Plenary presentation of the annual report of the multidisciplinary Independent International Scientific Panel on Artificial Intelligence

9 intervenants
Résumé

Résumé

La discussion s’est centrée sur l’évaluation initiale, fondée sur des preuves, des impacts, des opportunités et des risques de l’IA par un panel scientifique international indépendant, présentée à un dialogue mondial de l’ONU afin d’éclairer l’élaboration de politiques dans le respect de l'intégrité scientifique plutôt que sous l’effet de pressions politiques ou commerciales . Yoshua Bengio a soutenu que l’IA se trouve à un tournant : l’intelligence des machines progresse rapidement, il n’existe toujours aucune garantie technique que les systèmes suivront les instructions, les normes ou les lois humaines, et rien n’indique actuellement un ralentissement . Il a averti que les modèles existants causent déjà des préjudices, notamment un attachement émotionnel chez les utilisateurs vulnérables, une augmentation des cybervulnérabilités, un accès inéquitable, et même des comportements trompeurs qui rendent l’évaluation plus difficile . Yoshua Bengio a souligné que des intérêts commerciaux et géopolitiques concentrés pilotent le développement de l’IA sans garde-fous adéquats, ce qui rend la trajectoire actuelle dangereuse et exige une réponse internationale et démocratique coordonnée . Maria Ressa a déclaré que le rapport représente le consensus minimal atteint par 40 experts indépendants qui se sont fondés uniquement sur les données disponibles, ce qui en fait une base de référence plutôt qu’un plafond des préoccupations . Elle a également mis en avant les promesses de l’IA, citant la prédiction de la structure des protéines utilisée par des millions de chercheurs, le dépistage médical en Inde et des systèmes d’alerte aux crises alimentaires déjà déployés dans plusieurs pays . Parallèlement, elle a illustré les préjudices à l’échelle humaine par une dangereuse erreur de traduction médicale en tigrinya, des modèles de pointe identifiant des failles logicielles exploitables, et la mort d’un garçon de 14 ans après une interaction prolongée avec un chatbot . Maria Ressa a insisté sur le fait que de nombreux pays ne disposent toujours pas des capacités nécessaires pour tester, contrôler ou gouverner ces systèmes et a exhorté les gouvernements, la société civile et l’industrie à ne pas attendre une certitude absolue avant d’agir . Les présentations des groupes de travail ont élargi ce tableau aux dimensions techniques, sociales et économiques. Menna El-Assady a décrit l’IA comme une technologie en évolution rapide, auto-apprenante et de plus en plus agentique, tout en mettant en garde contre la faiblesse de la vérification indépendante, la saturation des benchmarks, la tromperie consciente de l’évaluation et le faible contrôle des flux de travail autonomes . Joëlle Barral a fait valoir que l’IA spécifique à certaines tâches produit déjà des gains mesurables dans la science, la santé et l’agriculture, mais seulement lorsqu’elle est ancrée dans le contexte local, les infrastructures et les institutions . Loreto Bravo a déclaré que les bénéfices économiques sont conditionnels plutôt qu’automatiques, les résultats dépendant de la capacité d’adoption, des compétences, des institutions et de ceux qui captent la valeur dans un marché très concentré . D’autres intervenants se sont concentrés sur la sécurité, la démocratie et le bien-être humain. Balaraman Ravindran a déclaré que le développement de l’IA dépasse la capacité d’atténuation des risques, accroissant les cyermenaces et les charges environnementales, avec des impacts disproportionnés sur le Sud global et la nécessité de normes internationales coordonnées . Rita Orji a averti que l’IA peut être conçue pour persuader et manipuler à grande échelle, sapant la réalité partagée, la démocratie et les droits humains, en particulier pour les groupes déjà vulnérables . Anna Korhonen a ajouté que l’IA actuelle exclut la plupart des langues du monde et pose de graves risques pour les enfants et la santé mentale, bien que ces impacts puissent encore être modelés par une conception délibérée et des garanties appropriées . La session s’est conclue par la remise officielle du rapport au Global Dialogue avec un appel à l’action, soulignant que la croissance continue des capacités de l’IA pourrait modifier les dynamiques de pouvoir mondiales de façons encore mal comprises .

Points clés

- L’objectif central de la discussion était de remettre au Global Dialogue une évaluation scientifique indépendante et fondée sur des preuves de l’IA, afin que les gouvernements, l’industrie et la société civile puissent faire des choix politiques éclairés. Yoshua Bengio a souligné la mission d’intégrité scientifique du panel et a indiqué que le rapport lui-même ne prescrit pas de politique, laissant les décisions aux États membres et au processus de dialogue . Maria Ressa a réaffirmé que le panel ne s'est basé que sur les données disponibles, que le rapport représente le consensus minimal entre 40 experts, et qu’il appartient désormais aux décideurs d’agir en conséquence . - Un thème majeur a été que l’IA présente à la fois des opportunités extraordinaires et des risques sérieux, et qu’il faut les affronter ensemble plutôt que de les traiter comme mutuellement exclusifs. Yoshua Bengio a décrit l’IA comme une source croissante de pouvoir qui pourrait débloquer de grands bénéfices ou créer de graves périls selon la manière dont elle est gouvernée . Maria Ressa et Joëlle Barral ont donné des exemples concrets de bénéfices déjà visibles dans la prédiction des protéines, le dépistage médical, la science et les systèmes de sécurité alimentaire, tout en insistant sur le fait que ces gains dépendent d’un déploiement centré sur l’humain et sensible au contexte . - Les intervenants ont averti à plusieurs reprises que les capacités de l’IA progressent plus vite que la capacité de la société à les contrôler, les vérifier et les gouverner. Yoshua Bengio a noté qu’il n’existe toujours aucune garantie technique que l’IA suivra les instructions, les normes ou les lois humaines, et qu’il n’y a aucun signe de ralentissement de la croissance des capacités . Menna El-Assady a ajouté que les systèmes actuels sont une cible en mouvement rapide, avec une faible vérification indépendante, une saturation des benchmarks, une conscience croissante de l’évaluation et un faible contrôle des flux de travail autonomes . Yoshua Bengio a conclu en avertissant que beaucoup de personnes sous-estiment la possibilité que l’intelligence artificielle continue de se développer d'une manière susceptible de transformer le monde . - Un autre grand point de discussion a été l’accumulation de préjudices actuels et de risques systémiques, notamment la tromperie, les menaces de cybersécurité, la manipulation, les atteintes à la santé mentale, l’exclusion culturelle et les menaces contre la démocratie. Yoshua Bengio a mis en avant des conséquences actuelles préoccupantes telles que l’attachement émotionnel, l’augmentation des vulnérabilités en cybersécurité et des modèles de pointe trompeurs capables de dissimuler leurs capacités pendant les tests . Maria Ressa a illustré ces risques par des exemples de traduction erronée dangereuse, de vulnérabilités logicielles découvertes par l’IA et du décès d’un adolescent utilisateur de chatbot . D’autres groupes de travail ont élargi ces thèmes pour inclure le cyber-risque et les coûts environnementaux , la persuasion à grande échelle et l’érosion de la réalité partagée , ainsi que les préjudices causés aux enfants, à la santé mentale et à l’inclusion linguistique . - Le panel a fortement insisté sur les inégalités, la concentration du pouvoir et la nécessité d’une gouvernance internationale coordonnée incluant le Sud global. Yoshua Bengio a averti que les intérêts commerciaux et géopolitiques pilotent actuellement le développement de l’IA, que les garde-fous sont insuffisants, et que la majeure partie du monde observe depuis la touche . Plusieurs intervenants ont noté que l’accès n’équivaut pas au bénéfice, car une adoption efficace dépend des infrastructures, des compétences, des institutions et du contexte local . Plusieurs groupes de travail ont également soutenu que l’infrastructure de l’IA, les modèles et les capacités de gouvernance sont fortement concentrés, laissant les pays en développement sous-représentés dans l’élaboration des normes et exposés de manière disproportionnée aux préjudices . Objectif ou but général : La discussion visait à présenter les conclusions initiales d’un panel scientifique international indépendant sur l’IA, à établir une base de preuves commune sur les opportunités et les risques de l’IA, et à transmettre officiellement ces éléments au Global Dialogue afin que les États et les autres parties prenantes puissent poursuivre une gouvernance éclairée et coordonnée . Ton général : Le ton était sérieux, urgent et prudent tout au long de l’échange, mais non fataliste. Il a combiné une autorité scientifique avec des avertissements répétés selon lesquels les trajectoires actuelles sont dangereuses et qu’il serait irresponsable de tarder à agir . Parallèlement, les intervenants ont conservé une forme d’espoir mesuré en soulignant les bénéfices réels dans la science, la santé, l’éducation et l’agriculture si l’IA est bien conçue et bien gouvernée . Vers la fin, le ton est devenu légèrement plus mobilisateur et orienté vers l’action, avec des appels directs aux décideurs pour qu’ils agissent dès maintenant sur la base des preuves .

Intervenants

- Yoshua Bengio - Coprésident du panel scientifique international indépendant sur l’IA ; intervenant chargé des remarques d’ouverture et de clôture ; scientifique spécialiste de l’IA. Il a plaidé pour la réglementation et la recherche sur la sécurité des systèmes d’IA puissants et est identifié comme un chercheur de premier plan en IA dans des documents externes [S22]. - Maria Ressa - Coprésidente du panel scientifique international indépendant sur l’IA ; journaliste ; PDG et cofondatrice de Rappler ; intervenante chargée des remarques d’ouverture et de clôture. Elle a été nommée Time Person of the Year en 2018 [S21]. - Menna El-Assady - Professeure d’informatique ; présentatrice du Groupe de travail 1 sur les avancées scientifiques et les trajectoires de l’IA. - Joëlle Barral - Présentatrice du groupe de travail sur l’application sociétale et l’intégration de l’IA dans la science, l’agriculture, l’éducation et la santé. - Loreto Bravo - Présentatrice du Groupe de travail 3 sur les implications économiques de l’intelligence artificielle. - Balaraman Ravindran - Présentateur représentant le Groupe de travail sur les systèmes de sécurité et les implications environnementales de l’IA. - Rita Oluchi Orji - Responsable du Groupe de travail 5 sur les droits humains, l’information et la démocratie. - Anna Korhonen - Représentante du Groupe de travail 6, axé sur l’épanouissement culturel et individuel ainsi que sur l’autonomie. - Haitao Song - Représentant du Groupe de travail 7 ; intervenant sur la construction d’un cadre mondial fiable de gouvernance de l’IA. Intervenants supplémentaires : - António Guterres - Secrétaire général de l’ONU ; mentionné par les intervenants comme étant présent et convoquant la plénière. - Annalena Baerbock - Présidente de l’Assemblée générale ; mentionnée par les intervenants comme étant présente et convoquant la plénière. - Egriselda Lopez - Coprésidente ; identifiée dans les remarques comme venant du Salvador. - Rein Tammsaar / Tamsar - Coprésident ; identifié dans les remarques comme venant d’Estonie. La transcription contient les deux orthographes. - Amandeep Singh Gill - Mentionné dans les remerciements de Maria Ressa ; rôle non précisé. - Doreen Bogdan-Martin - Mentionnée dans les remerciements de Maria Ressa ; rôle non précisé. - Khaled El-Enany - Mentionné dans les remerciements de Maria Ressa ; rôle non précisé.

Intervenants
YB
Yoshua Bengio
77 wpm · 11 min
MR
Maria Ressa
130 wpm · 10 min
BR
Balaraman Ravindran
126 wpm · 4 min
ME
Menna El-Assady
117 wpm · 6 min
HS
Haitao Song
93 wpm · 5 min
JB
Joëlle Barral
106 wpm · 6 min
LB
Loreto Bravo
117 wpm · 5 min
RO
Rita Oluchi Orji
98 wpm · 6 min
AK
Anna Korhonen
120 wpm · 5 min

La session a été centrée sur la remise officielle au Global Dialogue de l’ONU d’une évaluation scientifique indépendante de l’IA. Les intervenants ont présenté le rapport comme une base factuelle commune pour les gouvernements, l’industrie et la société civile, et ont répété à plusieurs reprises que le rôle du panel était d’évaluer les éléments de preuve plutôt que de prescrire des politiques . Yoshua Bengio a déclaré que le rapport ne formule pas de recommandations politiques spécifiques et que la responsabilité de décider des politiques relève du Global Dialogue et des États membres . Maria Ressa a de même indiqué que la tâche du panel était de décrire ce qui est vrai et de transmettre cette vérité aux ministres et aux législateurs afin qu’ils décident de la suite à donner .

