---
id: "lib-2026-10-26-task-expansion-with-generative-ai-experimental"
type: "library"
title: "Task Expansion with Generative AI: Experimental Evidence on Middle-Skilled, Early-Career Workers (CESifo WP 13045)"
lang: "EN"
org: "CESifo (ifo Institute)"
source_type: "Paper"
date: "2026-10-26"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-26-task-expansion-with-generative-ai-experimental/"
original_url: "https://d.repec.org/n?u=RePEc:ces:ceswps:_13045"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "A preregistered randomised field experiment by economists at ifo, TU Munich, Erlangen-Nuremberg and Germany's Federal Institute for Vocational Education and Training, with 673 apprentices, behavioural logs and robustness checks; short lab-style tasks, but the design and numbers are transparent."
summary_basis: "full-text"
original_minutes: "60"
tags: ["future of work", "enablement", "apprentices", "field experiment", "early career skills"]
---

# Task Expansion with Generative AI: Experimental Evidence on Middle-Skilled, Early-Career Workers (CESifo WP 13045)

Tier A (Strong evidence): A preregistered randomised field experiment by economists at ifo, TU Munich, Erlangen-Nuremberg and Germany's Federal Institute for Vocational Education and Training, with 673 apprentices, behavioural logs and robustness checks; short lab-style tasks, but the design and numbers are transparent.

Summary based on the full text.

## In short

A CESifo working paper (October 2026) asking whether generative AI lets middle-skilled, early-career workers do tasks beyond their training. In a preregistered experiment, 673 final-year IT apprentices in Germany were randomly given either a ChatGPT Pro account or no internet, then solved one task inside their curriculum and two that require early-career engineer skills. AI roughly doubled performance on all three, including 26-31 points on the tasks they were not trained for. Immediate unaided comprehension did not fall and on several tasks rose. Apprentices who were better at using AI gained more; prior schooling made no systematic difference.

## Key numbers

- 673: final-year IT apprentices at twelve German vocational schools, individually randomised
- 44 percentage points: gain in share of points achieved with AI on the task within the curriculum
- 26-31 percentage points: gain on the two tasks deliberately set beyond the apprenticeship curriculum
- 4%: share of points achieved without AI on the machine-learning task, where AI added 54 points
- 16%: share of apprentices who report learning effective AI use at vocational school

## What it reports

- AI raised the share of points achieved by 44 percentage points on the within-curriculum task and by 26 and 31 points on the two beyond-curriculum tasks.
- Gains were largest where unaided performance was lowest. Application Development apprentices scored 4% of points on a machine-learning task without AI; AI added 54 points. On the within task they rose 51 points from a control mean of 20%.
- System Integration apprentices gained most (40 points from 23%) on IPv4 subnet planning, which they already study; the authors suggest existing knowledge helps them check AI output.
- AI raised the chance of submitting any substantive answer by 11 points on the within task and 6 on one beyond task.
- Four multiple-choice comprehension questions answered without AI right after each task showed no decline and improvements on several tasks. Frustration fell and enjoyment rose.
- Effects were larger for apprentices with above-median self-reported AI literacy, driven by items on using AI well rather than understanding how it works. Only 16% say they learned these skills at school.
- Pasting task material into ChatGPT was the most common use; more intensive prompting and directly using AI output were associated with higher scores.

## How they got it

Individual randomisation within schools across six federal states; ethics approval from TU Munich and preregistration in the AEA RCT Registry. Each apprentice did a 10-minute baseline task, then three 20-minute tasks of 40 points each, built with rubrics reviewed by vocational teachers. Open answers were graded criterion by criterion by GPT-5.4 mini, blind to treatment; effects were similar under other grading models. Browser histories, timestamps and ChatGPT logs were recorded. The model behind ChatGPT Pro changed during fieldwork (GPT-4o, GPT-5.1, GPT-5.2) without systematic differences in effects.

## Limits

- The control had no internet, so the effect is AI plus web against nothing; against a web-only condition it would likely be smaller.
- Comprehension was measured immediately after each task, not days or weeks later.
- Short, timed, exam-like tasks in two IT occupations in one country.
- LLM grading, though checked across several models and for answer length.

## Why it matters for companies

The study tests a question behind many entry-level hiring plans: can AI let junior staff take on work that used to need more senior people? Here, at least on short tasks, it did, and without an immediate loss of understanding. The stronger signal for employers is the AI literacy gap. Gains depended on practical skill in using the tool, not on formal education, and few apprentices had been taught it. Training new hires in how to work with AI, and checking that skill when hiring, may matter more than degree level. The study does not show whether these juniors keep the skills when the tool is removed for weeks.

Original: https://d.repec.org/n?u=RePEc:ces:ceswps:_13045