Last week offered three scenes in which artificial intelligence made verification a pressing issue. In Geneva, CERN and Mistral AI signed a statement of intent to develop a foundation model for collider physics. In Brasília, Brazil’s Superior Electoral Court ordered Meta to treat both candidates in the presidential run-off equally, following a series of failures involving its platforms and AI assistant. And from 24 to 31 October, UNESCO will hold Global Media and Information Literacy Week, in which we at AIthropology Lab are taking part with a week of radio programming.
These are three very different settings: a laboratory, a court and the public. All three raise the same question: who has the data, time and tools needed to verify what an automated system does?
Geneva: a model for reading collisions
On 15 October, CERN and Mistral AI announced a statement of intent to establish a long-term strategic partnership to develop foundation models for scientific discovery [1]. At its centre is a pretrained Large Physics Model, described by CERN as the first foundation model for collider physics. It will be trained on billions of simulated proton–proton collisions at the Large Hadron Collider (LHC), and is intended to support tasks ranging from reconstructing collisions and identifying particles to flagging unexpected signals and assessing future detector designs [1]. According to a CERN spokesperson quoted by Reuters, the model will be called Flare and is intended to replace some of the specialised algorithms currently used to reconstruct and identify particles [2].
The novelty lies in the architecture. Machine learning has long formed part of the LHC’s reconstruction chain. A CERN spokesperson explained that the laboratory has used AI for around a decade to turn raw experimental data into representations of particle events, using different tools at successive stages [1] [2]. Flare aims to learn across that entire chain within a single model. The scale helps explain the ambition. CERN has already stored approximately one exabyte of experimental LHC data — one million terabytes — and expects the number of collisions to increase tenfold when the High-Luminosity LHC begins operating around 2030 [1].
Enrica Porcari, CERN’s Chief Information Officer, framed the partnership as a way to develop AI capabilities that could benefit the whole physics community and be shared [1]. In an opinion article published the same day in Le Monde, Mistral’s chief executive Arthur Mensch warned of Europe’s dependence if advanced AI models were left entirely to American and Chinese companies [22]. CERN’s announcement specifies neither a budget nor a delivery timetable or access conditions for the future model [1]. The partnership brings together the scientific and computing resources of an intergovernmental organisation and the model-training expertise of a private company. The terms governing access, use and the sharing of results have yet to be made public. In Autonomy, for whom? we examined Mistral’s promise of open weights. Here, the question returns on another scale.
Karin Knorr Cetina conducted ethnographic research into work in high-energy physics and described large experiments as collective arrangements in which authority and responsibility are distributed among hundreds of people. In her account, the knowing subject is no longer simply the individual scientist: the experiment itself becomes a continuously tested knowledge-producing arrangement [3] [4]. Flare would introduce into that community a system that learns across the entire reconstruction chain. Validating it will require researchers to be able to establish why the model identifies what it does, which data it was trained on, and under what rules it can be shared. Those conditions will depend on the agreement between CERN and Mistral, which has not yet been published.
Brasília: visibility before the court
Brazil will elect its president on 25 October in a run-off between Luiz Inácio Lula da Silva and Senator Flávio Bolsonaro. On 15 October, Justice Marco Aurélio Bellizze of the Superior Electoral Court (TSE) issued an interim order requiring Meta, which owns Facebook, Instagram and WhatsApp, to treat the two campaigns equally [5]. The court upheld the order unanimously the following day [6] [7]. The measure followed a complaint from Lula’s coalition, which alleges restrictions on the reach of its posts and different treatment from that given to its opponent.
Meta acknowledges four incidents, which it attributes to separate causes. Lula’s campaign advertising account was mistakenly disabled for a little over an hour on 23 September. On 7 October, a technical fault caused Reels view counters to freeze or display zero in more than a hundred countries. On 9 October, a code update temporarily prevented users from boosting existing Instagram posts; according to the company, the problem affected thousands of advertisers across multiple countries [8] [9]. The fourth incident came on 14 October. In tests circulated that day, Meta AI on WhatsApp generated campaign posters using “Flávio 22” but refused the equivalent request using “Lula 13” [5]. Meta attributes this to a gap in its safeguards: they blocked the candidates’ full names but could miss a first name combined with an electoral number. It says the same was possible with “Luiz” and the number 13 [8].
