From Learning to Delegated Work | AI Wars of 2026
In research: whether capability transfers. Source-backed analysis from AI Wars of 2026.
From Learning to Delegated Work
In research
This chapter asks whether capability built in education survives contact with unsupervised work. It is still in research, and the transfer question is the least evidenced link in this part.
Everything earlier in this book concerns distribution: which platform reaches the workplace, through which identity system, under whose governance. This chapter concerns the thing distribution cannot deliver. Judgment does not arrive through a license agreement. It arrives with the person.
The questions this chapter has to answer:
- Does capability transfer? A person who used an assistant competently under instruction — with a known task and a marked outcome — is not yet someone who delegates well. Ambiguous tasks with nobody checking the result are a different test.
- What do employers observe? Early-career hires are the first cohort to arrive having used these systems throughout their education. What, if anything, is measurably different?
- Who bears the cost of failure? If capability does not transfer, the cost appears somewhere: in review load on experienced staff, in errors reaching customers, or in work quietly not being delegated at all.
- Does experience level change the effect? Existing field evidence outside education finds learning effects concentrated among less experienced workers in one specific setting. Whether that pattern holds for people entering work is an open question, not an established finding. Source 1 Source 2
The customer-support study reports larger gains among less experienced workers in that setting. METR's early-2025 randomized study instead concerns experienced open-source developers working in familiar repositories, where AI-allowed work took longer despite participants' perceptions of speed. METR now labels that result out of date. Its early-2026 continuation suggests newer tools may help, but it reports selection effects and estimates too uncertain for a reliable point result. Different tasks, experience levels, and tool periods prevent these findings from becoming one general productivity verdict. Source 1 Source 3 Source 4
None of these studies follows capability acquired in school or university into later work. Learning while using an assistant on a job is not the same as transferring judgment from education. A study of that transfer would need to follow the same people and test their decisions, error detection, and need for supervision over time.
What will make this chapter publishable
Longitudinal evidence, or employer evidence about early-career capability, that measures something other than tool familiarity. Failing that, the chapter should say clearly that the transfer question is unanswered, which is a more useful statement than a confident answer assembled from adjacent studies.
This chapter must not forecast workforce displacement. The book does not predict outcomes, and this is the chapter most exposed to the temptation.
Sources
Source references
Source 1: Generative AI at Work
National Bureau of Economic Research
Published: 2023-04-24
Field study of a generative-AI assistant for customer-support workers. It reports an approximately 14% average productivity gain in that setting, with larger gains for less experienced and lower-skilled workers; it does not test autonomous agents or generalize to all knowledge work.
Source 2: Generative AI at Work
Quarterly Journal of Economics
Published: 2025
Field evidence from customer-support work showing heterogeneous performance and learning effects after AI adoption. It does not establish results for all knowledge workers or autonomous agents.
Source 3: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
METR
Published: 2025-07-10
Randomized study of 16 experienced open-source developers and 246 issues found AI-allowed work took 19% longer in its early-2025 setting. METR now labels this result out of date; see its 2026 continuation for current but still inconclusive evidence.
Source 4: Update on our early 2026 AI productivity study
METR
Published: 2026-02-24
METR's continuation of its developer-productivity research. Early-2026 results suggest developers may be faster with newer AI tools, but METR reports severe selection effects and a confidence interval too wide for a reliable point estimate.