IT-Azubis können mit KI Aufgaben auf Hochschulniveau lösen / IT apprentices can solve university-level tasks with AI
A · Strong evidence A randomised field experiment by the ifo Institute with BIBB, TU Munich and FAU Erlangen-Nuremberg, run in regular classes at twelve vocational schools with 673 apprentices and a no-aid comprehension check; graded on the press release, with the full method in a CESifo working paper.
Summary based on the full text·Original: about 4 minutes to read
In short
A German-language press release from the ifo Institute (29 October 2026) on a study with the Federal Institute for Vocational Education and Training (BIBB), TU Munich and FAU Erlangen-Nuremberg. In a field experiment at twelve vocational schools in six German states, 673 final-year IT apprentices were randomly split: one group worked with ChatGPT Pro and internet, the other without, as in exams. With AI they scored about 70% of possible points on tasks set at the level of graduate software and network engineers, against about 41% without. On final-exam-level tasks the share rose from about 22% to about 66%. A comprehension check straight after each task, without aids, showed no loss for the AI group.
Key numbers
- 673final-year IT apprentices randomly assigned to work with or without ChatGPT
- rund 70 Prozent ... rund 41 Prozentshare of possible points on university-level tasks with and without AI (about 70% vs about 41%)
- rund 22 auf rund 66 Prozentshare of points on final-exam-level tasks, without AI and with AI (about 22% to about 66%)
- 84 Prozentapprentices surveyed who already use AI at vocational school
- 16 Prozentof those who say school teaches them how to use it properly
As reported by the source.
English·Web page·opens ifo.de
What it reports
- University-level tasks: about 70% of points with AI, about 41% without.
- Final-exam-level tasks: about 22% without AI, about 66% with it. Many apprentices without AI did not attempt the hardest tasks at all.
- Apprentices who rated their own AI skills higher did especially well; most taught themselves.
- 84% of surveyed apprentices already use AI at vocational school; only 16% of them say they learn there how to use it properly.
- After each task all participants answered content questions without aids. The AI group did no worse anywhere and often better.
- The authors say longer-term retention needs further research.
How they got it
Field experiment from October 2025 to January 2026 in regular lessons. Participants were final-year trainees in application development and systems integration. A card draw assigned groups at random. Everyone did four tasks: a pre-test, one exam-level task and two university-level tasks. The full analysis is in CESifo Working Paper No. 13045 and two journal articles, which were not read for this digest.
Limits
- Graded on the press release; effect sizes are rounded and no confidence intervals are given here.
- The control group also had no internet, so the comparison is AI plus internet against neither.
- Comprehension was checked immediately after each task, not weeks later.
- One occupation, one country, one model.
Why it matters for companies
The study covers middle-skilled, early-career staff, the group most firms worry AI will displace. With a general-purpose assistant, apprentices did work normally assigned to graduate engineers, and did not lose understanding in the short term. Firms that hire from vocational training can widen what junior technical staff are trusted to do, provided they teach AI use deliberately: most apprentices learn it on their own, and those with better skills got more out of it.
Key passages in translation
Quoted in English
Mit ChatGPT holten die Azubis bei Aufgaben auf Hochschul-Niveau rund 70 Prozent der möglichen Punkte, ohne KI rund 41 Prozent.
With ChatGPT, the apprentices scored around 70 percent of the possible points on university-level tasks, and around 41 percent without AI.
Quoted in English
Viele Azubis ohne KI versuchten die schwersten Aufgaben gar nicht erst.
Many apprentices without AI did not even attempt the hardest tasks.
Quoted in English
Ob die Leistungssteigerungen tatsächlich ohne Einbußen beim Verständnis einhergehen und das Gelernte langfristig erhalten bleibt, muss weitere Forschung klären.
Whether the performance gains really come without any loss of understanding, and whether what was learned is retained in the long term, must be clarified by further research.
Summary based on the full text. Read the original for the full argument and data.Read the original on ifo.de