Can artificial intelligence extend students’ analytical capabilities without displacing their independent judgement? This column reports on a pilot study in a graduate seminar investigating weaknesses in European integration. An ‘AI-enriched contract’ required three elements for students’ presentations: an independent conceptualisation before AI use, an AI-assisted presentation, and a transparency statement. AI-tools accelerated routine execution, widened methodological ambition, and enabled exploration. Students also documented verification, rejected inappropriate suggestions and fabricated quotations, and retained responsibility for organisation and interpretation of findings. While the pilot offers qualified support for AI as a workmate rather than a master, several caveats apply.
Generative AI is rapidly changing what students can do, and what universities need to assess. It can help students analyse data, explore unfamiliar methods, and communicate ideas, but it also raises a central educational challenge: how can students benefit from these capabilities without outsourcing the independent judgement that graduate education is meant to develop? In the spring of 2026, we explored this question through a six-week graduate seminar at the Goethe University, where 31 master’s students used AI to investigate weaknesses in European integration.1 Students were asked to investigate why, at present, the EU seems more exposed to international shocks and political, financial, and industrial vulnerabilities than in the past.
Groups investigated several areas: defence and security, migration and labour markets, energy and the green transition, strategic autonomy and industrial policy, the Single Market, and the conduct of monetary policy under uncertainty.
Why an experiment on AI?
We knew that answering would be challenging. Views on artificial intelligence vary widely, from the utterly dystopian to the perhaps idealistic.2 We are increasingly surrounded by AI, and several human activities are increasingly affected by it and enveloped by prejudice about it. The literature points to several concerns:
- ‘AI deception’ and unreliable outputs, including hallucinations, fabricated results, and references (Novy-Marx and Velikov 2026, Campante et al. 2025).
- ‘Cognitive offloading’: the tendency to delegate specific cognitive tasks to technology and to remember where to find an answer rather than the answer itself (Chirayath et al. 2025, Galiani et al. 2026).
- ‘Cognitive surrender’: the more corrosive risk that people stop thinking for themselves altogether (Shaw and Nave 2026, The Economist 2026, Demirer et al. 2026).
These concerns matter well beyond classrooms. As AI takes over routine cognitive tasks, employers will place a growing premium on people who enjoy thinking hard and can reason critically and independently (The Economist 2026). Universities have been warned to ‘stress-test’ their assessments as a large majority of students — reportedly 92% in the UK — already use AI (The Guardian 2026).
In the seminar we adopted a pragmatic and optimistic stance, in line with two international agencies. UNESCO emphasises a human-centred, capability-building approach to generative AI in education: attentive to agency, equity, data privacy and responsible use (UNESCO 2026). The OECD similarly warns that assessment should focus on whether students are learning, not merely producing polished outputs.3
The challenge: AI — ‘Master and Commander’?
Can AI extend students’ analytical capabilities without displacing their independent judgement? Can AI support graduate inquiry without replacing the thinking? AI changes the process through which graduate students learn. Is there a risk of AI becoming a ‘Master and Commander’? The phrase originated in the 18th Century to describe Royal Navy officers who captained smaller warships — sloops-of-war carrying up to twenty guns. The master and commander was responsible for sailing the ship, managing the crew, and fighting when necessary: he was in total control of vessel and crew.
The analogy is deliberate. Are today’s powerful and rapidly improving AI tools becoming our master and commander? Will AI take over any task we begin, render us redundant, and ultimately control our principal activities — or can it remain a capable subordinate that we direct?
The design of an AI-enriched contract
Safeguarding human agency and getting “just the right amount of help from AI” turns out not to be straightforward. As the seminar made clear, useful collaboration with AI tools takes structure, discipline, preparation, and a level playing field. To enforce that discipline, we set a contract for every group.
Each group was asked to submit three elements of their project:
- A pre-AI outline — with a clear research question, motivation, initial sources, and analytical plan, written before using any AI-tool. To anchor ownership of the idea in the students.
- A core presentation, enriched by AI — the polished analysis and slides produced with AI assistance.
- A transparency statement laying out the AI-value added — which AI-tools were used and how, listing prompts, and describing what worked or failed.
The last lecture refined each group’s research questions and motivation, building on earlier class and email exchanges and helping students approach the presentations strategically (Table 1).
Table 1 Three dimensions of our assessment


The topics chosen
Thirteen groups presented, covering the main fault lines along which European integration remains incomplete, including defence and security, migration and labour markets, energy and the green transition, industrial policy, the Single Market, and monetary policy (for a detailed list, see Mongelli and Tohme 2026). Several projects asked how recent shocks – for example, Russia’s invasion of Ukraine, the 2022 energy crisis, renewed trade tensions – exposed gaps in Europe’s architecture; others assessed whether flagship EU initiatives such as REPowerEU, the Green Deal, or the Critical Raw Materials Act are on track. Methods ranged from policy analysis and surveys to panel regressions, difference-in-differences, and text analysis. This diversity matters for our purposes: it allowed us to observe how students used AI across very different research designs, from qualitative policy assessment to quantitative estimation.
What we found
The breadth of the presentations is itself a finding. Five patterns recur across groups.4
AI as an accelerator of routine work
AI compressed the ‘mundane’ parts of empirical work: cleaning and reshaping data, debugging codes, formatting charts, tidying slides, translating foreign-language sources, and summarising long documents. The gain was time: students reached usable outputs faster and spent the saved hours on argument rather than mechanics.
AI as a capability-extender
Several groups reached methods they did not previously know. AI did not do the thinking but helped students in using and executing new methods.
AI as an enabler of exploration and serendipity
Because the core pipeline ran quickly, students could afford to explore. Within a six-week seminar, that extra layer would otherwise never have been reached.
The human core: Judgement, interpretation, and fact-checking
Crucially, the groups drew the line in the same place. They reserved the research question, the economic reasoning, the choice of arguments and the final interpretation for themselves, and they verified what AI produced.
New literacies and emerging skills
First, prompting is a craft: an effective prompt states the goal, the role the AI should take, the context, the sources, the desired format and tone, what to exclude, and — where useful — an example. Second, and more important, is what one group called “playing the why game”. “Know your AI, prompt carefully, provide source material, fact-check sources and their reliability, and remain critical” — through practice rather than instruction.
Verdict: Friend and agent, not master and commander
Based on the evidence of this seminar, AI was not the master and commander. Used within the seminar’s contract, AI behaved as a capable subordinate, while the students retained command of the questions, the judgement, and the conclusions. On average, each student presented and answered questions for well over 15 minutes. AI appears to raise the ceiling of what graduate students can achieve without lowering the floor of what they must understand themselves.
Limitations
The cohort was small and self-selected into the seminar; it was not random. Much of the evidence consists of students’ reflections rather than objective measures of learning. There was no control group working without AI. Consequently, reported improvements in speed and ambition cannot be causally attributed to AI or to the contract, and durable learning remains unmeasured (Mongelli and Tohme 2026).
Future pilot studies will assess outlines and final outputs separately, require prompt logs systematically, and track whether critical questioning persists. It might be worth testing whether exposure to an AI-enhanced contract continues to support graduate studies while preserving independent and critical thinking well after the seminar. At the same time, the AI bar and graduate education are continuously moving, as discussed by Weder di Mauro and Ash and Hansen (Weder di Mauro 2026, Ash et al. 2023).
Source : VOXeu








































