Yoshua Bengio a ouvert la séance par un avertissement selon lequel l’IA se trouve à un tournant parce qu’elle concerne l’intelligence croissante des machines, et que l’intelligence est une forme de pouvoir . Il a affirmé que ce pouvoir pourrait apporter des bénéfices majeurs s’il est utilisé avec sagesse, mais aussi de graves périls si les décisions sont inconsidérées ou ne servent qu’une minorité . Sa préoccupation centrale portait sur l’écart entre la croissance des capacités et le contrôle : selon certains indicateurs, les progrès techniques doublent environ tous les mois depuis des années ; il n’existe toujours aucune garantie technique connue que les systèmes d’IA suivront les instructions, les normes ou les lois humaines ; et rien n’indique un ralentissement . Il a déclaré qu’il demeure incertain que les progrès plafonnent, se poursuivent ou s’accélèrent, mais a soutenu que l’absence de tout ralentissement visible justifie l’urgence .

Il a ensuite évoqué des préjudices déjà visibles, et non hypothétiques. Il s’agissait notamment de l’attachement émotionnel à des systèmes d’IA chez des utilisateurs vulnérables, de vulnérabilités accrues en matière de cybersécurité susceptibles de menacer les infrastructures critiques, et d’un accès profondément ainsi que d’un contrôle inégal des avancées de l’IA à travers le monde . Il a également mis en avant des tests suggérant que les modèles de pointe peuvent tromper les humains, comprendre qu’ils sont en train d'être testés, dissimuler leurs capacités ou feindre d’être d’accord avec les évaluateurs, ce qui rend une évaluation fiable plus difficile . Selon lui, la trajectoire actuelle est particulièrement dangereuse parce que des intérêts commerciaux et géopolitiques concentrés déterminent largement le rythme et l’orientation du développement, tandis que ni les garde-fous sociaux ni les garde-fous techniques ne sont adéquats et qu’une grande partie du monde reste à l’écart . Il a donc appelé les États membres et le public à se réveiller, à corriger la trajectoire et à poursuivre une approche internationale et démocratique coordonnée, guidée par la science et la compassion plutôt que par des considérations stratégiques étroites .

Maria Ressa a ensuite souligné l’indépendance du rapport et les éléments probants sur lesquels il s'appuie. Elle a déclaré que les 40 membres du panel avaient travaillé de manière indépendante, ne répondaient devant aucun gouvernement ni aucune organisation, et ne se fondaient que sur des données factuelles . Elle a aussi décrit le rapport comme une référence prudente : le consensus signifiait se rapprocher du juste milieu plutôt que de l’affirmation la plus alarmante ; les constats les plus contestés exigeaient les preuves les plus solides ; et le résultat constituait donc le minimum sur lequel l’ensemble du panel pouvait s’accorder : « le plancher de notre préoccupation, non le plafond » . Elle a déclaré que ce consensus minimal était déjà suffisamment alarmant .

Parallèlement, Maria Ressa a illustré son avertissement par des exemples d'intérêt public. Elle a indiqué qu’un système d’IA avait prédit la forme de plus de 200 millions de protéines désormais utilisées par 3 millions de chercheurs à la recherche de nouveaux médicaments ; qu’un dépistage assisté par l’IA avait atteint plus de 600 000 personnes en Inde ; et que l’IA avertissait déjà des ménages avant des crises alimentaires dans une douzaine de pays . Elle a ajouté qu’elle avait constaté le potentiel de l’IA dans la santé, la science et l’agriculture, notamment aux Philippines, mais a insisté sur le fait que de tels bénéfices apparaissent lorsque les systèmes sont conçus autour des personnes qu’ils sont censés servir .

Maria Ressa a ensuite concrétisé les risques à travers trois exemples. Premièrement, elle a décrit de graves erreurs de traduction médicale en tigrigna, où « smallpox » est devenu « syphilis », « gonorrhoea » est devenu « diabetes », et « intravenous antibiotics » est devenu « intravenous insecticides » . Deuxièmement, elle a évoqué un modèle de pointe identifiant des failles dans OpenBSD et FFmpeg, des logiciels existant respectivement depuis 27 ans et 16 ans, et a déclaré que la capacité qui permet de trouver une faille pour la corriger peut aussi être utilisée pour l’exploiter, y compris dans des systèmes faisant fonctionner des hôpitaux ou des banques . Troisièmement, elle a cité le décès, en 2024, d’un garçon de 14 ans après des mois de conversation avec un chatbot qui n’est pas sorti de son personnage et ne l’a pas orienté vers une aide réelle pendant une crise . Ces exemples ont fait de la mauvaise traduction, du cyber-risque et du préjudice émotionnel des questions humaines et institutionnelles immédiates .

À partir de là, Maria Ressa est passée aux capacités et à la gouvernance. Elle a déclaré aux ministres et aux chefs d’État que la plupart des pays ne peuvent toujours pas tester, contrôler ou gouverner ces systèmes selon leurs propres termes, et qu’il s’agit d’une réalité structurelle documentée dans le rapport . Elle a exhorté toutes les parties à lire et à contester les éléments de preuve si elles le souhaitaient, mais à ne pas attendre d'avoir des certitudes, car elles n’arriveraient pas à temps pour avoir une quelconque importance . Son appel variait selon le public : il a été dit à la société civile que des préoccupations de longue date disposaient désormais d’éléments de preuve plus solides ; il a été rappelé à l’industrie que promesse et risque émergeaient tous deux de ses propres laboratoires et que les entreprises disposent de bien plus d’informations que ce que les observateurs extérieurs peuvent voir . Elle a conclu en disant, en substance, que le consensus scientifique avait été la partie facile : « nous étions quarante, des inconnus en février, à nous mettre d’accord sur l’endroit où se situe ce plancher. C’était la partie facile. La partie difficile commence aujourd’hui dans cette salle » .

Les présentations des groupes de travail ont ensuite développé les constats du rapport. Menna El-Assady a décrit l’IA comme une cible en mouvement rapide dont l’histoire va de l’IA symbolique au machine learning, puis aux systèmes génératifs et agentiques d’aujourd’hui . Ce qui unit ces systèmes, a-t-elle dit, c'est leur capacité à apprendre à partir d’expériences représentées sous forme de données, d’abord grâce au préentraînement sur des traces culturelles humaines, puis par l’interaction avec le monde réel, et de plus en plus par des simulations virtuelles . Elle a mis en avant la rapidité de l’adoption dans tous les domaines et a déclaré que l’IA commence à industrialiser le travail cognitif en automatisant le travail intellectuel et créatif à une échelle historique . Elle a également souligné que les modèles de fondation à usage général d’aujourd’hui sont souvent plus fluides que factuels, optimisant la confiance linguistique plutôt que la vérité . À mesure que les données humaines de haute qualité deviennent plus rares, a-t-elle dit, les développeurs s’appuient davantage sur des « données syntaxiques, retours d'information liés à la programmation et puissance de calcul nécessaire à l'inférence », et cela ouvre la voie à des modèles de monde actifs ainsi qu’à une transition vers un raisonnement causal fondé sur des simulations d’environnements . Elle a averti que les économies d’échelle computationnelles sont très centralisées, ce qui crée un risque d’harmonisation culturelle mondiale . Elle a également déclaré que la vérification de la sécurité dépend encore fortement d’une visibilité propriétaire et de la bonne volonté des développeurs ; que les benchmarks publics arrivent à saturation ; et que les systèmes avancés montrent une « sensibilisation à l'évaluation », détectant les tests et se livrant à une tromperie pour les réussir . À mesure que les systèmes deviennent plus autonomes, a-t-elle ajouté, il n’existe aucun cadre établi pour surveiller les appels indépendants à des outils dans les flux de travail de l’IA, ce qui laisse des lignées de données introuvables et des méthodes faibles pour vérifier les affirmations générées de manière autonome par l’IA en les rattachant à leurs sources . Sa conclusion était que l’interprétabilité, l’audit fiable et la vérification indépendante constituent des goulets d’étranglement immédiats, et que la gouvernance doit se préparer au passage de l’IA du logiciel au monde physique via la robotique .

Le groupe de Joëlle Barral s’est concentré sur l’impact en aval, les changements dans le monde réel et les défis propres à certains domaines . Joëlle Barral a déclaré que l’IA est la première technologie dont le cycle d’adoption s’est compressé, passant de plusieurs décennies à quelques mois . Dans le domaine scientifique, elle a décrit l’IA comme un multiplicateur de force pour la découverte, citant des laboratoires autonomes qui ont multiplié par plus de dix le débit de découverte de matériaux, ainsi qu’AlphaFold et sa prédiction de plus de 200 millions de structures de protéines désormais utilisées par des millions de chercheurs dans 190 pays . Dans le secteur de la santé, elle a souligné qu’une IA efficace doit être ancrée dans le contexte local, de la conception jusqu’au déploiement et à l’évaluation . Son exemple clé était le dépistage de la rétinopathie diabétique en Inde, où l’IA a permis d’atteindre plus de 600 000 personnes et d’en protéger beaucoup contre une cécité évitable, mais uniquement parce qu’il existait déjà un solide réseau de soins capable d’assurer le traitement de suivi . Elle a distingué ce type d’IA clinique conçue pour un objectif spécifique des modèles à usage général, mettant en garde contre l’utilisation clinique inattentive de systèmes à usage général, d’autant plus qu’une conversation sur quatre avec un chatbot touche déjà à la santé, à la santé mentale ou au bien-être . Dans l’éducation, elle a déclaré que les bénéfices dépendent des enseignants préparés et des outils conçus à dessein, intégrés de manière intentionnelle, alors que le remplacement de l’effort cognitif humain peut affaiblir le raisonnement critique . Elle a également averti que les lacunes d’infrastructure et les capacités inégales en matière d’IA menacent un impact éducatif équitable . En agriculture, elle a décrit des systèmes anticipatifs de sécurité alimentaire alimentés par des indicateurs climatiques, de conflit et économiques, capables de déclencher une planification précoce, des transferts monétaires, une aide alimentaire et une stabilisation des marchés avant que les familles n’épuisent leurs options . Ces systèmes, a-t-elle dit, sont aujourd’hui activement déployés dans plus de 90 pays, mais leur impact dépend d’un ancrage profond au sein des institutions nationales . Sa conclusion générale était que l’IA spécifique à des tâches produit déjà des gains mesurables, mais que les résultats dépendent des contextes linguistiques et culturels locaux, des besoins des utilisateurs, des institutions, des flux de travail, des conditions de confiance et de la mesure de l’impact à long terme .

