Chapter 6 — What Software and Law Tell Us (plain-language version)

---

What the neighbouring professions show

Software developers had the problem first. Then lawyers. In both professions AI has been at work long enough to leave evidence about who is displaced, how newcomers learn and what institutions can do. This chapter draws four findings from it. It relies on published research rather than my own practice, so I speak less in my own voice here. 💡 The evidence supports the general claim: the mechanism works across knowledge professions, not only in European infrastructure advisory.

The software case

AI coding tools, widespread since 2022, have taken over the routine parts of a developer’s job: standard code, basic tests, style checks 1 2 3 . The displacement is selective 4 5 . The routine parts were also how juniors learned, through failed attempts and corrections. 💡 Yu and Moon interviewed junior and senior engineers in South Korea and found entry-level work being absorbed into the workflow of a senior working with a model. Juniors lose “the productive struggle through which expertise once developed” 6 . Protecting that path, they conclude, takes institutional design, not individual restraint.

The law case

Law shows the same pattern in research, drafting, contract review and due diligence. Law is regulated, with bar admission and formal apprenticeship, so the way lawyers are formed is visible, and the legal literature has a name for the change to that path: mediated evolution 7 . And law has responded. Bar associations, law schools and regulators in several jurisdictions have begun writing rules on AI, from disclosure to protecting how juniors are formed.

The AI-washing complication

Some firms have presented cost-cutting as AI replacing workers when the real reasons lay elsewhere. This is AI-washing. Analysts disagree on how much of the 2024-2026 technology layoffs AI really caused 8 9 10 , so layoff figures alone cannot show what AI is doing to jobs. 💡 The objection is fair, but my argument is about how experts are formed, not headcount.

The Susskinds: a different question

Richard and Daniel Susskind’s The Future of the Professions 11 is the opposing view. They predict that AI systems will increasingly deliver professional services, leaving humans narrower roles. They ask how services will be delivered. I ask how the next generation of experts is formed. They treat expertise as a delivery system that machines can copy. Following the sociologist Collins, I treat it as the ability to contribute to a practice, built through years of work with real consequences 12 . And they present the change as a neutral gain in efficiency, while I start from what the work owes, and to whom: the third box, the publics and future generations that infrastructure decisions affect. Their account is answered, not refuted.

Four findings carried forward

First, selective displacement. AI takes the routine analysis, the polished writing and the early learning tasks. It leaves judgment, relationships and responsibility for consequences. Second, the pipeline breaks whatever happens to headcount. The path that forms juniors changes even when employment holds. This is the most solid finding: the AI-washing objection does not touch it. Third, value moves unevenly, up to seniors and out to the owners of the systems. Fourth, professional legitimacy can no longer be assumed. Practitioners now have to say what the AI did and what they contributed. Audit shows a whole profession doing it: as AI arrived, the Japanese auditors’ association and the big firms’ AI teams wrote a new, upbeat identity for the auditor, Goto found 13 . A partner noted that juniors had begun to doubt the value of long manual work, which the new identity promised to remove; that work was also how auditors were formed 13 .

Medicine adds support. A review by Natali and colleagues finds core clinical skills at risk 14 , and separates losing a skill from never acquiring it. 💡 Ke and colleagues call the second never-skilling, and note that help which serves experts may harm novices. They state that “Direct evidence for never-skilling in clinical trainees remains absent” 15 , so I carry it forward as a hypothesis grounded in learning theory, not an observed result. They add that trainees “may come to view themselves as intermediaries who interpret and relay AI outputs”, and leave that open too 15 .

Four findings, one picture

The four findings are one mechanism seen from four sides: what AI takes, and what that does to formation, to the distribution of value and to legitimacy. Neither objection reaches it. AI-washing concerns layoff figures. The Susskinds look at service delivery, and from there the broken pipeline and what advisers owe the public cannot be seen. Chapter 7 applies the findings to each career stage of European infrastructure advisory.