Samuel Coavoux
LightNews LightNews
banner
scoavoux.bsky.social
Samuel Coavoux
@scoavoux.bsky.social
1.2K followers 700 following 110 posts
Sociologist. Assistant Prof @ENSAE & @crestsociology.bsky.social, Paris. I study cultural inequalities & digital technologies using mixed methods and CSS.
Posts Replies Media Videos
scoavoux.bsky.social
this language detection R package works extremely well
text in korean misclassified as esperanto by a text detection R package
September 20, 2025 at 7:56 AM
scoavoux.bsky.social
image exclusive de la passation de pouvoir à Matignon
dessin: deux personnes en costume de spiderman pointent un doigt l'un vers l'autre
September 10, 2025 at 9:01 AM
scoavoux.bsky.social
Specifically, we find that streaming users consume more varied music, movies and TV shows genres compared with similar non-users. We also find they have more cosmopolitan consumption, i.e. are more likely to consume items in a foreign language. The difference is higher for TV shows.
Secondary result plot from the paper. Difference in the consumption of foreign-language content between streaming users and non-users. Streaming users consume cultural content in foreign (non-French) languages more often than non-users, after controlling for confounders. The difference is small for music, intermediate for movies, and high for TV shows. Proportion difference before (light) and after (dark) adjustment through matching, with error bars indicating 95 percent confidence intervals. Main figure of the article. Difference in number of genres consumed, liked, and disliked between streaming users and non-users. Streaming users consume more genres than non-users after controlling for confounders. The difference is small for music (0.1 sd), moderate for movies (0.2 sd), and high for TV shows (0.46 sd). However, differences in number of genres liked or disliked are small or not significant. SMD before (light) and after (dark) adjustment through matching, with error bars indicating 95 percent confidence interva
September 5, 2025 at 8:40 AM
scoavoux.bsky.social
Do streaming platforms trap us in cultural filter bubbles? We like to think so but the evidence says otherwise. In a new paper @abelaussant.bsky.social and I find the use of streaming platform to be associated with an increase in consumption diversity. sociologicalscience.com/articles-v12...
Screenshot of the title and abstract of the article. the title is Streaming Platforms, Filter Bubbles, and Cultural Inequalities. How Online Services Increase Consumption Diversity. The abstract reads: Do digital technologies affect diversity in cultural tastes? Digital sociologists have warned of “filter bubbles,” whereas sociologists of culture have shown that diversity in consumption is valued as a marker of upper-middle-class status. We estimate the effect of using streaming platforms on the diversity of cultural consumption using a matching technique applied to 2018 survey data from France. We find a statistically significant positive effect of using streaming platforms on the diversity of cultural consumption as well as on cosmopolitanism, on three domains, music, movies, and TV shows. The magnitude of this effect is much higher for TV shows. The study brings new evidence against the filter bubble thesis; it shows that platforms do reinforce cultural inequalities by increasing the social gap in consumption diversity. It further suggests that the effect of technology on cultural consumption might mainly operate through its impact on cultural markets rather than changes in cultural experience. Main figure of the article. Difference in number of genres consumed, liked, and disliked between streaming users and non-users. Streaming users consume more genres than non-users after controlling for confounders. The difference is small for music (0.1 sd), moderate for movies (0.2 sd), and high for TV shows (0.46 sd). However, differences in number of genres liked or disliked are small or not significant. SMD before (light) and after (dark) adjustment through matching, with error bars indicating 95 percent confidence interva
September 5, 2025 at 8:40 AM
scoavoux.bsky.social
Faire de la philosophie pendant 50 ans et ne pas comprendre que socialement construit != non existant c'est quand même fort
August 29, 2025 at 6:33 PM
scoavoux.bsky.social
Great @theguardian.com article on Netflix movies that puts a lot in perspective. It is especially interesting to see how much "the algorithm" has just replaced the major studio producer in discourses criticizing how the industry limits creativity www.theguardian.com/media/2025/a...
August 29, 2025 at 6:53 AM
scoavoux.bsky.social
4.5M children, 10M adults will die from the policies of someone who claims "I’ve never physically hurt anyone"
Article abstract screen shot: "Forecasting models predicted that the current steep funding cuts could result in more than 14 051 750 (uncertainty interval 8 475 990–19 662 191) additional all-age deaths, including 4 537 157 (3 124 796–5 910 791) in children younger than age 5 years, by 2030."
July 1, 2025 at 10:00 AM
scoavoux.bsky.social
'Another article titled “Should I Stay or Should I Go?” or, the mass production of academic research titles' www.tandfonline.com/doi/full/10....