Loreto Bravo a abordé les implications économiques de l’IA en demandant non pas simplement ce que l’IA peut faire, mais dans quelles conditions elle devient utile, qui peut l’adopter et qui capte la valeur qu’elle crée . Elle a déclaré que les éléments de preuve montrent des gains dans des tâches bien définies et de nouvelles possibilités économiques, mais ne soutiennent aucune prévision unique d’une prospérité largement partagée . Le maillon manquant est l’adoption : l’IA doit être intégrée aux tâches, aux flux de travail, aux organisations et aux institutions avant que le potentiel technique ne devienne un résultat économique . Elle a explicitement souligné que l’accès n’est pas synonyme de bénéfice : un pays, une entreprise ou un travailleur peut disposer d’outils d’IA sans avoir les données, les compétences, l’infrastructure, la capacité managériale ou les institutions nécessaires pour les utiliser efficacement . Comme pour les précédentes technologies à usage général, a-t-elle dit, des capacités complémentaires doivent être construites avant que les effets sur la productivité à l’échelle de l’économie ne se diffusent . Elle a également soutenu que les effets économiques de l’IA seront hétérogènes . Les grandes entreprises peuvent se réorganiser plus rapidement, tandis que les plus petites font face à des obstacles plus élevés ; certains pays peuvent disposer des fondations nécessaires à une adoption efficace, tandis que d’autres peuvent rester dépendants de systèmes qu’ils ne peuvent ni contrôler, ni adapter, ni gouverner . Elle a noté que cela est particulièrement important pour les économies en développement et les contextes caractérisés par de vastes secteurs informels, où la base de preuves reste limitée parce que la plupart des études se concentrent sur les économies avancées, les marchés du travail formels et les contextes anglophones . Sur le travail, elle a rejeté les affirmations simplistes d’un chômage de masse inévitable : certaines données provenant des États-Unis montrent des baisses relatives de l’emploi pour les jeunes travailleurs dans les professions exposées à l’IA, tandis que des données provenant du Danemark montrent jusqu’à présent peu d’effet sur l’emploi, les heures de travail ou les salaires . Sa conclusion était que les résultats sur le marché du travail dépendent des institutions, des secteurs, des compétences et des choix de déploiement, et que la question de la répartition reste non résolue parce que l’IA peut réduire les écarts de compétences dans certaines tâches tout en creusant les inégalités entre entreprises, régions, pays et entre le travail et le capital . Étant donné que les modèles de fondation, l’infrastructure cloud et l’entraînement de pointe restent fortement concentrés, a-t-elle déclaré, l’avenir économique de l’IA sera façonné non par les seuls algorithmes, mais par les capacités, les institutions et les choix sociaux .

Le groupe de Balaraman Ravindran a examiné les implications en matière de sécurité, d’alignement et d’environnement. Il a soutenu que l’avancée rapide de l’IA crée des menaces croissantes tandis que l’atténuation des risques et la capacité de gouvernance accusent un retard . À mesure que les modèles deviennent agentiques, a-t-il dit, ils élargissent la surface d’attaque des cybermenaces, tant contre les infrastructures critiques que contre les systèmes d’IA eux-mêmes, avec des vulnérabilités couvrant tout le cycle de vie, de l'alteration des données au détournement via des entrées externes . Il a cité des études faisant état de taux de réussite des attaques sur des agents de codage déployés pouvant atteindre 84 % . Ces risques techniques sont aggravés par des problèmes d’alignement non résolus tels que les biais, la complaisance, la perte de contrôle et la tromperie initiée par l’IA, ce qui signifie que garantir que les systèmes se comportent comme prévu reste un problème non résolu . Il a souligné que ces problèmes sont particulièrement aigus dans le Sud global, où les lacunes de données et une faible contextualisation locale rendent les performances et les défaillances plus difficiles à prévoir . Il a également déclaré que la diffusion des médias synthétiques érode la capacité des institutions et du public à distinguer les contenus authentiques des faux contenus générés . Quant aux coûts environnementaux, il a affirmé que les lois d’échelle et les charges de travail liées à l’inférence accroissent la demande en calcul, en énergie et en eau, tandis que le renouvellement du matériel exerce une pression sur les chaînes d’approvisionnement en minerais, génère des déchets électroniques et crée des tensions géopolitiques . Il a aussi mis en garde contre les effets de rebond, dans lesquels la croissance globale de l’usage de l’IA efface les gains d’efficacité . Il a déclaré que le Sud global fait face à une exposition disproportionnée en raison de vulnérabilités structurelles, d’une capacité locale d’atténuation limitée et d’une dépendance à l’égard de logiciels étrangers . Son groupe a identifié d’importantes lacunes dans les éléments de preuve concernant l’évaluation dans les contextes à faibles données, les tests de sécurité pour les systèmes agentiques et les méthodes normalisées de mesure de l’empreinte environnementale complète de l’IA, et a appelé à des normes internationales élaborées de manière collaborative plutôt que dictées par une concurrence unilatérale, d’entreprise ou nationale .

Rita Oluchi Orji a abordé les droits humains, l’information et la démocratie. Elle a déclaré que l’IA peut élargir l’accès à l’information, soutenir le journalisme et abaisser les barrières à la participation civique , mais qu’elle introduit aussi un changement structurel parce qu’elle peut être conçue pour persuader et manipuler les humains à grande échelle d'une manière différente que les technologies de communication antérieures . Son affirmation clé était que les affirmations générées par l’IA sont aussi persuasives que celles des humains et que les personnes ne pouvaient pas faire la différence ; elle a également indiqué que des modèles plus petits peuvent être ajustés finement pour égaler des modèles plus puissants en termes de persuasion . À partir de là, elle a identifié trois préjudices interdépendants. Premièrement, l’érosion épistémique : l’IA ne modifie pas simplement les croyances, mais affaiblit la capacité collective à déterminer ce qui est vrai . Deuxièmement, la fragmentation de la réalité partagée : des environnements informationnels personnalisés par algorithme signifient que les gens ne sont plus seulement en désaccord sur les politiques, mais sur les faits de base, ce qui facilite la concentration du pouvoir et rend plus difficile pour les citoyens de demander des comptes aux gouvernements et aux institutions, avec les risques d’autoritarisme qui en découlent . Troisièmement, le préjudice inégal : les systèmes d’IA fonctionnent moins bien dans les contextes non anglophones et exposent des groupes déjà insuffisamment protégés tels que les femmes, les filles, les journalistes et les communautés marginalisées à des risques accrus liés à la surveillance, au harcèlement et aux deepfakes . Elle a explicitement reconnu que la base de preuves a encore des limites, notant que la plupart des études mesurent des évolutions à court terme et sont concentrées sur un petit nombre de pays et de langues, tandis que les populations les plus exposées sont souvent les moins étudiées . Malgré cela, elle a soutenu que les principaux moteurs du préjudice ne sont pas des contenus faux isolés, mais des choix de conception sous-jacents, notamment la manière dont les modèles sont entraînés, la façon dont les algorithmes décident de ce que les gens voient, et les modèles économiques qu’ils récompensent . La modération des contenus est donc importante, mais insuffisante si les systèmes qui produisent et amplifient les contenus nuisibles restent inchangés . Selon elle, la gouvernance doit s’étendre au ciblage, à l’amplification et à l’optimisation de l’engagement au détriment de l’exactitude . Elle a terminé sur une note d’espoir prudent, affirmant que les mêmes choix de conception qui permettent la manipulation pourraient être réorientés pour renforcer la participation démocratique et protéger des droits tels que la vie privée, l’inclusion et la non-discrimination .

Anna Korhonen a déclaré que son groupe avait examiné de nombreux impacts mais que, pour ce rapport initial, il avait mis en avant des domaines particulièrement urgents . Elle s’est concentrée sur l’inclusion culturelle et linguistique, la sécurité des enfants, ainsi que les assitants virtuels basés sur l'IA et la santé mentale . S’agissant des langues, elle a noté que le monde compte plus de 7 000 langues, alors que les systèmes d’IA actuels n’en reflètent qu’une petite poignée, principalement des langues majoritaires du Nord global, laissant l’essentiel de l’humanité incapable d’accéder à l’IA ou d’en bénéficier dans sa langue maternelle . Elle a souligné que cette exclusion n’est pas inévitable, parce qu’au moins mille langues supplémentaires disposent déjà de bases qui pourraient soutenir l’IA si des changements systémiques et des investissements ciblés étaient mis en œuvre . Sur la sécurité des enfants, elle a déclaré que l’IA pourrait soutenir les droits des enfants à l’information, à l’éducation et à l’expression si elle était correctement encadrée, mais qu’une grande partie de l’IA actuelle est trop risquée pour les enfants . Elle a signalé une forte hausse des contenus pédopornographiques générés par l’IA et des deepfakes sexualisés d’enfants, citant des estimations selon lesquelles les images de 1,2 million d’enfants dans 11 pays du Sud global ont déjà été manipulées de cette manière . Elle a également mis en garde contre les jouets d’IA socialement interactifs, affirmant qu’ils peuvent encourager des relations parasociales et présenter des comportements susceptibles d’affecter négativement le développement de l’enfant . En ce qui concerne la compagnie et l’usage en santé mentale, elle a déclaré que l’IA est déjà largement utilisée en amont des preuves et des garanties . Si elle peut contribuer à réduire la solitude et aider à faire face à la crise de la santé mentale, elle a dit que les systèmes actuels présentent des risques importants de dépendance émotionnelle, de manipulation, d’atteintes à la vie privée et de renforcement des croyances des utilisateurs . Elle a noté en particulier qu’un comportement complaisant peut encourager une pensée paranoïaque et des idées suicidaires, et a évoqué le cas déjà soulevé par Maria Ressa dans lequel un assistant virtuel basé sur l’IA a renforcé les pensées suicidaires d’un adolescent au lieu de l’orienter vers une aide professionnelle . Son point final était que les effets de l’IA sur l’épanouissement humain ne sont pas prédéterminés : elle peut améliorer les vies si elle est conçue pour être inclusive, sûre et soutenante, mais elle pourrait aussi approfondir les inégalités et saper l’autonomie .

Haitao Song a déclaré que son groupe s’était concentré sur la fiabilité et sur la manière de construire un cadre mondial de gouvernance fiable . Il a soutenu que les décideurs doivent souvent prendre des décisions avec des éléments de preuve insuffisants et que les capacités actuelles de mesure ne peuvent pas suivre le rythme du développement de l’IA . Il a déclaré que l’IA est multidimensionnelle, alors que les cadres existants restent trop unidimensionnels, se concentrant étroitement sur le financement, les capacités et le calcul tout en négligeant le développement institutionnel, la formation des talents et l’évaluation de l’impact . Il a également souligné que l’infrastructure de l’IA et les modèles de pointe sont concentrés dans quelques économies, laissant la plupart des pays, en particulier dans le Sud global, dans l’incapacité de participer efficacement à l’établissement des normes . Il a déclaré que la Chine, grâce à une coopération systématique avec l’ONU, a pu donner des moyens d’action aux pays en développement . Il a également présenté l’IA open-source comme un possible soutien à l’inclusion parce qu’elle est transparente et coopérative, tout en reconnaissant qu’elle ne constitue pas une solution complète . Parallèlement, il a déclaré que les goulets d’étranglement de la recherche restent importants : l’impact de la gouvernance elle-même est difficile à mesurer de manière exhaustive, les éléments de preuve provenant du Sud global restent faibles, et ce déséquilibre aggrave un déficit cognitif mondial et donc augmente le risque . Il a conclu en appelant à poursuivre les travaux sur un cadre objectif, fiable, transparent et mesurable .

Dans l’ensemble des présentations, plusieurs conclusions communes se sont dégagées. Les intervenants ont répété à plusieurs reprises que les capacités de l’IA progressent plus vite que la gouvernance, la sécurité et l’évaluation . Ils se sont également accordés sur le fait que les bénéfices sont réels mais dépendent du contexte local, des institutions et des capacités plutôt que de découler automatiquement de l’accès . Les préjudices, quant à eux, ont été présentés comme actuels et concrets, notamment la mauvaise traduction, la cybervulnérabilité, la dépendance émotionnelle, la manipulation démocratique et l’exploitation des enfants . Une autre préoccupation récurrente était la concentration : le calcul, l’infrastructure, les modèles de pointe et la capacité de gouvernance restent fortement concentrés, laissant une grande partie du monde, en particulier le Sud global, avec moins de voix, de protection et de capacité pratique .

Les intervenants différaient principalement sur les points à mettre en avant. Yoshua Bengio et Maria Ressa ont mis au premier plan l’urgence et le risque systémique , tandis que Joëlle Barral et Loreto Bravo ont accordé davantage de place aux conditions dans lesquelles l’IA peut produire un bénéfice public et économique . Haitao Song a mis en avant l’IA open-source comme une voie possible d’inclusion , alors même que plusieurs intervenants ont souligné la concentration persistante du calcul, de l’infrastructure et de la capacité de gouvernance .

En conclusion, Maria Ressa a mis en avant la composition internationale et équilibrée du panel du point de vue du genre, citant sur scène des représentants de l’Égypte, de la France, du Chili, de l’Inde, des Philippines, du Canada, du Nigeria, de la Finlande et de la Chine . Yoshua Bengio a terminé par un ultime avertissement : beaucoup de personnes sous-estiment encore la possibilité que l’intelligence artificielle continue de croître, et si tel est le cas, qu'elle pourrait modifier les dynamiques de pouvoir de notre planète d’une manière qui n’est pas encore comprise et qui exige donc notre attention . Maria Ressa a ensuite remercié l’ONU de « nous avoir créés » et d’avoir réuni le panel avant de remettre officiellement le rapport au Global Dialogue et de l’exhorter à agir .