On 16 October, Folha de S.Paulo reported internal messages in which Meta staff acknowledged restrictions affecting Lula’s Instagram accounts, while the platform was recommending Flávio Bolsonaro’s profile and WhatsApp channel [10]. The newspaper stressed that those conversations contained no admission of electoral interference. They documented problems, concern among staff about their seriousness, and delays in resolving them [8].
The court’s order bars Meta from applying unequal restrictions to the distribution, search visibility or metrics of content associated with either candidate. Restrictions on advertising accounts must undergo human review, or be reviewed within two hours if imposed automatically; Meta AI must also apply the same criteria to both campaigns. By 21 October, the company must restore paid promotion purchased by Lula’s campaign but not delivered. Non-compliance carries a daily fine of 500,000 Brazilian reais [8]. The court also ordered the blocking of electoral advertisements placed by unauthorised advertisers rather than candidates, parties, federations or coalitions [7] [9]. Its decision is preventive: it does not conclude that Meta deliberately manipulated content distribution [8]. Meta denies favouring either side [9].
Both campaigns have taken the dispute into other arenas. On 16 October, Flávio Bolsonaro’s campaign applied to join the proceedings as an interested third party, accusing Meta of negligence over paid negative adverts directed against its candidate [8]. On Saturday 17 October, Lula’s coalition asked electoral prosecutors and the Federal Police to investigate whether its rival had received an undeclared economic advantage and brought the matter to a United Nations special rapporteur [8] [11]. It assigned an indicative value of 7.44 million reais to the views it considered excessive; this is a hypothetical estimate, not an identified payment [8].
Google has not sold political advertisements in Brazil since May 2024 [12], and Diário do Nordeste describes Meta’s central position in the country’s paid digital election advertising [13]. A campaign’s visibility therefore depends to a significant extent on a private company. Letícia Cesarino, an anthropologist at the Federal University of Santa Catarina, has studied how digital platforms have reorganised Brazilian politics. In O mundo do avesso (2022), she examines algorithmisation and populism, arguing that understanding contemporary disinformation and populism requires attention to the technical infrastructure that supports them [14] [15].
The case illustrates how difficult it is to verify a platform’s behaviour from the outside. Establishing whether differences in reach stem from errors, design choices or audience behaviour requires data controlled by Meta. The TSE had to ask for those data, and Bellizze noted that the company’s technical explanations did not compare the two candidates’ situations [16]. Independent research is also taking place in Brazil. NetLab at the Federal University of Rio de Janeiro identified 1,613 political and electoral adverts placed between July and September from outside Brazil by unauthorised actors, paid for in foreign currencies, predominantly US dollars. The researchers documented how Meta reclassified many of the adverts as political content only after they had reached substantial audiences and generated revenue [17].
24–31 October: reading the world
UNESCO will mark Global Media and Information Literacy Week from 24 to 31 October under the theme We All Play a Part – Media and Information Literacy for the People. This year’s programme is decentralised and participatory, with regional and thematic events organised by UNESCO offices and partner organisations in Cairo, Nairobi, New Delhi, Bishkek, Abuja and Montreal, among other cities [18].
The calendar of registered activities shows how media literacy takes on different meanings in different places [19]. In Rio de Janeiro, MultiRio will launch the new ANDAR platform on 3 November, as part of the wider programme of activities: a student news agency that uses journalism by pupils in public schools as an educational tool. In Nairobi, JUKWAAA LA WASOMI plans to train student leaders from universities across Kenya ahead of the country’s 2027 general election. In Goma, Kinshasa and other cities in the Democratic Republic of the Congo, Soma Media Lab is organising workshops involving young people, students and people with disabilities. In Lima, the UNESCO-recognised organisation A Mí No Me La Hacen is presenting its sixth symposium as a live programme intended to keep circulating on video. In Madrid, Spain’s National Commission on Markets and Competition (CNMC) will hold the first national forum on media and information literacy on 27 October, focusing on public policy [19].
