The gain sits in one place, the cost in another
Nolan Lovett's paper in the journal Human Resource Development Review moves a familiar picture onto the labour side: a company that hands entry-level work to AI books 100 per cent of the efficiency gain, while the cost of eroded expertise is spread across every organisation hiring from the same pool. Lovett describes two separate routes. In the first, the post closes outright. In the second the post survives, but the new hire, helped by AI, reaches a level of productivity that used to take years and never spends the cognitive effort that builds deep knowledge.[1]
Lovett calls the mechanism's main consequence the validation tether: catching a domain-specific error in a model's plausible-looking output requires exactly the expertise that is wearing away. The paper still offers a framework rather than a measurement, and the figures it cites do not all point one way. A Federal Reserve Board study finds that growth in programming jobs has nearly halved since ChatGPT, while conceding the causal link cannot be proven; a study from January 2026 finds the decline began before ChatGPT. Anthropic's March 2026 study finds no measurable overall labour-market effect and flags only that the job-finding rate fell by half a percentage point among workers aged 22 to 25 in occupations heavily exposed to AI.[1]
The speedup is measured; the make-up of the work is not
The new edition of the alignment report Anthropic publishes every three to six months runs to 186 pages. It discloses that the company has two models more capable than Claude Mythos 5; the more capable of them, Model 2, is heavily used by staff to write software, generate training data and automate engineering work. The company estimates that its own models are raising its development pace, does not treat that as a risk, and writes that its threshold for recursive self-improvement, a doubling of the pace of progress relative to the pre-AI rate, has not been crossed. The same report lifts the risk of a model interfering with an organisation's systems from very low to low.[2]
The two headings the report scores are catastrophic harms and a model interfering with an organisation. Alongside the acceleration estimate there is no figure for how the work behind that speed is distributed across roles, and it is precisely such a figure that would make Lovett's mechanism testable. A company shifting its software work onto its own model can also mean fewer people at that company who would have learned the work by doing it. There is an ordinary explanation for the gap: the alignment report is a safety document, and role composition may be commercially sensitive as well as outside the document's subject. So the gap is no proof of concealment; it shows only that the organisation with the best data has not published the measurable side of the question.[2], [1]
What is actually measurable?
The studies on cognitive cost that the paper cites draw a sharper picture than the employment counts do. In an MIT study, more than 80 per cent of participants struggled to recall their own AI-assisted writing. In an Anthropic study with software developers, those with AI access scored 17 per cent lower on a knowledge test; the largest losses came from people who used the model purely as an answer machine, while those who asked for explanations learned markedly better. A Swiss study of 666 participants found a strong negative link between heavy use and critical thinking. Lovett's proposals fit that picture: AI-free learning environments, phased introduction, and separate competence certification by professional associations. There is no proposal for a ban.[1]
Five days ago in this column I wrote about a disclosed shift length that arrived without the headcount sharing it; the same missing denominator turns up here. Anthropic publishes the report every three to six months, so the next edition is due by the end of February 2027. The signal to watch is plain: does a figure for headcount or role composition appear beside the acceleration estimate? If it does, Lovett's mechanism becomes testable inside a single organisation for the first time. If it does not, the question of whose work the speed came out of stays an argument in which each side defends its own intuition.[2], [3]