In a four-week study at the MIT Media Lab, people who fact-checked news with an AI assistant became 21 percent more accurate. When the assistant was taken away, their unaided performance landed 15 percentage points below where they had started. About a quarter of them believed they had improved. It is a striking result, and a new essay from the consultancy triangility uses it to argue that companies must build judgment alongside tool skills, through a mix of workflow changes, redesigned entry-level roles and longer training.
We think the essay has the diagnosis right and the remedy too small. Judgment is what remains after years of junior work done badly, reviewed and redone, which is why a course cannot produce it. Companies that slow entry-level hiring to bank AI savings are removing the place where that judgment forms, and no training programme, including the judgment-focused kind, puts it back. Protecting judgment is a headcount and job-design decision, and it belongs with the executives who make those decisions.
Writing in triangility's magazine, Prof. Frank Widmayer pairs the MIT study, led by Valdemar Danry and Pattie Maes, with a Harvard analysis by Seyed Mahdi Hosseini Maasoum and Guy Lichtinger. That analysis covers 62 million workers at roughly 285,000 companies between 2015 and 2025. In firms that actively integrate generative AI, it finds that junior employment falls about 7 to 9 percent within six quarters relative to similar firms. The decline comes almost entirely from slower hiring, while senior employment keeps growing. Widmayer's prescription is what his firm calls "Authority-in-the-Loop", delivered through a six-month learning journey that his firm happens to sell.
The evidence is narrow. The mechanism is not.
Sixty-seven people checking news stories for four weeks tell you little about a finance director's year. A cross-firm hiring gap measured over a decade that included economic cycles and hiring swings may say more about which companies were already planning to hire fewer graduates than about what AI does to them. On this reading, the source has taken a small lab result and a correlation and presented them as a workplace crisis.
The sceptics are right about magnitude and wrong about direction. The Harvard comparison sets adopters against similar firms without AI integration, which should net out much of the shared economic cycle. That does not rule out selection, because firms with leaner hiring plans may have reached for AI first, but it makes "it's just the macro cycle" a weaker explanation than it first sounds. The MIT result also has company. A multicentre observational study in The Lancet Gastroenterology & Hepatology examined deskilling risk among endoscopists after routine AI exposure in colonoscopy, a professional setting with real stakes. A recent arXiv paper is titled, without much hedging, "AI Assistance Reduces Persistence and Hurts Independent Performance." These studies do not size the effect for managerial work, but they point the same way: regular reliance on AI is associated with weaker independent performance. The MIT study adds the more troubling finding that about a quarter of participants believed they had improved even as their unaided performance fell.
The pipeline argument also survives an unresolved causal debate. Whatever is driving it, firms actively integrating generative AI are hiring fewer juniors while their senior employment keeps growing, and the tasks AI absorbs first are the tasks through which today's seniors became senior. Whether AI or the economy closed the door matters less than the fact that it is closing.
Judgment is learned in the job, and a workshop cannot stand in for the job
Widmayer makes a good point when he says operational skill "can be taught in an afternoon" while judgment "matures through experience, feedback, and repeated confrontation with one's own mistakes." His prescription also reaches past the classroom. Among three leadership tasks, he urges firms to protect moments when people think before consulting AI, and to reinvent entry-level roles rather than eliminate them, giving young professionals harder work that builds judgment. That is the right instinct, but it sits on a list alongside configuring AI as a coach, it is framed as job design with no claim on headcount or budget, and the offer his firm actually sells is the six-month journey. A six-month journey beats a prompt workshop, but leaning on it repeats the original mistake on a slower schedule by treating judgment as something learning and development can supply. Judgment forms in the analysis a senior colleague takes apart and in the client call that went badly, and those things happen in jobs. A company that stops hiring juniors has not automated their work. It has cancelled the apprenticeship of its future senior staff.
The vendor vocabulary needs some discounting. "Authority-in-the-Loop" is human-in-the-loop with capital letters. "Transformation debt", which the essay says surfaces after about 18 months, arrives without evidence for the timing. Underneath the branding, the ideas are sound and old: deliberate practice, and thinking before you consult. The most useful contribution comes from the MIT researchers themselves, who distinguish AI used as a coach, which asks questions and pushes people to reason, from AI used as a crutch, which hands over answers. A separate arXiv paper, "Designing Against Deskilling", reports that metacognitive feedback reduces cognitive offloading to LLM assistants, so tool design can address part of the problem. A coach, however, needs someone to coach, and that is the person many firms have stopped hiring.
A company that stops hiring juniors has not automated their work. It has cancelled the apprenticeship of its future senior staff.
Make the junior tier a cost you choose
The saving from slower junior hiring is real and shows up immediately. The cost arrives years later, when the current senior bench retires or leaves and the people meant to replace them never did the work that would have prepared them. Leaders should treat entry-level seats the way they treat R&D: as spending on a capability the business will need and cannot buy at short notice. That shifts the conversation from "how many juniors does AI let us skip" to "what should juniors now do that AI cannot teach them." Widmayer gets to the second question; the answer has to show up in the headcount plan as well as the training calendar. One answer is to have juniors produce the first draft before any model sees the problem, with seniors reviewing both. Another is to test unaided performance periodically, because the MIT finding that a quarter of participants felt better while getting worse means confidence surveys will not reveal the problem.
Regulation adds pressure. Article 4 of the EU AI Act, in force since 2 February 2025, requires providers and deployers of AI systems to ensure a sufficient level of AI literacy among staff who handle those systems. Widmayer reads it as covering the ability to assess risks and make informed decisions, not just operate tools. If that reading holds, a compliance team that meets the requirement with a tool course has bought a certificate and little competence. Judgment is the hard part of the obligation, and slowed junior hiring is eroding exactly that part.
Firms that stop hiring juniors this year will look efficient for several years. The cost will show up on a specific day, when a model gives a confident and wrong answer and the person who should catch it never spent the years learning what wrong looks like. No training budget will have helped, because years earlier the company decided it did not need to hire that person at all.
