AI Took the Starter Tasks. Who Teaches Judgment Now?

The starter tasks are leaving. The judgment they used to teach is still on the scorecard. Headcount is easier to count, and it can't show where new hires will learn the job.

A Stanford Digital Economy Lab note dated 12 August 2026 finds employment for workers ages 22-25 in highly AI-exposed occupations about 19% below the path set by same-age workers in less-exposed occupations as of June 2026, up from 15% at the July 2025 vintage. The adjustment appears to run through reduced hiring, not separations, and young employment has fallen where work relies on codified knowledge. The authors call the patterns descriptive, not causal. 1

In Strada's 19 May 2026 report, 42% of employers that have explored AI and hire entry-level staff say AI increased analytical or judgment-based responsibilities, and 33% say it reduced foundational or skill-building tasks. 2

No source here measures whether judgment now forms more slowly. The open decision is who runs the repetitions those tasks used to supply.

1. 1. Young workers in AI-exposed jobs are falling further behind, and hiring looks like the channel. Stanford's peer gap went from 15% at the July 2025 vintage to about 19% as of June 2026. The authors call it descriptive. 1
2. 2. Codified work is where young workers lose ground; experienced workers gain where knowledge is learned by doing. Stanford calls that knowledge tacit, the kind that comes from practice and mentorship. 1
3. 3. Employers say AI is adding judgment to entry-level work and reducing skill-building tasks. Of 1,387 employers that have explored AI and hire entry-level staff, 42% report more analytical or judgment-based responsibilities and 33% fewer foundational or skill-building tasks; answers could overlap. 2
4. 4. When routine starter work goes, on-the-job learning can go with it. An OECD working paper argues from the literature that automating routine entry-level tasks may hinder learning on the job, though AI could also help knowledge pass. 3
5. 5. Leaders endorse the design work, and few say they lead it. In Deloitte's 2026 survey, 66% call intentional human-AI interaction design important to organizational success, and 6% say they are leading in it. 5

Brief 018 asked what a role is for once AI absorbs routine cognitive work, and answered for people already in the job: redesign the ladder around what remains. This Brief asks the earlier question. Those routine tasks were also how a new hire learned the judgment the ladder later rewards, so the subject here is the first rung. Brief 016 put an assistant inside a lesson; Brief 004 tracked the training calendar. This one is about the entry-level job's task mix.

The headcount fight is real and unsettled. The task shift is better supported. Payroll data shows young employment falling where work is codified, employers describe what left the job, and even the strongest counter-evidence here finds AI moving tasks but not pay or hours. 126

Young hiring fell where the work was codified

Brynjolfsson, Chandar, and Chen's note revises the Canaries work on ADP payroll records. The 19% is a peer gap, not a raw drop: in a footnote, ages 22-25 in the two most exposed quintiles fell about 11% from November 2022 to June 2026, while the three least-exposed quintiles grew about 10%. 1

The adjustment "appears to operate primarily through reduced hiring of young workers rather than increased separations," and it shows up in employment rather than base pay. Experienced workers show no comparable gap. Declines concentrate where AI automates tasks; where it complements the work, employment is flat or rising. Young employment has fallen in occupations that rely heavily on codified knowledge, while experienced employment has risen where knowledge is tacit, "acquired through practice, mentorship, and repeated exposure to real situations." 1

This Brief draws its own inference from that split: the codified tasks are the ones a tool can take, and doing them under supervision was where new workers began building the tacit kind. Beyond calling the patterns descriptive, the authors note that the gaps shrink when education is held constant, and they flag pre-trends, a sample with larger gaps than national benchmarks, and estimates that move with the specification. 1

Employers have already edited the entry-level task list

Strada's 19 May 2026 report describes a survey Artemis Strategy Group fielded March 3-22, 2026, of nearly 1,500 executives and senior talent leaders, weighted to US employers. The task question went to employers that have explored AI and hire entry-level staff (N=1,387), and they could mark more than one change. 42% say AI increased analytical or judgment-based responsibilities. 41% say it reduced routine or administrative tasks. 33% say it reduced foundational or skill-building tasks. 20% report no meaningful change. Strada did not test whether learning slowed. 2

