The question of exactly how many jobs Large Language Models (LLMs) have taken has long been an unsolved mystery causing deep anxiety among millions of workers.
Anthropic has just released research based on real usage data, clearly outlining the true situation of the current labor market facing AI.
Researchers moved beyond empty theoretical predictions, cross-referencing the US government's occupational database with real usage records of large models to accurately calculate the precise proportion of each occupation being replaced in reality.
The latest data reveals that workers with high exposure to machines are generally older, possess higher education levels, and earn higher salaries.
Although a widespread wave of unemployment, which the public constantly fears, has not yet arrived, the hiring rate for young people aged 22 to 25 in high-exposure occupations has already shown a visible decline.
Real Usage Data Redefines Unemployment Risk
Past research predicting the impact of technology on the labor market has always lacked sufficient foresight and accuracy.
Once, authoritative scholars claimed that 25% of US jobs were highly susceptible to offshoring shocks; a decade later, those positions remain vibrant.
Official predictions from government agencies also relied solely on extrapolating past linear trends, making it extremely difficult to capture sudden technological upheavals.
There is still fierce debate in academia regarding how much real unemployment industrial robots or international trade have actually brought about.
To avoid repeating past mistakes, researchers built a new measurement framework called "Observed Exposure."
This tool is specifically designed to accurately capture real-world technological substitution behaviors in the workplace and issue early warnings for high-risk occupations.
It acts like a high-precision radar, completely filtering out false technological scares and only recording actual industry changes that have taken root.
The new framework cleverly integrates core datasets from three dimensions.
It deeply integrates the O*NET (Occupational Information Network) database, which covers specific tasks for approximately 800 unique US occupations, and introduces Anthropic's own economic index, which records millions of real interactions.
Researchers simultaneously adopted the task exposure estimation standards established by academia in 2023, used to measure whether Large Language Models can theoretically double the speed of a specific task.
Previous research assigned clear scores to various workplace tasks: tasks that could be fully taken over by LLMs were scored as 1, those requiring additional software tools were scored as 0.5, and those with no substitution method were scored as 0.
What is theoretically possible and what is actually deployed in business processes are two completely different things.
Many jobs are theoretically capable of being performed by machines, but due to strict legal constraints, rigid software thresholds, or mandatory human review steps, their actual application is minimal.
Taking the authorization of pharmacy prescription refills as an example, past theories classified it as fully capable of being accelerated by AI, yet real usage records show no trace of it.
To make the measurement results extremely close to the real workplace, researchers meticulously compiled real Claude usage statistics and created charts.
Figure 1 below shows the distribution of Claude usage volume against theoretical task exposure.
It clearly presents the distribution of actual usage based on theoretical predictions.
Those high-risk tasks that LLMs can theoretically handle independently account for as much as 68% of real usage. Tasks deemed theoretically infeasible account for a pitiful 3% share.
Theoretical feasibility is highly positively correlated with actual usage volume.
97% of real business applications fall accurately within the theoretically feasible scope, confirming that past academic judgments were largely correct.
To intuitively present the gap between real-world application and theoretical limits, researchers created a comparative radar chart by occupational category, like taking an X-ray to probe the depth of machine penetration in various industries.
Figure 2 below shows the real comparison between theoretical capabilities and observed exposure across occupational categories.
The blue area representing the theoretical upper limit is exceptionally full for most occupations.
Computer and mathematical categories have 94% of tasks with theoretical penetration space, and office and administrative categories also have as high as 90% of tasks easily accelerated by AI.
The actual red coverage is far from reaching the theoretical limit; the current actual penetration rate in the computer and mathematical field is only 33%.
With technological iteration and commercial popularization, the blank area in the middle that has not yet been filled will inevitably be gradually encroached upon.
Physical labor in agriculture, such as pruning trees, or hard-core legal work like representing clients in court, still firmly remains outside the machine's capabilities.
Researchers ranked the highest exposure occupations based on multiple indicators, identifying the core tasks currently being replaced by automation.
The table below details the top ten most impacted occupations and the specific work content being largely taken over by AI.
The daily tasks of computer programmers, who rank at the top, are covered by a ratio as high as 74.5%.
This perfectly aligns with the current software development scenario, where massive code writing and routine testing work have long been handed over to large models for batch processing.
Customer service representatives and data entry clerks, who follow behind, are in equally dangerous situations; complex source file reading and system entry are being handed over to machines at an unprecedented speed.
Approximately 30% of workers still possess absolute job security.
Their exposure in the observation database is zero, and related tasks appear with extremely low frequency in real interaction records.
Chefs, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing room attendants all belong to the absolutely safe camp; daily pure manual labor has no risk of being replaced by virtual code.
Real Profile and Future of High-Exposure Occupations
Having identified who stands at the center of the storm, it is necessary to further clarify the macro trends of high-exposure occupations in the labor market.
The BLS (US Bureau of Labor Statistics) conventionally released employment projections for 2024 to 2034, and researchers rigorously compared these official expectations with the newly calculated real exposure levels.
The calculation results show a very weak negative correlation.
For every 10 percentage point increase in exposure, the official ten-year occupational growth expectation decreases by 0.6 percentage points accordingly.
The expectations given by analysts within the system, based on their professional judgment, are basically in sync with the conclusions deduced from real usage records.
The scatter plot in Figure 4 below very intuitively confirms this trend.
Each dot in the chart represents the average expected employment growth for occupations under different exposure levels.
As exposure climbs to the right, the overall employment growth expectation shows a slow downward trajectory.