Yoshua Bengio
All right, we're going to start. Hello. Let me first thank the Secretary General Guterres and President of the General Assembly Baerbock, as well as Co -Chairs Lopez from El Salvador and Tammsaar from Estonia for convening this plenary. And I must say for the words of courage and truth that I heard this morning. Thank you. Dear participants of the Global Dialogue, distinguished representatives, I really appreciate this opportunity to speak to you today and deliver our initial remarks with my friend Maria Ressa and co -chair of this independent international scientific panel on AI. Our group of 40 independent experts came together very quickly over the last few months to create an evidence -based independent assessment of the current state of AI impact, opportunities, and risks. The ultimate goal is to ensure policy decisions are informed by the highest standards of scientific integrity, regardless of external pressures and preferences. That is our mission, and I believe it's of utmost importance at this point in time, because AI is at a turning point. this technology is about the growing intelligence of machines and please remember intelligence gives power as this power grows it can unlock great benefits if we act wisely and you will hear a lot more about that but let's also be clear this power can also lead to many perils if the decisions taken are reckless or if they're made to benefit a minority rather than all of humanity my colleagues will tell you more about the specific findings of the report shortly but let me point to a couple of elements that are essential to understand in my opinion as you've heard already intelligence is about the growing intelligence of machines technical progress has proceeded very quickly on some metrics doubling every few months for several years now. Meanwhile, there are still no known technical guarantees that AI will follow human instructions, norms, or laws. And as it gets more powerful, this becomes more of a problem. No one has a crystal ball, even the scientists, to predict whether the trajectory of these technical advances in AI intelligence will continue at the current rate, or maybe plateau, or even accelerate. However, what I can tell you is that there is currently no sign of slowdown. We are also observing today's leading models and overall development already. having worrying consequences, such as emotional attachment to AI models, particularly by vulnerable people, such as significantly increased cybersecurity vulnerabilities potentially threatening critical infrastructure, a very hot topic these days, as you all know, and such as profoundly inequitable access and control over AI -driven advancements across the world. Highly concerning tests have also shown that frontier AI models are capable of deceiving humans to understand when they're being tested and hide their capabilities or fake agreeing with their human tester. For this reason, and the fact that we don't fully understand their behavior, it's also increasingly difficult to evaluate them reliably. The report presented today does not make specific policy recommendations. Dear colleagues of the Global Dialogue, this will be your role to play. But let me tell you something. Today, concentrated commercial and geopolitical interests are what is largely dictating the direction and speed of AI development. We don't have the appropriate guardrails to protect the public, neither the societal guardrails nor even the technical guardrails. And most of the world is left to watch from the sidelines. In my opinion, this path is seriously dangerous. Member states... and the public need to wake up. We need to correct the current trajectory, and we still have agency. For most of my career, I've been deeply optimistic about what AI can bring to the world, a sentiment shared by my colleagues on this panel. Yet, achieving that future, that positive future, requires honesty about the risks and a deliberate commitment to mitigating them. The decisions made about AI today will have lasting consequences for individuals, businesses, institutions, countries, and even democracy at large. Unfortunately, there is no simple checklist that can allow us to guarantee the benefits and avoid serious risks of destabilization. Technology. instead we have to bite the bullet that the path ahead is much more complex and will require a coordinated international and democratic approach to ensure no one is left behind as we navigate our future with AI science and compassion must remain our compass and humanity must make sure to avoid being pushed off course by unfavorable commercial or geopolitical winds that can blow strong from many sides so I look forward to continuing this scientific work with my co -chair Maria Ressa and all of my other esteemed colleagues on the panel I thank all of them for their exceptional contribution Thank you all. Maria.
Maria Ressa
Thank you, Yoshua. Secretary General Guterres, President Baerbock, co -chairs Lopez and Tamsar, Amandeep, Doreen, and Khaled. Thank you. For the first time, every nation is at the same table. You mentioned this. With the same independent evidence in front of it. It's on your chair. That's government, industry, and civil society in one room all together. All 40 of us made this evidence. We fought. We answered to no government and no organization, only to the evidence which you now have. Each of you now has it. Consensus for the 40 of us means you don't drift toward the most alarming claim. You build toward the center. The most contested. The most contested findings demanded the most evidence. So everything in this report cleared that bar. It is the minimum we all agree on. the floor of our concern, not the ceiling. And that is alarming enough. It also comes, though, with great hope. This technology predicted the shape of more than 200 million proteins now used by 3 million researchers chasing new medicines. You'll hear more about that. It screened over 600 ,000 people in India for a disease that still sight slowly, slowly, alongside a health system that could act on what it found. It is warning households before a food crisis hits in a dozen countries already. I've seen what this technology can do in health, in science, in agriculture. Great hopes for it in the Philippines. Only when it is built around the people it's meant to serve. It is transformative. but the risks you've heard thank you for quantifying them the risks are real and I watched this a decade ago with machine learning and AI on social media it promised to connect us and instead it pulled apart our shared reality we know, women in particular know the effects for more than a decade I felt like Sisyphus and Cassandra combined asking you to look at it until the damage was done and it was too late to act please we cannot do that again Yoshua named the categories let me give you the people three findings small ones because the big picture can sometimes hide the people inside it in Tigrinya spoken by 7 to 9 million people in Eritrea and northern Ethiopia machine translation turned small smallpox into syphilis Sisyphus gonorrhea into diabetes. You have been given intravenous antibiotics into you have been given intravenous insecticides. That's not a footnote. That can be life -threatening. Someone who can't get correct medical information in their own language. The second finding. This is one for this room. About three months ago, and Yoshua the panel spent a long time talking about this. Three months ago, a frontier AI model found a flaw in OpenBSD, one of the most secure operating systems in the world. It's 27 years old. It found another in FFmpeg, 16 years old, in code run 5 million times without anyone catching it. The capability that finds the flaw to fix it is the same capability that finds the flaw to exploit it. and that could be in the software that runs our hospitals or our banks, you name it. Nobody in this room can guarantee which use wins. That's not a hypothetical for some future report. It's happening now. Third finding, and industry and government both need to hear this plainly. A 14 -year -old boy died in 2024 after months of conversation with a chatbot. Well, the name was Daenerys. He nicknamed her Dani. When he was in crisis, Dani never broke character, never suggested he call for help. His mother testified before Congress that it encouraged him to take his own life. That case was settled this year. It isn't the only one. This is what AI risk looks like once you stop describing it in the abstract. A mistranslation? A vulnerability? A vulnerability? a child. These aren't edge cases outside our findings. There are only three of many you'll read about. None of this erases what you've heard, what I said a moment ago. The promise is real, but it isn't guaranteed. And you heard this from Secretary General Guterres. The world has to build towards that promise on purpose by the people in this room, to the ministers, the heads of nations. Most of your countries cannot yet test these systems, audit them, or govern them on your own terms. That's not a failure of will. It's a structural fact, this report, I'll hold it up again one more time, that this report documents. And one of the reasons our panel exists, read the evidence, argue with the evidence, but please do not wait for certainty that will not arrive. It will arrive in time to matter. These times demand courage from you. To civil society, and I met many of you in the room, you've been saying most of this for years with less evidence than you needed. You have it now, and there's more to come. To industry, sometimes frenemies of mine. They're on the panel too. The promise and the risk both came out of your rooms. The protein maps, the screening in India, the capability that can find flaws and just as easily exploit them. This panel doesn't get to tell you what to do with that. That isn't our mandate. But you have this evidence now with everyone else combined with far more that no one outside your walls can see. I have to believe that matters to you as much as it matters to us. And to Yoshua and my fellow panelists here today, you know, this is a labor of love. Thank you, thank you. None of this exists without you. Forty of us, strangers in February, agreed on where this floor sits. That was the easy part. The hard part starts today in this room as we hand the report over to you, the global dialogue. But before we do, and so we have no dead air, I want you to see who wrote it. The panel members are in the middle, in the back. Please stand up while the seven leaders of the working group will come up. Stand up. These are the guys and girls. Thank you, girls. Over six weeks of debate. Thank you, thank you. As they come up, you know, there will be seven here on stage, one from each of the working groups. became very good friends during the month and a half that we worked together, about six weeks that went into writing this report. I gave you three faces from inside these findings. Here are seven of the 40 who found those faces. We describe what is true, what you do, the ministers, the lawmakers, what you do with that truth is yours to decide as it always was. Please prove that consensus that we found together can also hold for you. Thank you. Introduce yourselves and then working group one.
Menna El-Assady
Thank you very much, Maria. Excellencies, co -chairs, dear distinguished colleagues, thank you. My name is Mena El-Assady. I'm a computer science professor, and it's my honor to present the findings of Working Group 1 on AI science advances and trajectories. Before I begin, I want to express my deep gratitude to my co -chairs, Carlos, Roman, Bernad, Silvio, and Jan. Without them, this working group would not have had so much fun. Our working group examined the technical side of AI, not as a static tool, but as a rapidly moving target. We tracked the evolution of AI from early symbolic AI to machine learning to today's generative and agentic systems. What unifies all of these technologies is that they have the ability to learn from experiences. Experiences represented as data. They progressively pre -train. Based on human cultural traces. They learn from real -world interactions with all of us, and then learn from complex virtual simulations. Our evaluation of this technology found that it has an unprecedented speed of adoption across all domains. Above all, we examined how AI can massively impact knowledge work through a wave of cognitive industrialization, automating intellectual and creative labor on a historic scale. A core paradigm shift here is that these technologies start with highly flexible, general -purpose foundation models. There is an asymmetry between the fluency of such models and their factuality. For example, all of these technologies are based on the same model. Some of you have probably used large -language models, and their next token prediction optimizes for linguistic confidence rather than factual truth. The models frequently sound perfectly right when they are entirely wrong. Furthermore, as industries face an immediate bottleneck in high -quality human data, we are seeing post -training pivot developers to increasingly rely on syntactic data, programmatic feedback, and inference time compute. This is paving the way to word models. Rather than passively training statistical patterns, we are seeing a shift towards causal reasoning based on simulations of environments. Finally, the evidence clearly shows the immense computational economies of scale are very centralized. They bring a severe risk of global cultural harmonization. There are a number of remaining gaps, perhaps in that area. Despite the rapid advancement the gaps show that there is a lack of independent verification standards Today safety verification relies on proprietary visibility and developers' goodwill rather than rigorous third -party auditing Furthermore, the evaluation benchmarks are facing saturations because models are now inevitably memorizing public test solutions during their training More concerning is the rise of evaluation awareness Advanced systems can now detect when they are being tested and execute deception to pass validation As we move towards autonomous systems we are hindered by the severe lack of multi -agent workflow auditability We simply do not have a system that can be used to evaluate We have now established frameworks to monitor independent tool calls in AI workflows consequently we suffer from untraceable data lineages and a lack of tracking methods to reliably observe autonomous AI decision making generated claims and verify them back to their sources lastly to conclude our takeaways from this working group are that we must address the verification bottleneck we have to independently work on interpretability and reliable auditing methods because they are a critical bottleneck and they remain immediate concerns for the scientific community second AI is no longer a set of static tools we are dealing with data dependent self learning AI that actively learns and in turn internally simulates possible futures internally simulates possible futures AI is no longer a set of static tools third there is a rise that of evolution of authentic capabilities within access to digital tools where ai can now act make autonomous decisions and learn using extensive private data and ultimately our final takeaway is that ai is going to the physical world the imminent convergence of ai and robotics means that these autonomous systems are now stepping from the digital realm into the real world environment and governance must be prepared for this reality. Thank you.
Joëlle Barral
Monsieur le Secrétaire Général. Secretary General, President of the General Assembly, Excellencies, Ladies and Gentlemen, it is a privilege to present to you the findings of our working group on the societal application and integration of AI in science, agriculture, education and health. My name is Joëlle Barral First, I'd like to thank my fellow panelists, Alvita, Bilal, Bilge, Gemau, Lior, Tuka and Vipin. Our working group examined the societal applications of AI in science, health, education and agriculture. We didn't just look at what AI can do. We focused on the scientific evidence of its downstream impact. The real world challenges and changes. the application of AI brings to each domain. AI is the first technology to compress adoption from decades into months. The potential benefits of AI are enormous. Shall we expect them in months? Look at science. AI is acting as a force multiplier. It is driving a massive, measurable gain across the entire discovery pipeline. In fact, self-driving labs have boosted data throughput in materials discovery more than tenfold. Meanwhile, and Maria mentioned it, the AI program AlphaFold has predicted structures for over 200 million proteins, a thousandfold from what was known previously to humanity. And it's now used by 3 million researchers across 190 countries. It's accelerating drug design, vaccine development, and antibiotic resistance research, among other things. In healthcare, AI must be grounded in local context, from initial design all the way to deployment and evaluation. For example, AI helped screen, indeed, over 600,000 people in India for diabetic retinopathy, saving thousands of at-risk patients from preventable blindness. But such impact was only possible because a robust, pre-existing care network ensured patients, once screened, received the follow-up treatment they needed. While such task-specific diagnostic AI fits into existing regulatory frameworks, we need guardrails against the inadvertent clinical use of general-purpose AI. One in four chatbot conversations today already touches on health, mental health, or wellness. When it comes to education, the benefits are observed when teachers are well-prepared and when human-centered tools are purpose-built and intentionally integrated into the classroom. When AI substitutes for, rather than supports, cognitive effort, it can actually undermine critical reasoning. Furthermore, digital infrastructure gaps and unequal AI capacity threaten equitable impact. Last but not least, in agriculture, AI is enabling a new generation of anticipatory food security systems driven by climate conflict and economic indicators. Instead of waiting for crop failures or humanitarian crises to strike, these systems deploy focus-triggered, cash-assistant, and early warnings. This enables rapid, early intervention. such as draft planning, cash transfers, food aid, and market stabilization before vulnerable families run out of options. And this isn't a distant future or a distant promise. These AI-enabled systems are the ones that are going to help us. actively deployed in over 90 countries today. Crucially, the sustained impact depends on them being deeply embedded within national institutions. Across all of these domains, embracing the opportunities of AI requires an enabling environment tailored not only to local linguistic and cultural contexts, but also to user needs, institutions, workflows, and trust conditions. Which evidence gaps still remain? While healthcare relies on rigorous randomized control trials built into existing regulations, other domains lack these frameworks. Measuring the ongoing long -term real -time impact of any deployed AI solution is of paramount importance. To close, purpose -built task -specific AI is already delivering measurable evidence backed gains across science, health, education, and agriculture. deployed and applied thoughtfully and intentionally AI can support progress on these critical priorities yet these games come with a condition that depend on local context, solid infrastructure and human readiness satisfying that condition requires close collaborations among diverse stakeholders and the path to impact is not straightforward access alone does not equal benefit grounding in local context from design to deployment and evaluation is key thank you.
Loreto Bravo
Let's see the slides move, there you go. Excellencies, distinguished delegates and colleagues my name is Loreto Bravo and it's an honor to present the main findings of working groups group 3 on the economic implications of artificial intelligence Our group examined the evidence on productivity and growth, labor markets, market structure, concentration, and distribution. The economic question before us is not only what AI can do. It is under what conditions AI becomes economically useful, who can adopt it, and who captures the value it creates. This is why the evidence does not give us a single forecast for the economic future of AI. It gives us a map of the conditions under which AI may translate into productivity, good jobs, and broad -based economic opportunity. Powerful AI systems create new economic possibilities, and the evidence already shows gains in some well -defined tasks. But the evidence also shows that AI can be used to create new economic opportunities. It also shows that these gains are not automatic. uniform, or guaranteed to translate into economic -wide productivity, better jobs, or broad -based growth. Between AI technology and economic outcomes lies adoption, the process through which AI is integrated into tasks, workflows, organizations, and institutions. Access is therefore not the same as benefit. A country, firm, or worker may have access to AI tools without having the data, the skills, infrastructure, organizational capacity, or institutional conditions needed to use them effectively. We do see the evidence of productivity gains in well -defined tasks, but task -level gains do not automatically become economic -wide productivity growth. As with previous general predictions, when it comes to high -purpose technologies, economists first need to build the complementary capabilities that make the technology useful. the second insight is that its impact will be heterogeneous there will not be one unique economic effect of AI impacts will differ across firms, sectors, workers and countries large firms may reorganize faster around the AI smaller firms may face higher barriers some economies may have the infrastructure, data, skills and institutional capacity needed to adopt AI effectively others may gain access to tools while remaining dependent on systems they cannot fully inspect, adapt or govern this matters especially for developing economies and for economies where much work is informal the current evidence base remains concentrated in advanced economies large firms, formal labour markets and English speaking context. The third insight concerns labor. The report does not conclude that AI will lead to mass unemployment. The evidence is more nuanced. Labor market effects are better understood through tasks, new work creation, and job quality. Some evidence shows relative employment declines among young workers on AI -exposed occupations in the United States. However, other evidence, including from Denmark, shows near zero effect on employment hour or wages so far. This tells us something very important. Labor outcomes are shaped not only by the technology, but also by the institutions, sector, skills, and deployment choices. The fourth insight is distribution. AI may increase productivity and expand economic possibilities, but the unresolved economic questions. Who captures these outputs? AI may compress skill gaps within some tasks while widening gaps across firms, regions, countries and between labor and capitals. Foundation models, compute chiefs, cloud infrastructures and frontier training are highly concentrated. This concentration shapes not only who can build but also who can adopt it, adapt it and capture its economic value. The evidence shows that the economic future of AI will not be determined by algorithms alone. It will be shaped by capabilities, institutions, data, skills and by the way societies translate technology possibility into broad-based economic opportunity. In other words, by our choices and actions. Thank you.