Andre and Schief, in OECD working paper 1870 (landing page 5 July 2026), explain what is at stake: automating routine tasks younger workers have traditionally performed "may complicate entry into the workforce and hinder the intergenerational transmission of knowledge, as young people get less opportunities to learn on the job." The same section says AI could also help pass knowledge on. It is a literature argument, and "may" is the operative word. 3

The design response is thin

Deloitte's 2026 Global Human Capital Trends survey covers more than 9,000 business and human resources leaders in 89 countries; the press release is dated 4 March 2026. In the design chapter, 66% say intentional design of human-AI interaction is important to organizational success, and 6% say they are leading in it. In the press release, 85% call it critical to build the organization's and the workforce's ability to adapt, and 7% say they are leading in helping their workforce continuously grow and adapt. Both are self-reports of "leading," not audits of what anyone built. Neither item asks about entry-level practice. 45

Autonoma forecast: Over the next 12 months, employers that keep taking codified starter tasks off entry-level jobs will still expect the judgment those tasks used to build, unless they give new workers another way to practice. 12345

The first five track the sources. The last two are local checks.

• The Stanford gap for ages 22-25 holds at or above 19%, still through hiring rather than separations. 1
• The codified-versus-tacit split survives education controls and new specifications, or fades under them. 1
• Employer surveys keep reporting more judgment work and fewer skill-building tasks for entry-level staff. 2
• Follow-up research tests the OECD's "may" in both directions. 3
• The share saying they lead human-AI design or workforce adaptation moves out of single digits. 45
• A new hire's first months include no supervised case, shadow, or redo of work the tools now finish.
• Pay and hours look stable, so a task shift is easy to miss until someone asks what the new person practiced. 67

For hiring managers and early-career supervisors: If the job now opens on judgment, name what replaces the tasks a tool finishes: a supervised case, a shadow, a redo. A shorter list is not a finished role. 12

For CHROs: Keep headcount and practice as separate questions. The evidence here does not support shrinking entry-level hiring, and none of it says where a new hire practices. 26

For CLOs and L&D: A course won't close this gap. The missing piece is practice inside the work, which the OECD paper says automation may erode. Design that practice with the same care as a course. 3

For people starting the work: Ask which tasks used to teach the judgment you'll be graded on, and who runs them now. If no one can say, the practice is on you.

For CIOs buying tools that take starter work: Put the practice path in the same decision that removes the task, and name its owner. 24

Weight: Strong. The strongest objection is that the labor data does not pin early-career losses on AI. Humlum and Vestergaard report precise null effects of AI chatbots on Danish earnings and hours. Brookings, on 6 October 2026, and NBER Working Paper 33777 (issued May 2025, revised March 2026) rule out average effects larger than 2%, two years after ChatGPT, in records through December 2024. AI-exposed occupations have seen falls in early-career jobs, but Brookings data show that firms adopting generative AI are not driving these trends. Strada finds 2.7 times as many senior talent leaders expect AI to increase entry-level hiring in 2026 as expect a decrease. 267

It goes further. The same research finds new tasks in content generation, oversight of AI, and integration of AI, and the OECD paper says AI could help knowledge pass. New workers may be learning judgment by checking a model's output. 36

Much of that stands. This Brief's case rests on the task list, and the Danish results show null pay and hours effects alongside changed tasks. More seats are not a practice path, and checking a model's output may build judgment or may assume it. Direct evidence that new workers still form the old judgment, on the thinner task list or through oversight work, would weaken this Brief. No source here provides it. 26

Entry-level jobs carried a curriculum nobody wrote down. The routine tasks were the work, and doing them under someone's eye was how a new hire learned what good looks like. When a tool takes the tasks, the curriculum goes with them, and no headcount report shows it.

Put last year's entry-level task list beside this year's. Mark what a tool now finishes on the first pass, then mark the judgment you grade in the first six months. Each graded judgment with no task left to practice it needs a named owner.

Who teaches judgment now? Whoever is named to run the supervised case, the shadow, and the redo. A model can help. It can't be the whole answer if it also finished the task.