A survey saying AI increased hiring and an analysis saying the barrier to employment for young people has not come down appeared in the same week. Anthropic released a labor market impact report on March 5 local time, stating that the probability of job seekers aged 22 to 25 newly entering occupations with high AI exposure has fallen about 14% since 2022. A day later, on March 11, Snowflake announced that in a survey of 2,050 decision makers in 10 countries, 77% of responding firms said hiring rose after they adopted AI. The two figures look at odds, but they are in fact measuring different things.
The core of the Anthropic report is a new indicator called "observed exposure." It is calculated by setting the share of tasks AI can theoretically perform against the share of tasks people are actually assigning to AI. The gap between the two shares is wider than expected. In computer and mathematical occupations, about 94% of tasks can in theory be carried out by large language models, but actual use stays at around 33%.
By occupation, computer programmers ranked highest at 74.5%. They were followed by customer service representatives at 70.1%, data entry clerks at 67.1%, medical records specialists at 66.7%, and market research analysts at 64.8%. At the other end are occupations that use the body, such as cooks, bartenders, motorcycle mechanics, and lifeguards.
The researchers also classified legal work such as courtroom advocacy, along with tree pruning and farm machinery operation, as areas AI still has difficulty handling. Legal constraints, software requirements, and procedures in which people verify the results were cited as factors slowing adoption.
What stands out is that unemployment in highly exposed occupations has not risen in any statistically meaningful way since ChatGPT was released. No large-scale employment shifts were detected either. If, even so, only the odds of young people landing new jobs have fallen, then as the researchers read it, the cause is more likely reduced hiring than layoffs. Companies adjust by not creating positions to fill rather than pushing people out.
The Snowflake survey shows what firms experienced. Some 46% saw a partial reduction in job roles, and among firms that experienced both, 69% assessed AI as positive for overall employment. Organizations that had built up several use cases said at a rate of 75% that they saw a net positive effect on staffing, while organizations in the early stages of adoption came in at 56%. The return per dollar of AI investment averaged 1.49 dollars.
Which roles grew and which shrank matters more. The areas with a net increase in jobs were IT operations at 56%, cybersecurity at 46%, and software development at 38%. The reduction effect was large in customer service and support and in data analysis, at 37% each.
Of the responding firms, 96% said they face obstacles in scaling across the company, and about 80% pointed to technology or data problems as the cause. Responses citing data scattered across departments came to 65%, and quality measurement and monitoring to 62%.
The correlation in which a 10 percentage point rise in AI exposure lowers the projected employment growth rate for the coming decade (2024-2034) by 0.6 percentage points emerged when the figures were matched against U.S. Bureau of Labor Statistics projections. Highly exposed occupations pay an average of 47% more per hour than unexposed ones, and the share of workers with graduate degrees or higher differs roughly fourfold, at 17.4% against 4.5%. That means jobs with good conditions are the first to be shaken. When the first step into those jobs narrows, the steps after it disappear along with it.
The policy clock is already running. The Ministry of Employment and Labor reported the "status of the establishment of a basic plan for employment stability in industrial transition" at the emergency economic ministers' meeting on March 11 and said it would finalize and announce the plan by June. The tasks it laid out include ongoing monitoring by industry, region, and occupation using real-time job posting data, an early warning system for employment crises, job transition consulting and incentives, reinforcement of the safety net including unemployment benefits, and AI skills training across the life cycle. The design presents principles for labor, management, and government to uphold together in place of a five-year plan pushed by the government alone, and divides policy by scenario given that no one knows where the technology will go.
The ministry is also reviewing a fact-finding survey on side effects of the transition, such as bias in hiring and evaluation algorithms, AI ethics guidelines for the labor field, and the institutionalization of the right to disconnect. Intensive discussion through an expert forum is set for April-May.
The scene of technology replacing people has not registered heavily in the statistics. Instead, the openings a 22-year-old can apply to are shrinking bit by bit. Whether the basic plan due in June stops at job transition support for incumbent workers or also covers those trying to enter a workplace for the first time will decide whether the plan succeeds. Filling in the entry-level postings that have disappeared cannot be done by individual effort alone.
