The debate over artificial intelligence and its effect on employment often feels like a pendulum swinging between utopian promises and dystopian fears. It is a discourse in need of data. A new report from Anthropic, titled “Labor market impacts of AI: A new measure and early evidence” [1] attempts to ground the conversation in reality. The authors have developed a new metric—observed exposure—that moves beyond theoretical speculation by combining AI’s potential capabilities with data on its actual, real-world usage.

The report contains a chart that tells a far more interesting and complex story than a simple narrative of job destruction or for that matter job creation. It visualises the gap between what AI could theoretically do and what it is actually doing in the workplace. This is the gap that educators, policymakers, and trainers must understand and navigate.

“Share of job tasks that LLMs could theoretically perform (blue area) and our own job coverage measure derived from usage data (red area).” Source: Anthropic [1]

In the chart, the expansive blue area represents the share of job tasks that Anthropic say a Large Language Model (LLM) could, in principle, make at least twice as fast. The small red area represents the tasks where AI is actually being used, according to Anthropic’s own data from professional settings. The difference is large. In “Computer & Math” occupations, the theoretical potential covers 94% of tasks, while actual usage covers just 33%. For “Legal” occupations, the gap is between 90% and 20%.

This gap can be interpreted in two diametrically opposed ways. Anthropic’s reading is one of impending transformation. They write: “As capabilities advance, adoption spreads, and deployment deepens, the red area will grow to cover the blue.” From this perspective, the gap is simply a measure of how much of the labour market is on the verge of being disrupted. It is a future waiting to happen.

However, an alternative reading sees the gap not as a sign of AI’s potential, but as a diagnosis of its current limitations. This view suggests the chasm between theoretical performance in a lab and practical competence in the messy and uneven reality of the workplace is significant. The hurdles of integration, regulation, organisational inertia, and the sheer complexity of human workflows are not trivial details to be ironed out; they are formidable barriers. The story, then, is not about the inevitability of the red area filling the blue, but about the immense friction that keeps them apart.

While this debate about future impact is critical, the Anthropic report also provides important data on the here and now. When the researchers looked for immediate effects on employment, they found little to support the most dramatic predictions of mass unemployment. Their analysis based on the US labour market data since late 2022 reveals no systematic increase in unemployment for workers in the most highly exposed occupations. The wave of AI-driven job losses has not yet materialised in the aggregate data.

However, the report does uncover a more subtle and potentially leading indicator. It finds “suggestive evidence that hiring of younger workers has slowed in exposed occupations” [1]. Specifically, the rate of new job starts for workers aged 22-25 in highly exposed fields has seen a statistically significant drop. This finding is also reflected in other recent research and points to an emerging dynamic: AI’s initial impact may not be felt in mass layoffs, but in the disappearance of entry-level opportunities. Companies may be using AI to automate or augment the tasks previously assigned to junior employees, thereby reducing the need to hire new entrants into the workforce.

Implications for Vocational Education and Training

For the vocational education and training (VET) community, these findings provide a challenge. The discourse needs to move beyond a binary of job replacement versus job augmentation and engage with what is. Going on in job markets

First, the gap between theoretical and actual AI use is a space of opportunity where VET can operate. The fact that AI is not being seamlessly adopted to its full theoretical potential highlights the importance of human skills in navigating real-world complexity. The challenge is not simply to train for AI-related tasks, but to cultivate human competencies that bridge the gap between the tool and the task. This includes critical thinking, complex problem-solving, collaboration, and the ability to adapt and integrate new technologies into existing workflows. These are the skills that make technology productive in a real-world context.

Second, the finding of a slowdown in hiring for young people is a direct challenge to both the VET system and Higher Education. It suggests that traditional pathways from education to entry-level employment are being disrupted. VET providers can no longer assume that the junior roles of today will exist tomorrow. This necessitates a rethinking of apprenticeships, internships, and work-based learning. The focus must shift from training for specific, narrow job roles to developing broader, more resilient skill sets. It also means VET must play a more active role in designing and supporting new forms of transition into the labour market, potentially through more integrated learning models or by fostering the entrepreneurial skills needed to create new roles.

Finally, the report reminds us that the impact of AI is not uniform. It affects different occupations, age groups, and demographics in different ways. The data shows that highly exposed workers are more likely to be female, older, more educated, and higher-paid. This underscores the need for a targeted and differentiated approach to upskilling and reskilling. A one-size-fits-all approach to AI training will fail. VET must develop programmes that are tailored to the specific needs of different learner groups, addressing not only their technical skill gaps but also the social and psychological dimensions of adapting to a changing work environment.

The Anthropic report does not give easy answers, but more data on which to develop future looking strategies in education and training

References

[1] Massenkoff, M., & McCrory, P. (2026, March 5). Labor market impacts of AI: A new measure and early evidence. Anthropic. https://www.anthropic.com/research/labor-market-impacts

About the Image

This image reflects on how AI systems interpret artworks in ways that differ from human understanding. For these works to become usable input for machine learning models, they must be converted into structured numerical data—a translation in which artistic, cultural, and social value is lost in numbers. This image also comments on the unconsented use of authors’ works in AI training. Here, AI is depicted as a tornado pulling artworks into its centre, evoking a sense of lost agency and control. The artworks featured are vintage magazine covers and advertisements selected from Rawpixel’s public domain collections. All numerical embeddings in this image are illustrative.