Le jeu du dictionnaire

The game is here.

Each round shows an obscure French word. You invent a plausible definition, which is mixed with the real one and a few AI-generated decoys. Everyone votes on the correct answer and you earn points for guessing right and for fooling others with your fake.

At launch, the game loads three static files: real word–definition pairs, a list of plausible fake definitions, and precomputed word embeddings. Nothing is generated live; I used ChatGPT once to create a few hundred entries. The embeddings come from the fra_news_2024 (1M) corpus (Leipzig Corpora Collection), trained with a basic Word2Vec model in gensim.

Each round, the game picks a random word. You submit a definition. The system combines the true definition, yours, and a few sampled fakes, shuffles them, and triggers voting. You choose manually; the AI players rely on embeddings.

To decide, the game averages the word vectors in each definition and computes cosine similarity with the real one. AIs select the highest-similarity option (no self-votes, slight randomness). Scoring: +2 for a correct guess, +1 per vote your fake attracts. It runs entirely in the browser ; a small semantic bluff duel.