10 years of AI in French politics on X.

Over ten years, artificial intelligence has moved beyond purely technical debate. It now appears in political statements about sovereignty, security, work and cultural policy. It is also entering the presidential campaign. Taken individually, political messages offer a fragmented picture of this trajectory.

I compiled 2 381 tweets, replies and quote posts with original commentary published by 364 French political accounts between 2016 and 2026. Simple reposts were removed. The main view covers the entire decade. A second tab focuses on the 48 hours of the Mythos US scandal, when American restrictions on Anthropic's advanced models triggered a tightly compressed sequence of political responses.

What the map shows

Each point corresponds to a complete statement, rather than a sentence taken out of context. Two nearby points indicate semantically similar wording or arguments. Colors are used to display themes, parties or argumentative stance. This reading makes it possible to separate the content of the discourse from political affiliation.

The projection helps identify families of discourse, their areas of overlap and cases that diverge from the group. It provides a reading of proximity between texts without turning the displayed points into ideological truth. Labels remain visible above the clusters of points. The topic filter also allows a specific subset to be isolated while keeping the rest of the corpus on screen.

Mythos US has a separate tab

The ten-year view tells the story of a slow conversation. It spans regulation, innovation, security, sovereignty and public services. Mythos US unfolds over a much shorter timescale. Within a few hours, an American decision focused reactions on European technological dependence, even when the proposed political responses went in opposite directions.

This tab complements the overall view. It shows how an external event can suddenly bring together vocabularies that, across the decade, remain associated with much more distant political and thematic families.

How to read the topic network

The graph positions themes based on their co-occurrences. Links behave like springs. The more often two topics appear together in the same tweets, the more they attract each other. At the same time, nodes repel each other. This mechanism produces an irregular geography and allows subgroups to form without imposing an artificial center.

Method and limitations

The corpus comes from public content published on X. Simple reposts are excluded. Replies and quote posts with original commentary are retained. Themes are generated using BERTopic. Proximities are based on multilingual embeddings, followed by UMAP. Expressions are extracted using keywords, after removing the generic terms AI and artificial intelligence.

The visualization published here contains only the aggregates and coordinates needed for the graph. The collection database and raw tweet text remain outside the publication. A point's account name and date are served separately on click through a restricted route. Party affiliation remains an aggregate data field in the graph. A small selection of public tweets is retained as verifiable examples.