AI is not destroying the ladder. It is removing the first rung.

The quiet exclusion happens before the layoff

The most important AI labor signal may be the one that never becomes a termination notice. A revised Stanford Digital Economy Lab working paper finds no widespread economy-wide displacement in payroll data covering millions of workers. But among people ages 22 to 25 in occupations most exposed to AI, employment is 19% below the path it would have followed had it kept pace with less-exposed peers.

The divergence appears mainly through reduced hiring, not increased separations. Experienced workers in the same exposed occupations show no comparable decline. That combination should change the question leaders ask. The risk is not only whether AI eliminates existing jobs. It is whether it quietly stops new people from ever entering the profession.

The first rung was doing more than producing output

Entry-level work has always contained tasks that looked inefficient from the top: draft the memo, check the numbers, trace the error, sit beside the customer, watch an expert revise the work. Those tasks produced value, but they also produced experienced people. If automation captures the output while institutions discard the learning process, productivity today can create a capability shortage tomorrow.

That is why the evidence from AI tutoring matters to the labor debate. In a preregistered six-week study of 1,059 introductory programming students, tutors steered through system prompts toward planning, monitoring, reflection, and deeper cognitive engagement changed some exploratory measures of activity. They did not produce statistically significant differences on the four preregistered learning outcomes. More interaction was not confirmed as more learning.

Do not automate the apprenticeship and call it training

Organizations cannot assume that an AI coach will replace the tacit learning embedded in real work. A system can explain, question, and simulate. It cannot automatically recreate the social obligation of a manager who reviews a novice's judgment, shares responsibility for mistakes, and decides when that person is ready for harder work.

The tutor study does not prove that scaffolding cannot help. It shows that system prompts alone did not deliver the measured learning gains. The burden therefore shifts to employers and educators: measure capability, preserve supervised practice, and test whether an automated learning intervention changes outcomes rather than merely increasing engagement.

Inclusion is a design decision

Google DeepMind's sign-language-to-text release offers a different model of deployment. The product was conceived by a Deaf employee and developed with Deaf data partners, researchers, user studies, and an advisory committee. Its initial launch is limited to American Sign Language and English on Pixel 11, but the process recognizes that access technology should be built with the people who will rely on it.

The contrast matters. AI can remove pathways when efficiency is measured only as headcount or output. It can widen pathways when institutions define the problem with users and treat participation as part of technical quality. Neither result is inevitable.

Rebuild entry before celebrating leverage

Experienced workers can become more productive with AI while the pipeline behind them dries up. That is not a stable victory. It is borrowed expertise. Companies that automate junior tasks should be required by strategy, and sometimes by policy, to explain how newcomers will acquire the judgment those tasks once taught.

The first rung does not need to look like the old job. It does need to be real, paid, supervised, and measurable. If nobody can enter, practice, fail safely, and become experienced, today's productivity gain becomes tomorrow's talent collapse.

• Track entry-level hiring and promotion by age and occupation, not only total headcount.
• Preserve paid apprentice roles when AI automates the tasks that once trained novices.
• Measure learning outcomes before replacing human instruction with an AI tutor.
• Give users affected by an AI system a formal role in its design and evaluation.
• Make every automation plan state how the next generation will acquire expert judgment.