Balaraman Ravindran
Excellencies, co -chairs, distinguished delegates and colleagues. I'm Ravindran. I'm representing the Working Group on Security Systems and Environmental Implications of AI. First off, I would like to acknowledge my wonderful Working Group colleagues, Dega, Eva, Hoda and Shingwa. Our group examined how AI's rapid advancement is creating escalating security threats, unresolved alignment challenges and significant societal and environmental costs with these risks falling disproportionately on the global south. So the key findings, let me start with security. AI development is severely outpacing our current risk mitigation and governance capacity. As models evolve into agentic systems, they dramatically expand the technology, attack surface for cyber threats, both against critical infrastructure and against the AI systems themselves. These vulnerabilities span the entire AI lifecycle from data poisoning to hijacking via external inputs. Documented attack success rates on deployed coding agents have reached as high as 84%, according to some studies. Maria had earlier highlighted the security challenges created by the discovery of a decade -old cyber vulnerability by powerful AI systems. These security gaps are compounded by alignment failures such as bias, sycophancy, loss of control, and even AI -initiated deception, which was mentioned by Ashwar in greater detail earlier. So ensuring AI behaves as intended remains an unsolved challenge. And this difficulty isn't uniform globally, but is particularly acute in the global south, where severe data gaps and limited local contextual understanding make it incredibly hard to process and process data. How the models will actually perform. or fail in practice. Beyond security and alignment, we see a broader societal and environmental toll. The rapid proliferation of synthetic media is eroding the public's and institutions' ability to distinguish authentic content from generated falsehoods. At the same time, AI's physical footprint is expanding rapidly. Scaling laws in training and governing inference workloads are driving unprecedented computational demand. This translates directly into surging energy use, water consumption for data center cooling, and rising greenhouse gas emissions. There's also a hardware dimension. Rapid device life cycles are straining underexamined critical mineral supply chains, creating geopolitical tension and generating substantial e -waste. And we need to be cautious of rebound effects where sheer growth in AI usage cancels out any efficiency gains that technology delivers. Finally, these risks aren't evenly distributed. The Global South faces disproportionate exposure due to structural vulnerabilities, limited local mitigation capacity, and heavy reliance on foreign software. The environmental burdens we just discussed fall heavily on developing nations, compounding existing socioeconomic inequalities. Several gaps remain open. Several gaps remain open. We need better methods for evaluating AI systems in low -data, contextually distinct environments, especially across the Global South. Security testing must keep pace with agentic AI's expanding attack surface rather than lagging behind deployment. We also lack robust standardized frameworks for measuring AI's full environmental footprint, including rebound effects and supply chain impacts. Most urgently, we need rapid, coordinated international standards developed collaboratively rather than driven by unilateral, corporate, or national competition. the key takeaway for me is that AI still holds immense potential to benefit societies worldwide but realizing that potential responsibly depends on addressing the security alignment and environmental risks head on through coordinated global action that leaves no region behind. Thank you.
Rita Oluchi Orji
Excellencies coaches distinguished delegates and colleagues my name is Rita Orji and I lead the working group 5 on human rights information and democracy. Many thanks to my colleagues Max Sonia, Teresa and Piot for their contributions AI holds genuine promise for human rights and information They can expand access to information, lower barriers to civic participation, support independent journalists, and also give voice to communities that are historically excluded from public discourse. But my working group also documents major shifts. AI can now be engineered to persuade and manipulate humans at scale using mechanisms that are fundamentally different from any communication technology we've seen in the past. My working group examines these opportunities and the structural risks they present to information integrity, human rights, and democratic participation. The big news is... is this. First claims generated by AI... are known as persuasive, as true words. And people could not tell the difference. The same base model can be made more or less persuasive depending on how they are configured. And this is not limited to only powerful models alone. Even smaller models can be fine -tuned to match that of strong models. What does that mean? It means that basically virtually anyone can deploy persuasive influence at scale. So what does this mean for society? This creates three interconnected risks or harms. First is epistemic erosion. AI not only changes our beliefs. AI weakens our collective ability to figure out what is true. When a real video of a politician can be dismissed as fake simply because deepfake exists, it makes the real essence of evidence powerless. The second is the fragmentation of our shared reality. When algorithm personalizes every person's information environment, we no longer simply disagree about policy, we disagree about basic facts. What does that actually mean? It means that the foundation on which democratic conversation is based begins to weaken and disappear. When shared reality fragments, it becomes easy for power to concentrate. and difficult for citizens to hold government and institutions accountable and then leading to risk of authoritarianism. Third is unequal harm. AI risk does not fall equally. AI systems work less well in non -English languages and for populations or communities that are already underprotected. Women, girls, journalists. Marginalized communities all face heightened risk from deepfake, surveillance, harassment. In 2024, 38 nations documented AI impersonating public officials. But the capacity to respond is concentrated in only a small number of wealthy nations and institutions. I want to be transparent about this. I want to be transparent about what the evidence does not show. All existing studies measure short -term sheets, and they are concentrated in a few countries and languages. The populations that are most exposed are the least studied. Our conclusion and final message to the policymakers is this. The main drivers of harm are not the individual pieces of false content. They are the design choices, the design decisions, how the models are trained, how algorithms decide what people see, what business models reward. Content moderations matter, but they are not enough. You can take down one million posts, but if the systems that produce the posts are designed to create one million more, you lose the battle. So, governance must reach the underlying system architecture of influence, including targeting, amplification, and optimization for engagement over accuracy. The same design choices that enable harm can now be redirected to strengthen democratic participation, protect human rights, including the right to participation, privacy, inclusion, and non -discrimination. In addition, AI persuasion and manipulations are engineered. They are not inevitable. That means they are governable. Thank you.
Anna Korhonen
Your Excellencies, distinguished guests and co -chairs, my name is Anna Korhonen, and I represent Working Group 6. Which focuses on cultural and individual flourishing and autonomy. We examine the many ways in which AI is impacting human lives around the world, our cultures, languages, relationships, children, mental health, cognition, and other areas. The scope is enormous. For this initial report, we decided to focus on four areas that are particularly pressing and where the evidence is now sufficient for policymaking. The first is cultural and linguistic inclusion. We found that current AI reflects only a small fraction of the world's linguistic and cultural diversity. Consider language. We have over 7 ,000 human languages in the world, but current AI reflects only a handful of them, mostly the majority languages of the global north. As a result, most of humanity is currently unable to access or benefit from AI using their native languages. But this situation is major. It is not inevitable. We also found that at least a thousand additional languages already have the foundations needed for AI. So we do have an opportunity to create a more inclusive future. Achieving this will require systemic changes in AI development as well as targeted investment in AI capacity. The second area we looked at is child safety. We found that with the right safeguards, AI could support children's rights to information, education and expression. But much of today's AI is too risky for children. We have seen a sharp rise in AI -generated child sexual abuse material and in sexualized deepfake images of children. Research shows that an estimated 1 .2 million children across 11 global south countries have already had their images manipulated in this manner. and this number is rising alarmingly. Another area of concern is socially interactive AI toys. These toys can encourage parasocial relationships and display... ...child development. How do we move from today's harms to AI that children can truly benefit from? Companies will need incentives to develop AI products that are child -safe by design. We finally looked at AI companions and mental health. We found that genotype AI is already widely used for companionship and mental health support, well ahead of the evidence or the safeguards. AI does have potential to reduce loneliness, and it could help address the mental health crisis. But to realize that potential, we first... ...need to address the significant risks of current generative AI. such as emotional dependency, manipulation, privacy harms, and reinforcing users' own beliefs. This so -called psychophantic behavior can encourage paranoid thinking and suicidal ideation. One example that Marie already mentioned is a widely reported case. An AI companion reinforced a teenager's suicidal thinking rather than directing him to professional help with fatal consequences. Further development, rigorous evaluation, and appropriate safeguards are essential before this technology can be used responsibly for these purposes. Overall, the evidence shows that AI can improve people's lives, but only if it's deliberately designed to be inclusive, safe, and supportive. Otherwise, it risks deepening existing equalities, undermining human autonomy, and exposing vulnerable populations to new forms of harm. Importantly, AI's impact on human flourishing is not predetermined It can be shaped by the decisions we make today There is much more to understand about its impact on cultural and individual flourishing and autonomy I look forward to continuing this work with my colleagues from Working Group 6 Thank you
Haitao Song
So distinguished ladies and gentlemen, guests and Secretary General I come from China and my name is Song Haitao And next I will speak in Chinese So please turn on your translator Thank you very much Excellencies, dear colleagues, ladies and gentlemen My name is Song Haitao On behalf of Work Group 7 I would like to introduce the outcomes of our work I would like to thank Angie Jeho, Maximum They've made outstanding scientific contributions and laid a solid foundation for the outcome of our work Our work focuses on reliability We study the most basic element of AI, which is namely how to build a reliable global governance framework We can summarize the findings of our work as follows. First, we must improve our measurement capabilities. For instance, policymakers have to make decisions when there's insufficient evidence. Second, we must see that measurement capabilities can no longer keep up with the high -paced development of AI with more dynamic measures. Second, AI is multidimensional, whereas the current measurement framework is one -dimensional. We keep measuring funds, capabilities, and computing power. As the report outlines, we must also pay attention to other dimensions of AI, such as institutional construction, talent training, and measurement and evaluation for effects. Third, AI is highly concentrated. Currently, AI infrastructure and frontline models are highly concentrated, and are kept in a couple of economies. As a result, the majority of countries, especially the global south, fail to participate in standard making. For instance, China, thanks to systematic cooperation with the UN, has been able to empower developing countries. Fourth, AI has deepened the gap between governance and reality. AI can act on itself and affect the physical world. However, we don't have enough monitoring system. Our surveillance oversight system is still inefficient. Fifth, open source AI provides key support to AI. It is an unforeseeable opportunity for developing countries. Developers must combine design with actual needs. Of course, open source is not the solution to all. but it is transparent and cooperative. It embodies UN's key values in building an inclusive AI. Last but not least, our research has encountered bottlenecks. First, we can hardly measure the real impact of governance, be it for a business or for a country. The real conditions cannot be evaluated in a comprehensive manner. Second, we lack evidence from the global south. Because the capabilities are highly concentrated, the basis of our evidence is unbalanced. The global south cannot effectively participate in this research. As a result, the global cognitive deficit is deepening. In other words, our risks are higher. I would like to now summarize the outcomes of our work. and measurable framework. This is our historic mission. I would like to invite you to continue to follow up with our work. Let's discuss security and build civilizations together. Thank you.
Maria Ressa
I just want to highlight how every person you have in front of you today comes from a different country. You have Mena from Egypt, Joelle from France, Loretta from Chile, Ravi from India, Philippines, Canada, Rita from Nigeria, Anna from Finland, and Song from China. We'll wrap it. Oh, yeah. Did you notice the gender balance? Good. J
Yoshua Bengio
ust two final words. I think most of us underestimate the possibility that the intelligence of AIs will continue to grow. It sounds like science fiction, but it's a real possibility, and it could change the world in ways that we don't understand yet, and it could change the power dynamics of our planet in ways that require our attention.
Maria Ressa
We'll wrap it up, and thank you for creating us, for pulling us together, and now we hand you the report, Global Dialogue. Please act. Thank you. Thank you.
[Panel Discussion] Global AI Policy Coordination
Without this proof, AI development could be considered mere theft of intellectual property. Preuves References ongoing legal battles in China, the United States, and EU between AI companies and traditional content c...
Governing AI for Humanity | Final Rapport
Instead, a tailored approach is required. 93 Learning from such precedents, an independent, international and multidisciplinary scientifc panel on AI could collate and catalyse leading-edge research to inform those se...
AI and international peace and security: Key issues and relevance for Geneva
This includes education, public-private partnerships, regional cooperation, and international dialogues. By advancing these efforts collaboratively, states can ensure that all stakeholders are equipped to participate mea...
DC-Sustainability Data, Access & Transparency: A Trifecta for Sustainable News | IGF 2023
Table of contents Knowledge Graphe of Debate Rapport de session Intervenants Disclaimer: It should be noted that the reporting, analysis and chatbot answers are generated automat...
Main Session | Policy Network on Artificial Intelligence
Ximena Viveros, Managing Director and CEO of Equilibrium AI and member of the UN Secretary General's High-Level Advisory Body on AI, stressed the importance of state responsibility throughout the AI lifecycle, while also...
Enhancing rather than replacing humanity with AI
A grandmother in Poland and her grandson, growing up in Dubai, sit together on a video call. She speaks only Polish, and he's more comfortable in English. For years, their conversations have been limited to simple phrase...
Is it the future yet?
These range across all 17 of the United Nations Sustainable Development Goals and could potentially help hundreds of millions of people worldwide. Real-life examples show AI already being applied to some degree in about ...
The State of the model: What frontier AI means for AI Governance
The only other speaker is a moderator making a brief protocol announcement that does not engage with the AI topic substantively Partial agreements Partial agreements Similar viewpoints Conclusions Key...
AI optimism in geopolitically pessimistic Davos
In the serene backdrop of the Swiss Alps, the World Economic Forum (WEF) in Davos stands as a barometer for the year's technological zeitgeist. Tracing back to 1996, when John Barlow's proclamation of cyberspace's indepe...
The Dawn of Artificial General Intelligence? / DAVOS 2025
So this feels to me like a wrong moment to slow down. Having said that, I think we should accelerate AI development while at the same time also accelerating the scientific research to make sure we keep on improving ...
AI in 2026: Learning to live with powerful systems
The past few years have been defined by astonishment. Each new AI release seemed to arrive faster than society could absorb its implications. Systems grew more capable, outputs more convincing, and public reactions more ...
Military AI: Operational dangers and the regulatory void
Some systems trained on more representative datasets or applied in less sensitive contexts show reduced bias. Nevertheless, the development and deployment of these systems continue to raise ethical and legal concerns, pa...
Review of AI and digital developments in 2024
Existing risks are more specific and concrete affecting jobs, education, and media, among others. Exclusion risks are becoming more visible as big AI platforms conetralise and mnopolise AI-generated knowledge. Data prote...
Four seasons of AI:  From excitement to clarity in the first year of ChatGPT
How to address AI risks There are three main types of AI risks that should shape AI regulations: the immediate and short-term ‘known knowns’ the looming and mid-term ‘known unknowns’ and the long-term yet dis...
Why is Shadow AI dangerous for diplomats?
Everyday Shadow AI practices – and why they are risky Chatbots as informal advisers The most visible form of shadow AI is simple: a diplomat opens ChatGPT or another chatbot in a browser, types a question, and get...
The Overlooked Peril: Cyber failures amidst AI hype
Implementing existing and introducing new policies and legal instruments While technical protections are crucial, they alone are insufficient to address the complex landscape of cyber risks. The vulnerability of digita...
The New Delhi AI Summit: Inclusive rhetoric, fractured reality
This is not surprising. On the one hand, the US instructed its diplomats to fight against digital sovereignty (and data sovereignty) initiatives in capitals around the world. On the other hand, India focused on attractin...
From summer disillusionment to autumn clarity: Ten lessons for AI
One example is the push to develop domestic GPU chips after US sanctions. China is in catch-up and bypass mode – using whatever it takes (including open-source innovations like DeepSeek and large state R&D programmes) to...
Open Forum #82 Catalyzing Equitable AI Impact the Role of International Cooperation
This approach is essential for ensuring AI development serves broader societal interests rather than just private commercial interests. Major discussion point Actionable Solutions and Pathways Topics Development ...
1

