AI Will Not Replace Junior Developers

The Illusion of Total Substitution by AI

For several months, Dario Amodei has expressed a vision that, within six to twelve months, AI models could perform "most, perhaps all" of the work of software engineers. This perspective has sparked strategic reflections within companies, where the idea of no longer hiring junior developers seems appealing. However, this approach overlooks a crucial aspect: the need to train senior developers capable of supervising and improving these AI systems.

The Numbers Speak

Data from the Stanford Digital Economy Lab, in their report Canaries in the Coal Mine, reveals a significant decline in the employment of young developers in the United States. Between the end of 2022 and mid-2025, employment for developers aged 22 to 25 has dropped by nearly 20%. This decline is unique to this age group, while jobs for those aged 30-50 remain stable or are increasing. The Stanford AI Index 2026 confirms this trend. Meanwhile, PwC's AI Jobs Barometer highlights a 7.5% increase in jobs requiring AI skills, with a salary premium of 56%, and productivity that has quadrupled in exposed industries.

In France, Apec forecasts 61,160 recruitments of IT executives in 2026. However, the growth in hiring for junior profiles is limited to just 1%. The market continues to hire, but young developers are increasingly less in demand.

Learning: An Irreplaceable Process

AI automation primarily concerns explicit tasks, but it cannot replace the tacit learning of the profession. Sociologist Yann Ferguson, scientific director of LaborIA (Inria and the Ministry of Labor), has brought back into the debate a concept coined by Michael Polanyi in 1966: "we can know more than we can say." The transmission of knowledge between juniors and seniors relies on skills that cannot be documented. Juniors learn through observation and mentorship, methods that AI cannot replicate.

The Stakes of Work Quality

Ferguson also warns against the risk of prioritizing functional solutions at the expense of work quality. Code that works is not necessarily well-designed. As juniors learn, they acquire the ability to distinguish between functional code and well-crafted code, an essential skill for maintenance and regulatory compliance.

The Real Risk: Not Training Tomorrow's Seniors

The right financial question is not "how much do I save by not hiring juniors anymore?" It is "what will be the cost, in 2030, of not having seniors because I did not train them between 2024 and 2028?" This column never appears in a quarterly cost plan, but it is already materializing in the tensions around architect, lead developer, and security expert positions, which remain among the hardest to fill in France.

Reinventing the Role of Juniors

The myth of substitution poses the wrong question. The right one is: what does an augmented junior look like, and how do we evaluate them? Five shifts are being discussed at the executive committee:

• First, shift the measurement of productivity towards measuring learning: fewer resolved tickets, more progress on a public, shared, verifiable skills grid.

• Second, change the content of mentorship: the senior no longer shows how to do something, but shows how to verify, read an output, isolate a bias, reproduce a result, confront two approaches. This shift brings the developer's role closer to that of a researcher; a proposal is accepted when it is reproducible, tested on a separate validation set, and reviewed by a peer.

• Third, prioritize open tools in training, so that juniors learn a transferable profession rather than a proprietary product.

• Fourth, treat senior time as an investment, not a cost, since the same knowledge that is lacking in mentoring a junior will be needed tomorrow to govern an agent.

• Fifth, finally formalize the boundary between human and agent, as the European AI Act, whose main obligations regarding high-risk systems come into effect in August 2026, will impose traceable human supervision that no one will provide without a trained pool in advance.

Knowledge Transmission Becomes a Strategic Issue for the Entire Sector

The organizations that will succeed in the decade will not be those that have replaced their juniors with agents the fastest; they will be those that have redefined the junior role around method, verification, and reproducibility, that is, around what fundamentally constitutes a culture of adult engineering.

Transmission is a common good; an industry that drains its entry-level pipeline does not create a rent; it produces a systemic risk for the entire sector. The tools that will train the juniors of 2030 will be those that are readable, open stacks, models whose learning curves can be inspected, public metrics, benchmarks that can be replayed.

A company that entrusts the entire skill development of its young engineers to a proprietary black box does not train them; it binds them to a supplier. The question to ask before closing a work-study promotion or freezing a junior recruitment plan is threefold: who will know, in 2030, if the agent is right, who will know how to train them, with what open tools? If no one around the table can answer, it is not the junior who is surplus; it is the strategy that is lacking.