We at AIthropology Lab are taking part with AIthropology FM: Thinking Critically with AI, a special week of programming on our online radio station, listed in the Spain section of UNESCO’s calendar [19]. It will bring together rights-cleared talks, interviews, research-based audio, commentary and curated resources on AI literacy, synthetic information, algorithmic mediation, education, culture and the social consequences of generative AI. Like the other listed activities, ours is independently organised, and we remain responsible for its content [19].
In 1984, Brian Street distinguished between an autonomous model of literacy, treating it as a set of technical skills detached from context, and an ideological model, understanding it as a social practice shaped by power relations [20]. UNESCO’s calendar makes more sense through the latter lens. In Goma, Nairobi or Rio, learning to verify an image depends on who pays for an internet connection, which language information circulates in, and which institutions command trust. Paulo Freire put it in a phrase that remains useful, speaking at the opening of the Brazilian Congress of Reading in Campinas in November 1981: “reading the world precedes reading the word” [21]. With systems capable of generating images and deciding what becomes visible, reading the world now also means asking who controls those systems.
Who gets to verify?
The three scenes reveal an unequal distribution of the capacity to verify. At CERN, the validation of Flare will depend on access conditions that have yet to be announced. In Brazil, a court needed data from the company under scrutiny in order to decide, while university laboratories attempt to audit platforms using the information they can obtain. During Global Media and Information Literacy Week, verification is presented as a task for everyone, although the material conditions are very different in Goma, Madrid and Rio.
When an automated system intervenes in science, an election or what someone believes they have seen, the anthropological question includes who can examine it, with what resources, and with what authority. We will follow Brazil’s presidential run-off on 25 October and any access conditions published for Flare. From 24 to 31 October, we will be broadcasting on AIthropology FM.
References
[1] CERN. “CERN and Mistral join forces to accelerate AI for science”. 15 October 2026. https://home.cern/cern-and-mistral-join-forces-to-accelerate-ai-for-science/
[2] Marchandon, Leo. “CERN and Mistral partner to build Flare, a new AI model for physics”. Reuters, 15 October 2026. https://www.reuters.com/science/cern-mistral-partner-build-ai-foundation-models-physics-2026-10-15/
[3] Knorr Cetina, Karin. Epistemic Cultures: How the Sciences Make Knowledge. Harvard University Press, 1999. ISBN 978-0-674-25894-5. https://books.google.com/books?vid=ISBN9780674258945
[4] Giere, Ronald N. “Discussion Note: Distributed Cognition in Epistemic Cultures”. Philosophy of Science, 69(4), pp. 637–644, 2002. https://doi.org/10.1086/344627
[5] SBT News. “TSE determina que Meta dê mesmo tratamento a Lula e Flávio”. 15 October 2026. https://sbtnews.sbt.com.br/noticia/eleicoes/tse-determina-que-meta-de-mesmo-tratamento-a-lula-e-flavio
[6] Gazeta do Povo. “TSE confirma decisão que mandou Meta garantir igualdade entre Flávio e Lula nas redes”. 16 October 2026. https://www.gazetadopovo.com.br/eleicoes/2026/tse-confirma-decisao-que-mandou-meta-garantir-igualdade-entre-flavio-e-lula-nas-redes/
[7] Folha de S.Paulo. “TSE confirma decisão que obriga Meta a dar tratamento igual a Lula e Flávio em suas plataformas”. 16 October 2026. https://www1.folha.uol.com.br/poder/2026/10/tse-confirma-decisao-que-obriga-meta-a-dar-tratamento-igual-a-lula-e-flavio-em-suas-plataformas.shtml