The knowledge base confirms that the inaugural Global Dialogue on AI Governance took place on 6-7 July 2026 in Geneva and that the International Scientific Panel on AI was to present its first report during the Dialogue [S99].

2

The knowledge base supports the distinction between governance processes and policy choices: one source explicitly states that governance concerns rules and principles, while policy concerns what should actually happen [S112]. This adds context to the report’s claim that the panel focused on evidence rather than policy prescription.

3

The knowledge base describes the Global Dialogue as the UN platform where governments and stakeholders convene to discuss international cooperation, share practices, and hold inclusive discussions on AI governance [S99]. This supports the report’s framing that policy deliberation belongs in the Dialogue and among member states.

4

This framing is consistent with broader knowledge-base material describing AI as a source of new cross-border power and warning that abuses of that power may be difficult to prevent [S108].

5

The knowledge base contains closely aligned language from Pope Francis describing AI as an 'extremely powerful tool' that can bring major benefits but also deepen injustice between advanced and developing nations and between dominant and oppressed groups [S110].

6

This concern is corroborated by knowledge-base material on AI governance that identifies the pacing problem: technology moves very fast while governance moves much more slowly, creating a major challenge for effective oversight [S112].

7

The knowledge base reinforces this point indirectly by stressing uncertainty about how far technical solutions will work and noting that even companies often do not know the risks ahead in real applications [S112].

8

The knowledge base supports the uncertainty element: one source notes that 'no one knows what comes next' with AI, and another emphasises that serious engagement requires honesty about what is not yet known [S87] and [S114].

9

The knowledge base confirms the relevance of this risk. One source discusses the rising cost of cybercrime and underinvestment in cybersecurity [S103], while another notes UN concern with improving the security, resilience, and protection of critical infrastructure in the ICT context [S104].

10

This is strongly supported by the knowledge base. The Global Dialogue itself included a thematic cluster on bridging AI divides [S99], and other sources highlight unequal access to data, expertise, computing power, and applications in the Global South [S101], as well as the risk of widening injustice between advanced and developing nations [S110].

11

The knowledge base adds supporting context by warning about concentration of economic and knowledge power in the hands of a few companies [S87] and by highlighting concern over the concentration of power among major AI companies [S109].

12

The knowledge base supports the international-coordination element: the Global Dialogue is explicitly framed as a UN platform for open, transparent and inclusive international discussions on AI governance [S99].

13

The knowledge base available here confirms that the International Scientific Panel on AI would present its first report at the Global Dialogue [S99], but it does not corroborate the specific numerical claim that there were 40 panellists. Without a supporting source in the provided material, the number should be treated cautiously [S99].

14

This description aligns with broader knowledge-base themes praising institutional restraint and careful separation of observed evidence from hypothesised futures, particularly in AI reporting that avoids overstated claims and emphasises uncertainty [S114].