Cashier positions, which are already shrinking, are located in the bottom left, while programmers and customer service representatives are deeply trapped in the high-risk and low-expectation zone on the right.
Researchers retrieved detailed data from the CPS (Current Population Survey) for the three months prior to the launch of ChatGPT to deeply characterize the real profile of the affected population.
They extracted workers in the top 25% of exposure and those with zero exposure risk to complete an extremely detailed demographic comparison.
The high-exposure group and the safe group are like coming from two completely different parallel worlds.
In the high-exposure group, the proportion of women is 15.5 percentage points higher, the proportion of whites is 10.6 percentage points higher, and the proportion of Asians is almost double.
They generally possess more decent incomes and higher educational backgrounds; their average hourly wage is 47% higher than the safe group, and they rarely join labor unions.
We can see the more micro-level underlying differences clearly from the demographic survey comparison table below.
In the non-exposed group, those with graduate degrees account for only 4.5%, while in the highest exposure group, this proportion of high-educated elites surges to 17.4%.
This nearly four-fold disparity completely breaks the objective law that previous mechanical revolutions targeted low-educated blue-collar workers; instead, high-paid knowledge workers have become the hunting targets precisely locked in by algorithms.
With a precise profile, researchers focused their final attention on the most ruthless unemployment rate indicator in the workplace.
When a person loses a job and urgently needs to find the next one, it most intuitively reflects the economic trauma brought by sudden technological changes.
Researchers used historical feedback from population surveys to conduct extremely strict unemployment tracking on the highest-risk groups.
Figure 6 below fully displays the historical trajectory of unemployment rate evolution for the two groups starting from 2016.
The red line in the upper part marks the high-exposure group, and the blue line represents the non-exposure group.
During the COVID-19 pandemic, the non-exposure group, engaged in offline physical services and insulated from new technologies, suffered huge unemployment pain, with the unemployment rate soaring cliff-like.
The difference-in-differences model in the lower part perfectly confirms the smooth convergence of the two groups in the post-pandemic era.
Since ChatGPT emerged at the end of 2022, the unemployment rate gap between the high-exposure group and the safe group has shown no waves.
The average change in the two sets of data is at a statistically negligible level; the unemployment rate in high-risk positions has only caused a tiny ripple that cannot be confirmed as caused by technology.
Workers can finally temporarily put their hearts at ease.
The so-called white-collar recession, spoken of with certainty, has not happened at all; during the Great Recession from 2007 to 2009, the overall unemployment rate doubled from 5% to 10%.
Even if the unemployment rate of the highest exposure group only increased from 3% to 6%, an extremely sensitive difference-in-differences model could instantly capture the waveform oscillation.
The current coverage of large models is only slowly seeping into some links, far from directly taking away workers' entire livelihoods.
Young People's Job Search Dilemma in High-Exposure Occupations
While the overall unemployment pie looks as steady as Mount Tai, the underlying currents for specific groups are changing direction rapidly.
Young people play the role of extremely sensitive canaries in the extremely involutionary labor market, often being the first to sense the thinness of industry oxygen.
Relevant academic reports found that the employment rate of young people aged 22 to 25 in occupations with extremely high exposure has shown a significant decline of 6% to 16%.
Scholars directly point the finger at the culprit of the decline being a significant shrinkage in corporate recruitment, rather than simple and crude direct layoffs.
Many newcomers who have just stepped into society have not yet had time to hold a fixed job title in official statistics; ordinary unemployment rate indicators are difficult to see the deterioration of their situation.
Newcomers who cannot find jobs often choose to continue their studies or quietly withdraw from the labor force, thereby becoming invisible people in statistics.
Researchers utilized the unique panel tracking attributes of the CPS (Current Population Survey) to directly calculate the exact probability of young people receiving new job offers.
They calculated month by month how many young people aged 22 to 25 successfully started a new high-exposure or low-exposure job they did not have before.
Figure 7 below depicts the real job search situation of young people, which is like ice and fire, with a very clear trend.
The upper part intuitively presents the monthly hiring rate changes for young people in two types of positions, while the model calculation in the lower part accurately measures the cruel gap between the two lines.
Completely eliminating the extreme oscillation interference caused by the pandemic in previous years, these two lines symbolizing vitality moved towards an irreconcilable divergence in 2024.
Safe positions insulated from the technological storm have extremely strong absorption rates for young people, stably accepting 2% of fresh blood every month.
The door to high-exposure positions is slowly closing on young job seekers; the overall hiring rate has dropped by about 0.5 percentage points.
Entering the post-ChatGPT era, the successful hiring rate for high-exposure positions accepting young people has dropped by a full 14% compared to 2022.
This job search resistance only works on junior workplace newcomers under 25 years old and has no killing power on veterans who are over 25 and have worked in the workplace for many years.
After enterprises introduce APIs to fully take over junior programming, customer inquiries, or basic data processing, they indeed no longer need to recruit interns in large quantities to share the hard labor.
Automation substitution was not achieved by sending layoff letters to old employees; the vast majority was achieved by silently freezing recruitment, directly cutting off the entry qualification for a new generation of job seekers.
Those young people shut out of high-paying industries may eventually only be able to wander around marginal positions.
Extremely calm data has sounded a warning bell for everyone: machines have not immediately robbed veterans of their jobs, but are silently swallowing the growth stepping stones of workplace newcomers.
Enterprises used AI tools to achieve human-machine collaboration, raising the upper limit of output for existing teams, and naturally conveniently cut off the channel for newcomers to enter.
References:
https://www.anthropic.com/research/labor-market-impacts
END