[8] Vitorio, Tamires. “Lula x Flávio Bolsonaro: entenda as acusações contra a Meta às vésperas do 2º turno”. Exame, 17 October 2026. https://exame.com/brasil/lula-x-flavio-bolsonaro-entenda-as-acusacoes-contra-a-meta-as-vesperas-do-2o-turno/
[9] Meta. “A Meta não está interferindo nas eleições brasileiras. Estes são os fatos”. 16 October 2026. https://about.fb.com/br/news/2026/10/a-meta-nao-esta-interferindo-nas-eleicoes-brasileiras-estes-sao-os-fatos/
[10] Reuters. “Meta employees said Lula’s accounts were hindered during Brazil election, Folha reports”. 16 October 2026. https://www.reuters.com/world/americas/meta-employees-said-lulas-accounts-were-hindered-during-brazil-election-folha-2026-10-16/
[11] SBT News. “Lula pede à ONU e à PGR que apurem atuação da Meta”. 17 October 2026. https://sbtnews.sbt.com.br/noticia/eleicoes/lula-pede-a-onu-e-a-pgr-que-apurem-atuacao-da-meta
[12] Meio & Mensagem. “Google proíbe anúncios de conteúdo político no Brasil”. April 2024. https://www.meioemensagem.com.br/midia/google-proibe-anuncios-de-conteudo-politico-no-brasil
[13] Diário do Nordeste. “Após Meta admitir falhas que afetaram campanha de Lula, TSE mantém ordem para isonomia nas redes”. October 2026. https://diariodonordeste.verdesmares.com.br/pontopoder/apos-meta-admitir-falhas-que-afetaram-campanha-de-lula-tse-mantem-ordem-para-isonomia-nas-redes-1.3799365
[14] Cesarino, Letícia. O mundo do avesso: verdade e política na era digital. São Paulo: Ubu Editora, 2022. Bibliographic review: https://periodicos.unb.br/index.php/anuarioantropologico/article/view/52039
[15] MATRIZes. Review of O mundo do avesso by Letícia Cesarino. MATRIZes, 18(1), pp. 295–302, 2024. https://doi.org/10.11606/issn.1982-8160.v18i1p295-302
[16] Borges, Gabriel Ferreira. “TSE referenda liminar de Bellizze e manda Meta dar tratamento igualitário a Flávio Bolsonaro e Lula”. O Tempo, 16 October 2026. https://otempo.com.br/eleicoes/2026/2026/10/16/tse-referenda-liminar-de-bellizze-e-manda-meta-dar-tratamento-igualitario-a-flavio-bolsonaro-e-lula
[17] NetLab UFRJ. “Meta lucra com anúncios irregulares promovidos por redes coordenadas para influenciar as eleições no Brasil”. 15 October 2026. https://netlab.eco.ufrj.br/post/meta-lucra-com-an%C3%BAncios-irregulares-promovidos-por-redes-coordenadas-para-influenciar-as-elei%C3%A7%C3%B5es-no
[18] UNESCO. “Global Media and Information Literacy Week 2026”. https://www.unesco.org/en/articles/global-media-and-information-literacy-week-2026
[19] UNESCO. “Global MIL Week 2026 – Celebrations around the World”. Updated 16 October 2026. https://www.unesco.org/en/weeks/media-information-literacy/around-world-2026
[20] Street, Brian V. Literacy in Theory and Practice. Cambridge University Press, 1984. https://books.google.com/books/about/Literacy_in_Theory_and_Practice.html?id=R0UdWQ5thf8C
[21] Freire, Paulo. A importância do ato de ler. Delivered at the opening of the Brazilian Congress of Reading, Campinas, November 1981. Published in São Paulo by Cortez in 1982 and later editions. Audio edition in the Paulo Freire Institute archive: https://acervo.paulofreire.org/items/32a70d70-8cd8-42a2-a15e-51a2a5fdfb45
[22] Mensch, Arthur. “Laisser les modèles d’IA aux Américains et aux Chinois placerait l’Europe dans une dépendance sans équivalent dans l’histoire des civilisations”. Le Monde, 15 October 2026. https://www.lemonde.fr/idees/article/2026/10/15/arthur-mensch-directeur-general-de-mistral-ai-laisser-les-modeles-d-ia-aux-americains-et-aux-chinois-placerait-l-europe-dans-une-dependance-sans-equivalent-dans-l-histoire-des-civilisations_6794078_3232.html