AI helps with architecture tasks beyond the curriculum
Hanna Brosch, Christina Langer, Philipp Lergetporer, Sandra Mueller et al. · Stanford Digital Economy Lab · 7 Oct 2026
Source finding
On software or security architecture tasks beyond the curriculum, IT apprentices with ChatGPT achieved about 65 percent of available points on average, versus about 39 percent without AI or internet. This measures task scores, not the share of people succeeding.
Our interpretation
AI can make more demanding tasks accessible. Good AI-assisted solutions do not establish durable skills.
Method and evidence
Preregistered field experiment: 673 final-year apprentices at twelve German vocational schools, randomly assigned within schools. Data collected October 2025–January 2026. Open answers graded by AI against expert-reviewed rubrics. Strong evidence for this short-term comparison; working paper, not peer-reviewed.
Limitations and interests
Control group without internet; not a comparison with normal work including web search. Two IT apprenticeships, mostly men. Immediate comprehension assessed, long-term learning not tested. Possible AI-grading bias; alternative grading models checked. Funding from the Dieter-Schwarz Foundation and Dr. Hans-Riegel Foundation.
What businesses can do
Trial more demanding tasks with AI. Check the results and ask employees to explain their solution. After a few weeks, test again what they can do without AI.
Where to start
Table 2 for tasks, section 3.5 for grading, section 5.1 and Table A.4 for scores. Original PDF