AI and Labor / The Gap Between Theoretical Capability and Implementation / Thoughts on the Slowdown in Youth Hiring

I. Not what AI can do, but where it is infiltrating work

The first thing I feel when reading Anthropic's new report is that, before considering the magnitude of unemployment rates, there is value in the fact that they have reorganized the units of measurement. Previous studies have relied on estimates of capability, specifically how fast large language models can theoretically perform certain tasks. A representative example is the study by Eloundou et al. (2023), which evaluates whether an LLM can increase processing speed by more than double for tasks in each occupation using a beta value from 0 to 1 [Note 1].

In contrast, Massenkoff and McCrory (2026) overlay actual Claude usage data onto theoretical possibilities, and by weighting automated use more heavily than auxiliary use, and professional use more heavily than personal use, they establish an indicator called observed exposure [Note 2]. There is a very strong implication in the fact that they have brought the scope of the discussion back from general theories of capability to implemented forms of usage. What matters for job loss is not cases where AI assists human judgment, but cases where automated use that directly replaces human labor is progressing, and the substantive significance of this indicator is that it incorporates that distinction into its measurement.

1. Maps of capability and traces of implementation

The implication of this indicator is clear. If one discusses the labor market by looking only at a map of AI capabilities, one will overestimate the speed of change. In the report, while the theoretical coverage rate reaches 94% for computer and mathematical occupations, the observed coverage rate remains at only 33% [Note 2]. The occupations listed as having high exposure are those involving digitized descriptive work or routine judgment, such as computer programmers (74.5%), customer service representatives (70.1%), data entry clerks (67.1%), and medical records specialists (66.7%) [Note 2]. Legal professions also have high theoretical potential, but the act of representing a client in court remains outside the scope of implementation.

From the perspective of legal policy, this distance between capability and implementation is the point of contention. Regulations, liabilities, and employment adjustments must respond not to abstract capabilities, but to implemented forms of usage.

2. Measurement is both the collection of facts and a selection

However, observed exposure is not a neutral recording device. Judgments such as what degree of usage is considered significant, how much more heavily automation is weighted than assistance, and how similar tasks are bundled together all influence the shape of the index. Anthropic itself frankly acknowledges this point, and explains that even if the settings are changed, the correlation of rankings by occupation remains extremely high [Note 3]. When law and policy use indicators, it is not enough to read only the numbers; one must also read the design philosophy that produced those numbers. It is also necessary to frankly recognize that the structure of conflicts of interest—the structure in which Anthropic researchers analyze the impact of their own products using their own data—is not neutralized by efforts to ensure transparency through data disclosure.

II. Widespread unemployment is not yet visible, but the narrowing of entry points appears first

1. Impact on unemployment rates: Limited at present

The conclusion of this report is quite far from catastrophic theories. Using the Current Population Survey to compare unemployment rate trends since the advent of ChatGPT between workers in the top quartile of exposure and workers with zero exposure, the trends for both groups are generally parallel, and even in estimates using the difference-in-differences method, the change in the gap after the advent of ChatGPT remains at a level indistinguishable from zero statistically [Note 2]. Yale's Budget Lab also states that at this stage, no identifiable employment disruption can be seen in the economy as a whole [Note 4].

It is too early to immediately derive reassurance from this. The current analysis is designed to detect unemployment rate changes of about 1 percentage point or more, but it is possible that changes smaller than that are being missed. Furthermore, it is estimated that if all workers in the top 10% of exposure were laid off, the unemployment rate for the top quartile would reach 43% from 3%, and the aggregate unemployment rate would reach 13% from 4%, which provides context for what the current 'no impact' result indicates.

2. What the slowdown in youth hiring indicates

On the other hand, a different movement is appearing at the entry level. The report observes that the rate at which young people aged 22 to 25 are newly entering high-exposure occupations has fallen by an average of 14% compared to 2022. While statistically on the borderline, Brynjolfsson et al. also report a relative decline in employment in high-exposure occupations for the same age group [Note 5], and when the two are placed side by side, what is happening now appears to be a slowdown in first jobs and youth hiring rather than a wave of layoffs. For companies, adjusting by tightening new hiring first is more natural than letting existing workers go all at once.

The report also explains why the slowdown in hiring does not immediately appear as an increase in the unemployment rate. This is because many young workers are new entrants to the labor market, so if they are not hired, they may exit the labor force rather than being counted as unemployed. There is a segment here that the unemployment rate indicator cannot capture.

3. The gap between the distribution of exposure and the distribution of disadvantage

Furthermore, what is hard to overlook is that the attributes of existing high-exposure workers do not match the group that is affected first at the entry level. According to the report, the high-exposure group has a female ratio about 16 percentage points higher than the zero-exposure group, wages are 47% higher on average, and the ratio of those with graduate degrees is about four times higher [Note 2]. In other words, at present, those most exposed to AI are relatively highly educated, high-income female workers, but it is the younger generation where the employment slowdown appears first. The distribution of exposure and the distribution of disadvantage are misaligned. It is not enough to look only at which occupations are close to AI; one must also look at who enters those occupations and how the entry points there become narrower.

III. Units that legal policy should track

What this document shows is the crudeness of talking about the relationship between AI and the labor market in a single diagram. Theoretical capability, actual implementation, and impact on employment are each at different stages. Compared to the BLS employment projections from 2024 to 2034, there is a weak correlation where for every 10-point increase in observed exposure, the growth projection falls by 0.6 points, but that correspondence does not emerge from theoretical beta values alone [Note 6].

The immediate focus of legal policy lies here. It is not to talk about danger based only on maps of capability, nor to talk about harmlessness based only on unemployment rates, but to separately track how implemented usage spreads to hiring, first jobs, job changes, and training demand. The areas where the law should intervene should also, for the time being, rely on grasping entry routes for young workers, the burden of retraining costs, changes in the value of degrees and qualifications, and transitions that are difficult for unemployment insurance and employment placement services to pick up, rather than on the post-processing of mass layoffs.

It may be time to dismantle the crude question of whether AI will take away jobs. Which jobs, in which aspects, and for whom will the entry points narrow? It is at that level of specificity that we should be looking first.

IV. In Conclusion

The Anthropic report is not merely a document stating that AI has not yet significantly disrupted the labor market. Rather, it is a document that updates how we perceive the units of disruption. It does not conflate high capability with depth of adoption, nor does it conflate depth of adoption with employment damage; instead, it measures the distance between them. Only with this methodology can the discussion of AI and law move away from grand prophecies and closer to the places where changes are actually emerging.

(Magazine) "Random Thoughts on AI and Law"