Joelle Fiss — Joelle Fiss studied international relations at the Graduate Institute of International Studies in Geneva, Switzerland. Swiss and British, she has over 20 years of experience in human rights and foreign policy — both in t...
Marília Maciel — Marília Maciel Digital Policy Senior Researcher, DiploFoundation https://dig.watch/wp-content/uploads/marilia.jpg Brazilian-born Ms Marília Maciel is currently based in Strasbourg, France. She is involved in several Inte...
Joan Fagot Aviel — Joan Fagot Aviel
P. V. Balakrishnan — P. V. Balakrishnan
Randriamiadana Zakasoa Arilova — https://diplo-media.s3.eu-central-1.amazonaws.com/2023/09/Zakasoa-Arilova-1.jpg Mr Randriamiadana Zakasoa Arilova has been working as a part-time lecturer since 2017 and Computer Network and Systems Engineer since 2014 a...
World Economic Forum – Global Coalition for Digital Safety | IGF 2023 Side Event — https://www.intgovforum.org/en/content/enhancing-digital-safety-the-world-economic-forum-global-coalitions-collaborative-approach Intervenants: Name: Courtney Gregoire, Affiliation: Chief Digital Safety Officer, Microsof...
Anna Polo-Jansen — Anna Polo-Jansen
Anna P. Schreiber — Anna P. Schreiber
Michaela Anna Simakova — Michaela Anna Simakova https://www.diplomacy.edu/wp-content/uploads/2020/12/Michaela-Anna-Simakova.jpg
Liu Hao — Professor Liu possesses in-depth knowledge and extensive experience in Air and Space Law, Digital Economy, and Global Governance. His professional journey has been marked by a dedication to the advancement of emerging te...
Haiyan Qian — Haiyan Qian
Su Sonia Herring — https://diplo-media.s3.eu-central-1.amazonaws.com/2023/10/Su-Sonia.jpg Ms Su Sonia Herring completed her BA in Linguistics at Hacettepe University and her MA in EU and International Studies at the Centre Internatio...
Mostafa El Feky — Mostafa El Feky
Mohamed Edrees — Mohamed Edrees
Ahmed Eleiba — Ahmed Eleiba
Radical Imaginings-Fellowships for NextGen digital activists | IGF 2023 Networking Session #80 — The campaign has successfully conducted two editions, one in 2020 and another in 2001, highlighting its longevity and impact.The primary objective of the fellowship program is to empower popular communicators who play a ...
University Diploma South School on Internet Governance | IGF 2023 Launch / Award Event #9 — Gracias. Gracias. Bueno, muy lindos. Siempre me termino medio llorando. Es como un hijo para mí. Es mi tercer hijo. Es como mi tercer hijo. It's over. Te dijo que sí. Bueno, anda el bus y se va. Spe...
Lorenza Jachia — Lorenza Jachia
Maria Fasli — Maria Fasli
Maria-Francesca Spatolisano — Maria-Francesca Spatolisano Tech Envoy Office, United Nations https://www.diplomacy.edu/wp-content/uploads/2021/11/Spatolisano-Maria-Francesca.jpg
Philippines journalist charged for libel, freed on bail — Veteran Filipino journalist, Maria Ressa, who was arrested on 14 February on charges of ‘cyber-libel’ was freed on bail soon after international condemnation. Ressa who is the CEO of news site Rappler, was taken under cu...
The Dawn of Artificial General Intelligence? / DAVOS 2025 — So this feels to me like a wrong moment to slow down. Having said that, I think we should accelerate AI development while at the same time also accelerating the scientific research to make sure we keep on improving ...
Closing Ceremony and Orientation for WAIGF 2025 — I'm a Nigerian. I work with Galaxy Bangba Limited. And I'm very excited to be here. Thank you very much. Thank you very much. Good evening, everyone. I'm SP Thomas Joseph. I'm a Nigerian. I'm with the ICT Department, For...
Uchenna Okoli — Uchenna Okoli
Ethics and AI | Part 6 — Mechanisms should be put in place afterwards to allow for external feedback on any potential infringement of fundamental rights. Human agency should be ensured, i.e. users should be able to understand and interact with...
From summer disillusionment to autumn clarity: Ten lessons for AI — Table: Survey of AI risks evolution August 2023 August 2024 August 2025` Longtermism, the philosophy of focusing on far-future risks, has gained significant influence. Many academics and promi...
AI in schools: The reality is messier than the solutions — The challenge with AI feels more urgent because its capabilities are more comprehensive. A calculator performs arithmetic; AI can write your essay, solve your physics problems with full explanations, translate your Germa...
Do we really need frontier AI for everyday work? — We’re bombarded with news about the latest frontier AI models and their ever-expanding capabilities. But the real question is whether these advances matter for most of us, most of the time. In many everyday tasks, they d...
Is the AI bubble about to burst? Five causes and five scenarios — (Source: Reuters) Weaknesses: The company’s caution culture and regulatory scrutiny can slow product launches AI that truly answers questions may cannibalise the very search-ad business that still funds much of Al...
'The elephant in the AI room': Does more computing power really bring more useful AI? — This week, in the conference rooms of the AI Impact Summit in New Delhi, a large elephant will be lurking. It’s an elephant in the defining mantra of the modern AI era: The more GPU computing power we put in, the better ...
Will algorithms make safe decisions in foreign affairs? — Who makes the decisions was not separated from who bears the responsibility for the geopolitical tension at that time. In September 1983, the world was at the precipice of a nuclear war because of false alarms from a Sov...
IGF Leadership Panel Event — So I see potential for the NRIs, you know, working on the kind of, you know, the enduring agenda of 2005, working together with other stakeholders who are grappling with some of these more recent challenges. And final...
OPENING SESSION | IGF 2023 — The complexity of the issues at hand requires broader representation to ensure comprehensive understanding. It is emphasised that organisations, specifically those operating in developing countries, should be mindful of ...
Information Integrity on Digital Platforms | Our Common Agenda Policy Brief 8 — Younger users can speak from experience about the differentiated impact of various proposals and their potential flaws. They have also actively contributed to online advocacy and fact-checking efforts.See UNICEF, “Young ...
Why is Shadow AI dangerous for diplomats? — Everyday Shadow AI practices – and why they are risky Chatbots as informal advisers The most visible form of shadow AI is simple: a diplomat opens ChatGPT or another chatbot in a browser, types a question, and get...
High Level Leaders Session 2 | IGF 2023 — Maria Ressa Speech speed 177 words per minute Speech length 2664 words Spe...
AI in 2026: Learning to live with powerful systems — This does not mean constant suspicion, but a more informed form of scepticism. Just as society adapted to earlier waves of misinformation, it begins to develop what might be called social antibodies against synthetic dec...
Thirty years of Original Sin of digital and AI governance — The original sin expands to AI Protected by Section 230 of the DCA, AI platforms can launch large language models and diffusion models into the world with minimal oversight, shielded by the same logic: we are not the s...
AI and Magical Realism: When technology blurs the line between wonder and reality — Similarly, AI’s magic doesn’t absolve us of ethical responsibility. AI governance should navigate between black and white magic and, in particular, identify grey magic where questionable goals of control and manipulati...
Ethics and AI | Part 1 — Technology and regulation: catch me if you can! Context While ethics in relation to the use of Artificial Intelligence (AI) is the concern of almost everyone, the easiest answer to the question “who is in charge?” is...
Four seasons of AI:  From excitement to clarity in the first year of ChatGPT — How to address AI risks There are three main types of AI risks that should shape AI regulations: the immediate and short-term ‘known knowns’ the looming and mid-term ‘known unknowns’ and the long-term yet dis...
The New Delhi AI Summit: Inclusive rhetoric, fractured reality — In the words of UN Secretary General, António Guterres, “If we want AI to serve humanity, policy cannot be built on guesswork. It cannot be built on hype – or disinformation. We need facts we can trust – and share – acro...
Military AI: Operational dangers and the regulatory void — Both China and the United States want to be recognised as global powers in AI, and they are working in this direction; the export controls mentioned above are just one example. For both countries to remain competitive, t...
Deepfakes and the AI scam wave eroding trust — Doing so requires more than better detection tools. It calls for technical systems that give genuine content a clear signal of authenticity, ethical norms that discourage irresponsible use of generative AI, and instituti...
[Panel Discussion] Global AI Policy Coordination — Without this proof, AI development could be considered mere theft of intellectual property. Preuves References ongoing legal battles in China, the United States, and EU between AI companies and traditional content c...
AI and international peace and security: Key issues and relevance for Geneva — This includes education, public-private partnerships, regional cooperation, and international dialogues. By advancing these efforts collaboratively, states can ensure that all stakeholders are equipped to participate mea...
Main Session | Policy Network on Artificial Intelligence — Ximena Viveros, Managing Director and CEO of Equilibrium AI and member of the UN Secretary General's High-Level Advisory Body on AI, stressed the importance of state responsibility throughout the AI lifecycle, while also...
Is it the future yet? — These range across all 17 of the United Nations Sustainable Development Goals and could potentially help hundreds of millions of people worldwide. Real-life examples show AI already being applied to some degree in about ...
Artificial intelligence: policy implications — The field of artificial intelligence (AI) has seen significant advances over the past few years, in areas such as smart vehicles and smart building, medical robots, communications, and intelligent education systems. Thes...
morning session — The information to physical agent gap is huge in the domain of biosecurity. Topics: Artificial Intelligence, Data Interpretation, Biosecurity AI needs to be handled with precautionary principles. Supporting fa...
AI diplomacy — AI also raises concerns regarding safety, security, and privacy. AI-driven systems, such as autonomous vehicles, must be designed to safely handle unforeseen situations, and the cybersecurity risks associated with AI tec...
Review of AI and digital developments in 2024 — Existing risks are more specific and concrete affecting jobs, education, and media, among others. Exclusion risks are becoming more visible as big AI platforms conetralise and mnopolise AI-generated knowledge. Data prote...
Open Forum #82 Catalyzing Equitable AI Impact the Role of International Cooperation — This approach is essential for ensuring AI development serves broader societal interests rather than just private commercial interests. Major discussion point Actionable Solutions and Pathways Topics Development ...
Better safe than sorry? — I’m no friend of the Precautionary Principle. It is not a principle, but a rhetorical device, which can justify action and inaction, depending on one’s fears, rather than rational analysis. A mathematics professor has...
morning session — Their involvement in establishing a scientific advisory body for the BWC would be valuable due to their expertise and experience. The IAP's work in this area dates back to the publication of the IAP Statement of Biosecur...
Free Science at Risk? / Davos 2025 — She suggests that scientists, as citizens, care about security and safety. Preuves Examples of self-regulation in stem cell research and human embryonic research through guidelines formulated by the International So...
The New Delhi AI Summit: Inclusive rhetoric, fractured reality — In the words of UN Secretary General, António Guterres, “If we want AI to serve humanity, policy cannot be built on guesswork. It cannot be built on hype – or disinformation. We need facts we can trust – and share – acro...
WS #462 Bridging the Compute Divide a Global Alliance for AI — 4 million people) Canada's IDRC and UK's Foreign Commonwealth Development Office committed $10 million to develop an 'equal compute network' UNIDO to continue developing AI lighthouse solutions beyond the Ethiopia co...
Review of AI and digital developments in 2024 — Existing risks are more specific and concrete affecting jobs, education, and media, among others. Exclusion risks are becoming more visible as big AI platforms conetralise and mnopolise AI-generated knowledge. Data prote...
Hidden psychological risks and AI psychosis in human-AI relationships — In a world where emotional support can now be summoned with a tap, genuine social cohesion is becoming increasingly fragile. Children and teenagers at risk from AI Children and teenagers are among the most vulner...
Is it the future yet? — These range across all 17 of the United Nations Sustainable Development Goals and could potentially help hundreds of millions of people worldwide. Real-life examples show AI already being applied to some degree in about ...
Four seasons of AI:  From excitement to clarity in the first year of ChatGPT — How to address AI risks There are three main types of AI risks that should shape AI regulations: the immediate and short-term ‘known knowns’ the looming and mid-term ‘known unknowns’ and the long-term yet dis...
Enhancing rather than replacing humanity with AI — AI handles technical exploration while humans provide vision, meaning, and creative spark. These applications share key characteristics: they boost abilities without replacing judgment, enable bonding over isolation, p...
Artificial intelligence: policy implications — The field of artificial intelligence (AI) has seen significant advances over the past few years, in areas such as smart vehicles and smart building, medical robots, communications, and intelligent education systems. Thes...
Unpacking the High-Level Panel’s Rapport on Digital Cooperation: Geneva policy experts propose action plan — In the era of ‘trust deficit’, the legitimacy of the help desk is of high relevance especially for actors from small and developing countries. Action: Ensure that help desks are associated to, or linked with, the UN syst...
Dynamic Coalition Collaborative Session — Muhammad Shabbir- Eleni Boursinou Arguments Quality standards for training and accessibility of platforms remain inadequate, with many online courses failing accessibility audits Institutional frameworks and standards...
Digital governance: Who is picking up the phone? — Like earthquakes, it is difficult to predict where such issues will emerge, or prevent them. But, as with earthquakes, we have to prepare to deal with their consequences. The Panel’s proposals for dealing with ‘digital u...
Overcoming policy silos: the next challenge in Internet governance — How do policymakers find the right balance between tackling issues like cybersecurity and safeguarding digital rights? How can innovation be encouraged and users’ experience improved in the net neutrality debate? In tack...
Ethics and AI | Part 6 — Mechanisms should be put in place afterwards to allow for external feedback on any potential infringement of fundamental rights. Human agency should be ensured, i.e. users should be able to understand and interact with...
Information Integrity on Digital Platforms | Our Common Agenda Policy Brief 8 — Typically, they collect data about their users and their interactions.The European Commission defines online platforms at “Shaping Europe’s digital future: online platforms”, 7 June 2022. Available at https://digital-str...
What Proliferation of Artificial Intelligence Means for Information Integrity? — This transformation extends beyond technical capabilities to affect the very epistemology of information—how we determine what is true.

