Pour qui votent les IA ?

Pour qui votent les IA ?

by Underscore_

Formal & Physical Sciences AI & Society UBDigital and Information technology: general topicsUBJDigital and information technologies: social and ethical aspects
🎙 Underscore_ 👥 951K 📅 March 11, 2026 ⏱ 43 min 👁 558K 📄 science communication 🧭 2026-08-03
Available in: English (current) Français

Keywords

LLMWerewolfelectionsmanipulationbenchmark

Summary

The video features an interview with Raphaël, a French student researcher, about his experiments with large language models (LLMs) playing the game Werewolf and participating in simulated presidential elections. The goal was to move beyond traditional benchmarks in math and code to assess social and emotional intelligence. In the Werewolf experiments, models like GPT-5 demonstrated advanced strategic planning, including creating detailed action plans during night phases. Gemini 2.5 Pro showed an emergent behavior of publicly apologizing and being forgiven, which was unprecedented. The election simulations revealed surprising political leanings, with some models showing extreme or unexpected positions. The discussion highlights concerns about AI’s ability to manipulate and the ethical implications of training models for social engineering. The video also includes a sponsored segment for Mammouth AI, a platform aggregating multiple AI models. The overall tone is informative and critical, aiming to make AI research accessible to a broader audience.

149 words

Critical Evaluation

The video provides a compelling and accessible overview of novel AI research, focusing on social and political behaviors of LLMs. The guest, Raphaël, presents his methodology clearly, explaining the design of the Werewolf game and the election simulations. The strength of the content lies in its originality: it moves beyond standard benchmarks to explore emergent behaviors like manipulation, strategic planning, and even apology. The examples given, such as GPT-5’s detailed night-phase planning and Gemini’s apology, are illustrative and thought-provoking. However, the scientific rigor is somewhat limited by the lack of peer-reviewed publication details; the results are presented as preliminary and based on a specific setup. The discussion of political leanings is interesting but may be influenced by the models’ training data and the specific prompts used, which are not deeply analyzed. The video also includes a clear commercial segment for Mammouth AI, which is disclosed but may introduce bias. The production quality is high, with good pacing and visual aids. The title accurately reflects the content, and the video successfully bridges technical research and public understanding. Overall, it is a valuable contribution to AI discourse, though viewers should be cautious about overgeneralizing the findings.

194 words

Title / Content Match

The title accurately reflects the content, which focuses on AI behavior in elections and social deduction games.

Quality & Reliability

8/10

The video presents original research by an invited researcher, with clear methodology and references to public benchmarks. The discussion is balanced and includes critical perspectives. However, the commercial partnership and lack of peer review for the presented results slightly reduce the score.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

Dissenting Sources

  • No direct discordant sources found — The video does not present opposing viewpoints, but the lack of peer review could be considered a limitation.

External References

Contribution & Novelties

The video presents a novel approach to evaluating LLMs by using social deduction games and political simulations, revealing emergent behaviors not captured by traditional benchmarks. It highlights the potential for AI to manipulate and the ethical considerations this raises.

Pour aller plus loin :

  • Werewolf-bench GitHub — Directly related to the benchmark used in the video.
  • LLM Politics — Website with the election simulation results.
  • AlphaGo — Historical context of AI in games, but with perfect information.
  • Social engineering — Concept relevant to the manipulation concerns raised.

87 words

Radar Profile

The radar profile shows high scores in information quantity and quality, with a slightly lower technical level, indicating a good balance between depth and accessibility. The overall reliability is strong, making it a trustworthy source for understanding AI social behaviors.

Reliability 8/10

💬 Positif. Les commentaires sont très favorables, saluant la qualité de la vidéo et l'intérêt du sujet, avec quelques interrogations sur les implications éthiques.