A five-minute opinion collection with 26 Sciences Po students, coded by theme and stance. Every tile is one opinion: colour is its stance, fill height is how highly the student ranked it.
Select a tile to read the opinion.
Themes coded from the text of each entry. Bars show the stance mix inside each theme. Select one to filter the full list at the bottom of the page.
There is no straightforward optimist in this cohort. The room divides into students who are ambivalent and students who are negative — and the ambivalence is not fence-sitting. It is a specific, repeated shape: the benefit first, the fear immediately after.
Anna is the clearest case. Her top-ranked opinion is that AI is practical for revising lectures. Her second is that relying on it damages our cognitive function. Sofia, Gilad, Marco and several anonymous participants do the same thing.
| Rank | Benefit | Harm |
|---|---|---|
| 5 — their top item | 12 | 9 |
| 4 | 7 | 16 |
| 3 | 2 | 10 |
| 2 | 6 | 3 |
| 1 | 3 | 3 |
Benefits arrive first because they are what students use. Harms follow because they are what students believe.
Time saved on low-value work, always described in the first person — summarising, revising, drafting, translating. Then AI as a thinking partner, always conditioned on the user's skill. Then medicine: cancer detection appears twice, unprompted, in almost identical words. Then, from one student only, emancipation — sovereignty, democratised access, freedom from the need to produce capital.
Hope is about efficiency far more than about capability, and almost never about a collective good.
Losing the capacity to think (14). Jobs (6). Water, energy and data centres (5). Unreliability — hallucination, altered inputs, context bleeding between chats (3). Information integrity (4). A market crash when the bubble deflates (2, and it was their highest-ranked item). Power concentration (3). Homogenisation and the dehumanisation of war (3). Existential risk (1, ranked last).
The widely shared fears are proximate and personal. The systemic ones are carried by three or four students each.
Assumptions that no entry argues for, and that the whole corpus rests on.
Almost every entry treats AI as a generative text assistant. No entry mentions recommendation systems, ranking, credit scoring, insurance pricing, predictive policing or automated administrative decisions — the AI that already governs their lives. A field has been collapsed onto the product they personally use.
Offloading effort is assumed to destroy capacity rather than redirect it, and it is stated as fact rather than hypothesis.
the brain is a muscle that must be trained, after all, and will weaken without it
One conditional runs through the corpus: if used properly, if prompted well, if we have the will power. Responsibility sits with the individual user. This is what explains the near-total absence of governance — if the problem is willpower, law is beside the point.
Benefits are described in the first person, costs in the third. No entry connects the two: nobody notices that their own lecture summaries are a claim on the water and energy they criticise elsewhere.
Adoption is something to adapt to, not something built by identifiable firms making identifiable choices — though the same students who call it irreversible also expect its market to collapse.
AI is the modern-day fire… there is nothing that can be done to undo it
People get lazy, society is destabilised, unemployment is universal. Only three entries name a who: those with access to better models, superpowers, and children. Distributional thinking — the core move of their own discipline — is largely absent.
An argument from origin, asserted rather than defended.
It will never replace human intelligence since it is made by it
Nearly every claim about society is extrapolated from two years of using a chatbot. This is the corpus's honesty and its central weakness at the same time.
In a room of political science students, the political dimension of AI is the least-represented theme in the corpus. It follows directly from the discipline assumption above: if the harm comes from users lacking self-control, then no institution has anything to do.
Every entry as written, with the student's own ranking. Text is unedited, including typos and entries cut off by the five-minute limit.