage points.
On average over the period since ChatGPT’s launch, this is a 14% decline compared with 2022. Among workers older than 25, no similar decline is observed.
The researchers caution against overinterpreting this: young people who are not hired may stay in their current jobs, move to other industries, or return to education. In addition, survey data on job changes may contain measurement errors.
Nevertheless, this result echoes the findings of another study (Brynjolfsson et al., 2025), which identified a 6–16% drop in employment in vulnerable occupations among workers aged 22–25.
How Bureau of Labor Statistics forecasts confirm the metric
The U.S. Bureau of Labor Statistics (BLS) regularly publishes employment forecasts. In 2025, it released a forecast for 2024–2034. The researchers compared it with the observed exposure metric.
Result: every 10 percentage-point increase in exposure reduces projected employment growth by 0.6 percentage points. The correlation is weak but significant — and it is an important confirmation that the new metric really reflects real economic trends.
Notably, the theoretical Eloundou metric (without data on real usage) does not show such a correlation. This suggests that it is precisely the data on real-world AI use that adds predictive power.
What this means for workers, employers, and policymakers
For workers in highly exposed professions
Good news: right now, your job is probably not at risk. Bad news: the long-term employment outlook in your profession may worsen. Especially if you are just starting your career.
Practical recommendations:
- Develop skills that complement AI rather than compete with it
- Focus on tasks that require judgment, empathy, and physical presence
- Monitor how AI is changing your industry and adapt ahead of time
For employers and HR professionals
Data on «observed exposure» is a valuable tool for strategic workforce planning. Occupations with high coverage are candidates for restructuring, but not necessarily for downsizing: many tasks are shifting from automation to augmentation.
For policymakers
The study creates a basis for monitoring: if youth hiring continues to decline, this will become an early signal for policy interventions — retraining programs, income support, educational reforms.
FAQ: frequently asked questions about the impact of AI on the labor market
Will AI replace my job?
According to Anthropic's study, there is as yet no systematic wave of layoffs in highly exposed professions. The risk exists in the long term, especially for young professionals entering the labor market. Full automation of occupations takes time and depends on technological development, the regulatory environment, and economic conditions.
Which professions are safest from AI?
About 30% of workers are in professions with zero exposure to AI. These are typically manual labor and work requiring a live presence: cooks, mechanics, rescue workers, healthcare workers, construction workers. Tasks that require fine motor skills, social interaction, and the physical world remain beyond the reach of modern AI systems.
Why doesn't AI reach its theoretical potential?
The gap between what AI can theoretically do and what it actually does is explained by several factors: model limitations for specific tasks, legal restrictions (for example, issuing prescriptions), the need for human verification, high integration requirements, and slow organizational adoption.
Why are young professionals at risk in particular?
According to the researchers, companies may be holding back hiring for entry-level positions — the very ones previously performed by young professionals — by replacing them with AI tools. At the same time, experienced employees keep their jobs because their role is more complex and requires judgment. This effect is still preliminary and requires further confirmation.
How is AI's impact on the labor market measured in this study?
The researchers use three sources: the O*NET database (task descriptions for ~800 occupations in the US), data on real-world use of Claude from the Anthropic Economic Index, and theoretical exposure estimates from Eloundou et al. (2023). Combining these data produces a metric of «observed exposure» that takes into account both the possibility and the reality of automation.
Conclusion: AI and the labor market — evolution, not revolution (for now)
AI's impact on the labor market is real, but so far not catastrophic. Anthropic's study shows that despite the enormous theoretical potential for automation, the real penetration of AI into workers' tasks is much more modest, and there is no systematic wave of unemployment.
Nevertheless, the signals are there: long-term employment forecasts are worse for highly exposed professions, and youth hiring in these professions is slowing. This requires attention — from workers themselves, employers, and the state.
The value of this study lies not only in its results, but also in its methodology. The authors created a basis for regular monitoring: as AI spreads, this same metric will make it possible to track when and where real labor disruptions begin — before they become obvious.
Follow updates to the Anthropic Economic Index — the company plans to update this data regularly, creating a live picture of how AI is changing the labor market in real time.