May 12, 2025 at 12:28 PM
scoavoux.bsky.social
si on se met à écouter les recommendations, au bout d'un mois, les préférences des usagers sont très déformées, au détriment des français/québecois (la part de chansons locales décroît); aux USA, les robots qui ne veulent pas de musique locale finissent par en écouter beaucoup au bout d'un mois. 3
même graphique que le skeet précédent. graphique montrant l'évolution de la part de contenus locaux recommandé par pays et par condition de l'expérience. Les conditions: les robots arrivent avec une proportion de préférence pour la musique locale (définie comme produite dans le pays et chanté dans la langue vernaculaire principale, français en France et au Québec, anglais aux US). Pour chaque pays et préférence, le graph montre deux courbes: l'une pour les robots qui n'écoutent pas de recommendations, l'autre pour ceux qui le font. Le résultat principal est décrit dans les skeets. On note que les robots qui écoutent les titres recommandés se voient proposer de moins en moins de chansons locales en France et au Quebec; aux USA, le déclin est moins prononcé. Les robots dont la préférence locale est nulle (ne veulent pas de chansons locales) voient leur préférence respectée en France et au Quebec, mais se voient proposer des chansons locales aux USA et finissent par en écouter à hauteur de 20% au bout d'un mois.
April 8, 2025 at 7:51 AM
scoavoux.bsky.social
Le résultat? Les algos respectent plutôt les préférences des usagers au départ, et étonnamment mieux en France et au Quebec qu'aux USA (plus de musique française/québecoise et francophone pour les français/québecois qui en veulent que de musique anglaise et US pour les américains). MAIS... 2
graphique montrant l'évolution de la part de contenus locaux recommandé par pays et par condition de l'expérience. Les conditions: les robots arrivent avec une proportion de préférence pour la musique locale (définie comme produite dans le pays et chanté dans la langue vernaculaire principale, français en France et au Québec, anglais aux US). Pour chaque pays et préférence, le graph montre deux courbes: l'une pour les robots qui n'écoutent pas de recommendations, l'autre pour ceux qui le font. Le résultat principal est décrit dans les skeets. On note que les robots qui écoutent les titres recommandés se voient proposer de moins en moins de chansons locales en France et au Quebec; aux USA, le déclin est moins prononcé. Les robots dont la préférence locale est nulle (ne veulent pas de chansons locales) voient leur préférence respectée en France et au Quebec, mais se voient proposer des chansons locales aux USA et finissent par en écouter à hauteur de 20% au bout d'un mois.
April 8, 2025 at 7:51 AM
scoavoux.bsky.social
arxiv.org/abs/2409.14202
Abstract of paper titled Mining Causality: AI-Assisted Search for Instrumental Variables. The abstract reads: The instrumental variables (IVs) method is a leading empirical strategy for causal inference. Finding IVs is a heuristic and creative process, and justifying its validity--especially exclusion restrictions--is largely rhetorical. We propose using large language models (LLMs) to search for new IVs through narratives and counterfactual reasoning, similar to how a human researcher would. The stark difference, however, is that LLMs can dramatically accelerate this process and explore an extremely large search space. We demonstrate how to construct prompts to search for potentially valid IVs. We contend that multi-step and role-playing prompting strategies are effective for simulating the endogenous decision-making processes of economic agents and for navigating language models through the realm of real-world scenarios. We apply our method to three well-known examples in economics: returns to schooling, supply and demand, and peer effects. We then extend our strategy to finding (i) control variables in regression and difference-in-differences and (ii) running variables in regression discontinuity designs. Simpsons character Ralph Wiggum chuckling "I'm in danger" with a "econometricians" as a title.
April 7, 2025 at 3:35 PM
scoavoux.bsky.social
Dans le même numéro qui fait sa une sur les nouveaux camions de nettoyage à Lausanne on trouve une chronique qui rit de la mobilisation contre les féminicides parce qu'ils n'ont fait "que" 19 mortes en Suisse en 2024 et qu'il faudrait parler des vrais sujets importants.
Chronique prétendument satirique qui se moque de ce que les féministes dénoncent les féminicides alors qu'il y a d'autres causes de mortalité plus importante.
March 17, 2025 at 8:47 AM
scoavoux.bsky.social
Last version of our forthcoming paper on why we cannot use large language models to answer surveys as a diverse sample -- and it's not what you think. w/ Julien Boelart @ppraeg.bsky.social @eollion.bsky.social and Ivaylo Petev @crestsociology.bsky.social osf.io/preprints/so...
Title page for Machine Bias. How Do Generative Language Models Answer Opinion Polls? Abstract: Generative AI is increasingly presented as a potential substitute for humans, including as human research subjects in various disciplines. Yet there is no scientific consensus on how closely these in-silico clones could represent their human counterparts. While some defend the use of these “synthetic users,” others point towards the biases in the responses provided by the LLMs. Through an experiment using survey questionnaires, we demonstrate that these latter critics are right to be wary of using generative AI to emulate respondents, but probably not for the right reason. Our results i) confirm that to date, models cannot replace research subjects for opinion or attitudinal research; ii) that they display a strong bias on each question (reaching only a small region of social space); and iii) that this bias varies randomly from one question to the other (reaching a different region every time). Besides the two existing competing theses (“representativity” and “social bias”), we propose a third one, which we call call “machine bias”. We detail this term and explore its consequences, for LLM research but also for studies on social biases.
