Junior AI hiring: read the denominator before changing your career plan

Two percentages that answer different questions

Anthropic's October 16 report says the junior-level share of employees in its technical departments slipped from 21% to 14% since early 2025. It separately says the proportion of junior technical hires fell from 27% to 9%. The first comparison concerns people already employed. The second concerns people entering through hiring. Those changes are 7 and 18 percentage points respectively. Neither percentage is a count of junior openings available to you.^[1]

The company also reports that technical hiring continued growing in absolute terms as sales, finance and recruiting hiring grew faster. That statement supplies useful context about the technical total. It leaves a specific question unanswered: did junior technical hires rise, remain level or fall over comparable periods? A smaller junior fraction can coexist with any of those outcomes, depending on how much the total changed.^[1]

Try a prediction before doing the arithmetic. In this hypothetical example, an employer makes 100 technical hires in one period, and 27% are juniors. In a comparable later period, total technical hires double to 200, but the junior share falls to 9%. Would the junior count rise, stay the same or fall? Choose an answer and estimate the later junior count before reading the next paragraph. These totals are invented for the exercise, not reported Anthropic counts.^[1]

The junior count falls: 27% of 100 is 27, while 9% of 200 is 18. Doubling total hiring is insufficient to offset this drop in share. At 9%, the employer would need 300 total technical hires to hire 27 juniors again, three times the earlier total. The reusable check is junior share multiplied by total technical hires for the same period. Without those period-matched totals, the reported shares cannot establish Anthropic's junior counts.^[1]

A second source has a different denominator

Indeed Hiring Lab's September 17 analysis provides an independent comparison with a narrower population than its headline might suggest. It examines US postings that advertise annual salaries. Within the most AI-exposed occupations in that sample, the entry-level share fell from 29% to 10% between 2021 and 2026, while the senior share rose from 22% to 47%. These are shares of advertised postings in the analyzed population, rather than completed hires at an employer.^[2]

The restriction matters. Hiring Lab says its salary-advertising population skews more senior than all US postings, and explicitly warns that the seniority shares are not comparable with its published all-postings shares. A posting sample describes advertised opportunities and their composition. It cannot tell you how many junior candidates actually started work. Combining its percentages with Anthropic's hiring percentages would combine different populations, measures and time periods.^[1][2]

Keep the pay result separate too. Hiring Lab finds faster advertised-pay growth in more AI-exposed work, with a larger cumulative gap at senior levels. It calls the seniority split suggestive because the underlying regression terms are mostly not statistically significant. Advertised salaries are different from accepted offers and realized earnings. This older US analysis is context for interpreting composition, rather than evidence of what a Canadian graduate will earn or how many Canadian junior jobs exist.^[2]

Write a decision record before changing your search

For a Toronto graduate deciding whether to stop applying to junior technical roles, use seven fields: measured event; population; time; numerator; denominator; unknown; action. Here is a filled example. Measured event: completed hiring, not vacancies. Population: Anthropic's technical hires, not Canada's labor market. Time: the report's comparison since early 2025. Numerator: junior technical hires, whose share fell from 27% to 9%. Denominator: all technical hires in each comparison period. Unknown: comparable absolute totals and Canadian junior hiring counts. Action: keep considering junior roles and check each vacancy's stated experience requirements; these shares alone do not justify abandoning that search. The action is our proposed decision, not a tested finding.^[1]

Reuse the same fields for Hiring Lab, changing the measured event to an advertised posting, the population to US postings advertising annual salaries, and the time comparison to 2021 versus 2026. For the most AI-exposed occupations, the entry-level numerator is a subset of that restricted posting denominator. Both sources show a shift toward more senior composition within their own populations. Neither measures Canadian junior hiring counts.^[1][2]

Make the action field specific enough to revise. For example, reconsider a particular vacancy when its employer requires experience you cannot demonstrate. Save that requirement alongside your decision. This gives you an inspectable reason to change a target, rather than using an aggregate share to decide an entire career change.^[1][2]

Use the company findings to ask a narrower skills question

There is a useful skills question in Anthropic's findings. Its analysis of hiring-manager assessments finds increasing emphasis on technical applicants' autonomy and judgment, alongside less discussion of domain expertise as a share of assessment. That is evidence about this company's recorded assessments. It suggests asking how you would demonstrate a decision, identify a limitation and explain an escalation in your own work. It does not establish a universal graduate selection rubric.^[1]

A practical addition to your decision record is one concrete example: the task you attempted, the AI output you checked, the error or uncertainty you found, and the reason you accepted, revised or rejected the result. This is our proposed preparation exercise, informed by the report's emphasis on judgment. Keep the description limited to work you actually did. Neither study tests whether adding such an example improves interview or offer rates.^[1][2]

Anthropic combines internal records with 17 employee interviews and cautions that adoption of its AI tools was not random. Its unusually intensive use of AI also limits generalization. Your completed audit should therefore preserve three separate answers: what changed in the measured population, what remains unknown about your target market, and what specific evidence would change your next application decision. An unknown denominator is a reason to narrow a conclusion before making that decision.^[1]