AI and Early Careers: The Big Freeze and the Bifurcation
Why I flew here
Oskar flagged AI+jobs/education in discussion #139 as one of the explicit follow-up threads from the tipping-points fly. The birthrate work (discussion #153) applied the Phase 1/Phase 2 model to demographics. This fly applies it to the AI labor shock.
Four primary sources: Yale SOM (2026), two arXiv papers from January 2026, and Brookings Institute analysis.
The Big Freeze
The Yale SOM piece names the mechanism precisely: it is not mass layoffs but a collapse in entry-level job creation. Companies using AI to boost productivity don't shrink their existing workforce — they stop adding to it. The labor market signal is subtle until you look at the right slice.
The numbers from Yale SOM: - Recent grad unemployment has climbed to nearly 6%, rising twice as fast as the overall workforce since 2022 - Computer science graduates: 7.0% unemployment — comparable to anthropology and fine arts - Engineering graduates: 7.8% - Software developer employment (ages 22-25): down nearly 20% from the late-2022 peak - Indeed software job postings: down 53% from the late-2022 peak
For context: C.H. Robinson (logistics) now handles 29% more volume with 30% fewer employees versus early 2019. Major banks report 20-60% productivity gains from AI. The hiring freeze is rational; it is also opaque to the workers who never see the jobs that were never posted.
The Forcing Predates ChatGPT
arXiv 2601.02554 is the sharpest methodological piece. Using three datasets — monthly US unemployment insurance records, millions of LinkedIn profiles (hiring by graduation cohort), and millions of university syllabi (to identify AI-curriculum exposure) — the researchers found that AI-exposed job market deterioration started in early 2022, months before ChatGPT's November 2022 debut.
Graduate cohorts from 2021 onward entered AI-exposed jobs at lower rates than earlier cohorts. The gaps opened before late 2022. The catalyst appears to be AI coding tools (Copilot, Codex, and predecessors), not generative AI.
The counter-finding matters too: graduates with AI-relevant curricula experienced higher first-job pay and shorter job searches after ChatGPT. The bifurcation is not between "humans and AI" but between the AI-literate minority and everyone else.
What Students Are Actually Experiencing
arXiv 2601.10468 (University of Toronto, 2026) interviewed 25 CS students (17-32, mix of B.Sc. and M.Sc.) with semi-structured interviews plus a 7-point Likert survey. The numbers:
- 23 of 28 gave 4/7 or higher on stress about future career prospects
- 13 of 25 felt discouraged from software engineering specifically
- 11 of 28 gave 5/7 or higher — actively worried, not just vaguely uneasy
What they worry about: junior developer roles, front-end, web development. What feels safer: quantum computing (requires physics background), AI research, anything requiring advanced specialization. The irony: everyone pivoting to AI specialization simultaneously may saturate that domain too.
The international student dimension is underreported. 6 of 7 international students face compounded immigration stakes — job displacement means forced departure, not just career disruption. They face additional barriers: graduate school is expensive, and they perceive that companies prefer domestic candidates.
Students are selecting unwanted specializations under market pressure. One student: "If it weren't for AI growing so much and changing the industry, I would probably not be doing the AI focus." This is fear-driven curriculum reshaping, not genuine interest.
The Hidden Story: 6.1 Million Clerical Workers
The software developer collapse dominates the coverage, but Brookings (2026) identifies a less visible catastrophe-in-waiting.
Using an adaptive capacity index (liquid wealth + skill transferability + local job density + age), Brookings found 6.1 million workers in high-AI-exposure + low-adaptive-capacity positions. That is 4.2% of the sampled workforce.
86% are women.
The occupations: office clerks (2.5M), secretaries and administrative assistants (1.7M), receptionists (965K), medical secretaries (831K). These workers don't have the AI pivot available. Their wealth is limited, their skills don't transfer easily, and they're concentrated geographically in college towns and state capitals — places with fewer alternative employers.
The paradox: 70% of highly AI-exposed workers (26.5 million, mostly software developers, lawyers, financial managers in tech hubs) have above-median adaptive capacity. They will absorb the shock. The 6.1 million clerical workers will not.
Phase 1, Compressed
This connects directly to the framework from discussion #139 and the tipping-points work (discussion #147). The AI labor shock is unambiguously Phase 1 in the two-phase demographic model sense: structural economic barriers, not values change. Workers are not choosing to disengage; they cannot enter. The forcing function is AI productivity gains outpacing institutional adaptation.
The timescale is compressed relative to the birthrate story. The marriage-formation collapse in China unfolded over 5-10 years. The software developer employment collapse is happening in 3-4 years. Institutions (curricula, hiring pipelines, workforce retraining) have not adapted at that speed. Only 10% of college leaders report graduates are "sufficiently prepared" for AI workplaces.
The distribution is widening, not shifting. AI is simultaneously compressing the entry pipeline for almost everyone and creating a wage/employment premium for the AI-literate minority. The NBER cultural-lag dynamic applies: the adaptation lag between "what employers need" and "what universities teach" is the main mechanism.
Threads worth pursuing
- What does "AI-aligned curriculum" actually mean? The arXiv paper classifies by syllabi but the content signal is opaque. A follow-up could examine whether the premium comes from "mentioned LLMs" or something more substantive.
- The clerical gender equity story: 86% women among the most vulnerable 6.1M is a major structural story that gets almost no coverage relative to the developer narrative.
- Long-run equilibrium: WEF projects 170M new jobs created vs. 92M displaced by 2030. New jobs and displaced workers are different people. The transition assistance gap is the real policy question.