March 4, 2025 at 6:48 AM
scoavoux.bsky.social
60% of CO2 emissions come from autocracies, relative to 40% in 1992. There are of course many confounders, this is not causal... But the illiberal turn in the western world is currently accompanied with policies that will make the earth unsuitable for humans journals.sagepub.com/doi/full/10....
plot disaggregating the total CO2 emissions by year by the nature of countries' political regime. Left hand side: V-dem index shows a ris from closed autocracy and electoral autocracy regimes from about 40% to about 60% of total emissions from 1992 to 2022. Right hand side, Freedom House Index, documents a rise of "not free" political regimes from 17% to 49% in the same time frame.
January 29, 2025 at 8:32 AM
scoavoux.bsky.social
freedom of speech
Elon Musk doing a nazi salute
January 21, 2025 at 6:32 AM
scoavoux.bsky.social
Screenshot of text: When the printing press was introduced to the masses around the sixteenth century, Erasmus of Rotterdam warned that all the new books would distract people from what was really important in life – reading Aristotle (Erasmus & Barker, 2001).
January 20, 2025 at 12:34 PM
scoavoux.bsky.social
je ne suis pas sûr que les chiffres pointent clairement une augmentation récente. fr.wikipedia.org/wiki/Rentr%C...
Wikipedia: Lors de la rentrée littéraire de septembre 2014, 607 livres sont sortis, c'est moins bien que la moyenne de 2005 à 2012 qui porte le chiffre à 676. En septembre 2012, 646 ouvrages de fiction étaient publiés. En janvier 2013, 525 autres ouvrages s'y ajoutaient. Les rentrées littéraires de septembre 2011 et 2010 comptaient respectivement 654 et 711 nouveaux ouvrages. Bien qu'en baisse par rapport au début des années 2010, les chiffres de 2021 avec 521 ouvrages retrouvaient les niveaux pré-COVID (524 ouvrages en 2019) après une baisse en 2020 (511 ouvrages). Toutefois, 2022 a amené une nouvelle fois une baisse avec le plus bas niveau en 20 ans avec 490 ouvrages publiés.
January 20, 2025 at 8:38 AM
scoavoux.bsky.social
"Denial of a wanted abortion has both immediate and lasting detrimental
effects." uc26d743563e946263a301a7d5d9.dl.dropboxusercontent.com/cd/0/inline2...
Denial of a wanted abortion has both immediate and lasting detrimental effects. It increases a woman’s risk of death by 2.5 p.p within nine months, mainly due to unsafe abortion procedures, and raises the likelihood of carrying the pregnancy to term by 31 p.p. Tracking outcomes up to 15 years later, we find that women denied an abortion experience more health issues, lower educational attainment, reduced labor-force participation, and higher rates of single motherhood, poverty, and reliance on government assistance. Existing children, born before their mother sought an abortion, are less likely to attend school and more likely to engage in child labor.
November 22, 2024 at 2:26 PM
scoavoux.bsky.social
Living in a place with fewer restriction on teenage abortion improves women later life outcomes doi.org/10.1177/0003...
We find that women who lived in a location with fewer abortion restrictions in adolescence, and women who had an abortion, compared to a live birth, in adolescence, are more likely to have graduated from college, have higher incomes, and have greater financial stability at two time-points over an almost 25-year period. Our results provide evidence that policy environments allowing access to abortion, and teenagers having the option to use abortion to avoid early parenthood, are important axes along which women’s economic lives are shaped. Our research implies that the widespread abortion bans and restrictions in the United States are likely to lead to lower educational attainment and adult economic stability among women living under such restrictions, as compared to women in locations with better access to abortion.
November 11, 2024 at 11:10 AM
scoavoux.bsky.social
Vous êtes bien bêtes de regarder des séries et de croire que vous vous amusez alors qu'en fait vous ne faites qu'alimenter le capitalisme et ne vous amusez pas vraiment vous auriez pu passer ce temps à lire Guy Debord bande de nullos.
laviedesidees.fr/La-productio...
Citation: La critique du temps passé à enchaîner les épisodes à partir de la notion de « travail abstrait » (p. 75) chez Marx, qui fait à la fois sens du « temps d’attention » (p. 75) extorqué à l’utilisateur et du dégoût ressenti quand il estime avoir consacré trop de temps à une série qui n’en valait pas la peine, est particulièrement féconde. Elle dévoile en effet le décalage qui existe, d’une part, entre ce que croit faire le spectateur (se divertir) et ce qu’il fait vraiment (produire de la donnée), et d’autre part, entre le « capital chronométrique » (p. 72) investi et la satisfaction réelle éprouvée.
November 8, 2024 at 1:21 PM
scoavoux.bsky.social
In Denmark, rich people use libraries, even the 1% wealthiest osf.io/preprints/so...
June 5, 2024 at 2:59 PM
scoavoux.bsky.social
Sanctions imposed on people who receive unemployement benefits decrease employement probability and earnings in the long run. academic.oup.com/ser/advance-...
March 8, 2024 at 5:51 AM