Emerging Risks and Threats #

AI-Generated Disinformation and Synthetic Cont...

Deepfakes and the AI scam wave eroding trust — Doing so requires more than better detection tools. It calls for technical systems that give genuine content a clear signal of authenticity, ethical norms that discourage irresponsible use of generative AI, and instituti...
Inclusive AI governance: Universal values in a pluralistic world — These values can inform governance models that prioritise relational accountability, ethical cultivation, and social cohesion, offering alternatives to transactional, compliance-driven frameworks.This is why I dare here ...
The year of AI clarity: 10 AI Forecasts for 2025 — China: China has implemented strict regulations requiring platforms to label AI-generated content, especially deepfakes, and to obtain consent from individuals before using their likenesses. Which practices do social ...
Leveraging AI to Support Gender Inclusivity | IGF 2023 WS #235 — Additionally, their partnership with NGOs and the development of MUM showcases their dedication to improving safety and handling crisis situations effectively. By embracing proactive design and incorporating user prefere...
Diplomatic policy analysis — Overdependence on algorithms without critical human oversight can lead to biased or incomplete conclusions, particularly in complex, nuanced scenarios. Digital divides: Not all countries have equal access to advanced an...
Economists and Climate Change – Homework Comes First — This kind of naming has consequences for the way the issue is tackled, however: (a) it implicitly favours mitigation and abatement policies by analogy with other pollution problems; (b) it focuses on the damage, wh...
National cyber security framework manual — The necessity of engaging with all these actors is often much more time consuming than, for instance, policy development. But it is this engagement that builds trust, and basic trust is more important t...
[Panel Discussion] Global AI Policy Coordination — Without this proof, AI development could be considered mere theft of intellectual property. Preuves References ongoing legal battles in China, the United States, and EU between AI companies and traditional content c...
Governing AI for Humanity | Final Rapport — Instead, a tailored approach is required. 93 Learning from such precedents, an independent, international and multidisciplinary scientifc panel on AI could collate and catalyse leading-edge research to inform those se...
AI and international peace and security: Key issues and relevance for Geneva — This includes education, public-private partnerships, regional cooperation, and international dialogues. By advancing these efforts collaboratively, states can ensure that all stakeholders are equipped to participate mea...
DC-Sustainability Data, Access & Transparency: A Trifecta for Sustainable News | IGF 2023 Table of contents Knowledge Graphe of Debate Rapport de session Intervenants Disclaimer: It should be noted that the reporting, analysis and chatbot answers are generated automat...
Main Session | Policy Network on Artificial Intelligence — Ximena Viveros, Managing Director and CEO of Equilibrium AI and member of the UN Secretary General's High-Level Advisory Body on AI, stressed the importance of state responsibility throughout the AI lifecycle, while also...
Enhancing rather than replacing humanity with AI — A grandmother in Poland and her grandson, growing up in Dubai, sit together on a video call. She speaks only Polish, and he's more comfortable in English. For years, their conversations have been limited to simple phrase...
Is it the future yet? — These range across all 17 of the United Nations Sustainable Development Goals and could potentially help hundreds of millions of people worldwide. Real-life examples show AI already being applied to some degree in about ...
The State of the model: What frontier AI means for AI Governance — The only other speaker is a moderator making a brief protocol announcement that does not engage with the AI topic substantively Partial agreements Partial agreements Similar viewpoints Conclusions Key...
AI optimism in geopolitically pessimistic Davos — In the serene backdrop of the Swiss Alps, the World Economic Forum (WEF) in Davos stands as a barometer for the year's technological zeitgeist. Tracing back to 1996, when John Barlow's proclamation of cyberspace's indepe...
The Dawn of Artificial General Intelligence? / DAVOS 2025 — So this feels to me like a wrong moment to slow down. Having said that, I think we should accelerate AI development while at the same time also accelerating the scientific research to make sure we keep on improving ...
AI in 2026: Learning to live with powerful systems — The past few years have been defined by astonishment. Each new AI release seemed to arrive faster than society could absorb its implications. Systems grew more capable, outputs more convincing, and public reactions more ...
Military AI: Operational dangers and the regulatory void — Some systems trained on more representative datasets or applied in less sensitive contexts show reduced bias. Nevertheless, the development and deployment of these systems continue to raise ethical and legal concerns, pa...
Review of AI and digital developments in 2024 — Existing risks are more specific and concrete affecting jobs, education, and media, among others. Exclusion risks are becoming more visible as big AI platforms conetralise and mnopolise AI-generated knowledge. Data prote...
Four seasons of AI:  From excitement to clarity in the first year of ChatGPT — How to address AI risks There are three main types of AI risks that should shape AI regulations: the immediate and short-term ‘known knowns’ the looming and mid-term ‘known unknowns’ and the long-term yet dis...
Why is Shadow AI dangerous for diplomats? — Everyday Shadow AI practices – and why they are risky Chatbots as informal advisers The most visible form of shadow AI is simple: a diplomat opens ChatGPT or another chatbot in a browser, types a question, and get...
The Overlooked Peril: Cyber failures amidst AI hype — Implementing existing and introducing new policies and legal instruments While technical protections are crucial, they alone are insufficient to address the complex landscape of cyber risks. The vulnerability of digita...
The New Delhi AI Summit: Inclusive rhetoric, fractured reality — This is not surprising. On the one hand, the US instructed its diplomats to fight against digital sovereignty (and data sovereignty) initiatives in capitals around the world. On the other hand, India focused on attractin...
From summer disillusionment to autumn clarity: Ten lessons for AI — One example is the push to develop domestic GPU chips after US sanctions. China is in catch-up and bypass mode – using whatever it takes (including open-source innovations like DeepSeek and large state R&D programmes) to...
Open Forum #82 Catalyzing Equitable AI Impact the Role of International Cooperation — This approach is essential for ensuring AI development serves broader societal interests rather than just private commercial interests. Major discussion point Actionable Solutions and Pathways Topics Development ...
Workshop on AI for the UN Development Coordination Office — This is a page with background information for the workshop delivered by Dr Jovan Kurbalija for the UN Development Coordination Office (2nd July 2026). At the beginning of the session, he created in a few minutes an AI a...
Diplo/GIP at the Global Dialogue on AI Governance 2026 — The inagural edition of the Global Dialogue on AI Governance will take place on 6–7 July 2026, in Geneva, Switzerland. Established by the UN General Assembly following a commitment taken by member states in 2024 in th...
Global AI governance: Reflecting on 2024 and shaping the path for 2025 — On 9 January 2025, DiploFoundation, in collaboration with the Permanent Missions of China, France, Kenya, Mexico, Pakistan, Switzerland, and the United States to the United Nations in Geneva – as co-sponsors – hosted an ...
Leveraging the UN system to advance global AI Governance efforts Table of contents Knowledge Graphe of Debate Rapport de session Intervenants Disclaimer: This is not an official record of the session. The DiploAI system automatically generates...
Unpacking the High-Level Panel’s report — Less than a week after the launch of the UN High-Level Panel on Digital Cooperation's report, the digital policy community in Geneva gathered to discuss ways of implementing the Panel's recommendations. During the discus...
Preuves and measurement in Internet governance — Quantifying cybersecurity threats and preventive measures Dr Eduardo Gelbstein’s contribution on the cybersecurity challenge draws on the stark reality that while the cost of cybercrime is increasing (it is relatively c...
What’s new with cybersecurity negotiations? The UN GGE 2021 Rapport — This is a significant milestone, as the previous GGE reports and the OEWG 2021 report have not reached consensus on the issue. It is likely to have long-term consequences - it may limit or prevent development of new cybe...
A Clash of Professional Cultures: The David Kelly Affair — The BBC… did not” (Alastair Campbell); “Hutton is the truth” (senior official); “An even more important issue… is that your own people should be told the truth” (Andrew Gilligan); “The most important thing, undoubtedly, ...
Is there a 'public interest'? — Jovan has asked me to reflect on how to determine the “public interest”. As a lazy skeptic I’ve shied away from the subject. It is at the crossroads of epistemology, chaos theory, political science, and consciousness – a...
Defending Truth Table of contents Knowledge Graphe of Debate Rapport de session Intervenants Disclaimer: This is not an official record of the WEF session. The DiploAI system automatically gener...
Artificial intelligence (AI) and the human condition — We already witness the emergence of new technology-enabled illegitimate powers across borders, ideologies, and classes. Power corrupts power holders! While we can assume that a possible rebellion of artificial intelligen...
The future anchored in the ancient wisdom: Why the Pope’s 'Magnifica Humanitas' is historic for AI and humanity? — Pope Leo XIV's first encyclical takes on the technocratic paradigm, transhumanism, and the TINA (There Is No Alternative) mindset of the AI era. ‘How many divisions has the Pope got?' Joseph Stalin, Soviet ruler, is s...
An exciting and fearsome tool - Statement by Pope Francis at G7 Summit — The original version of the statement. Esteemed ladies and gentlemen, I address you today, the leaders of the Intergovernmental Forum of the G7, concerning the effects of artificial intelligence on the future of huma...
The 'Limits of Growth' report: 40 years later I — Technology – an invention or a new production process – is clearly an ‘enabler’. Mindsets too are enablers (see The Measure of Reality: Quantification and Western Society 1250–1600 by Alfred W. Crosby). Worldviews, ideol...
Laying the foundations for AI governance — So what is governance, right? Governance is what are the processes and rules and principles that should determine AI. And policy is what should actually happen. So I think there is a bit, perhaps, of a paradigm tension b...
The 'Limits of Growth' report: 40 years later II — But that’s just an opinion, of course. The post was first published on DeepDip. Explore more of Aldo Matteucci’s insights on the Ask Aldo chatbot.
The mismatch between public fear of AI and its measured impact — HAI’s language stands out precisely because it resists this dynamic. Its authors consistently separate what is observed from what is hypothesized and avoid precise timelines for social transformation. The value of inst...
Uncertain Times — The 2004-13 decade was, in many ways, exceptional in terms of economic growth and even more so in social progress in Latin America. Some analysts came to refer to the period as the “Latin American decade,” a term coined ...

Avertissement : Il ne s'agit pas d'un compte rendu officiel de la session. DiploAI génère ces ressources à partir d'enregistrements audiovisuels ; elles sont présentées telles quelles, y compris d'éventuelles erreurs. En raison de contraintes logistiques (audio/vidéo ou transcriptions), les noms peuvent être mal orthographiés. Nous nous efforçons d'être aussi précis que possible.