WSIS Forum 2026
Rapport généré par l'IA

Fit for Whom?

10 intervenants
Résumé

Résumé

Cette discussion, animée par Caitlin Kraft-Buchman de Women at the Table, a porté sur l'exclusion systémique des femmes de la recherche médicale et des données, ainsi que sur l'effet amplificateur de ce phénomène lorsque l'intelligence artificielle est construite sur ces bases défaillantes. Les intervenantes Oriana Kraft, de FemTechnology, et Yu Ping Chan, du PNUD, ont analysé comment les biais de genre sont intégrés à chaque niveau de la chaîne des soins de santé et des technologies. Oriana Kraft a décrit l'ampleur du problème en médecine, soulignant que des différences biologiques entre les sexes existent dans chaque cellule du corps humain, alors que la recherche s'est historiquement concentrée presque exclusivement sur des sujets masculins . Les animaux mâles sont cinq fois et demie plus nombreux que les femelles dans les études, en raison d'une hypothèse erronée selon laquelle les cycles hormonaux féminins les rendraient trop imprévisibles . Les conséquences sont graves : les femmes reçoivent un diagnostic en moyenne quatre ans plus tard que les hommes pour 770 maladies , ont 50 % plus de risques de mourir à la suite d'une crise cardiaque , et sont 50 à 75 % plus susceptibles de présenter des effets indésirables médicamenteux . Des pathologies telles que l'endométriose nécessitent en moyenne sept ans avant d'être diagnostiquées, et les femmes doivent consulter en moyenne cinq médecins avant d'obtenir un diagnostic précis, ce qui entraîne l'accumulation de données inexactes dans les bases de données médicales . Lorsque les modèles d'IA sont entraînés sur ces données historiques biaisées, la précision diagnostique diminue de 11,3 % , et le problème est encore aggravé par l'utilisation de données synthétiques, qui efface progressivement les signaux relatifs à la santé des femmes aux marges des distributions de données . Yu Ping Chan a ajouté qu'une enquête du PNUD menée dans 80 pays a révélé que près de 90 % des hommes comme des femmes entretiennent au moins un préjugé à l'égard des femmes, ce qui signifie que le substrat même sur lequel reposent les systèmes numériques est déjà biaisé . Elle a averti qu'un écart numérique de genre de 10 à 15 % en matière de connectivité à l'échelle mondiale, atteignant 30 à 40 % dans les pays en développement, implique que la poursuite de la transformation numérique sans remédier à ces biais ne fera qu'aggraver les inégalités . Parmi les solutions proposées figurent l'établissement de la représentativité comme norme scientifique dans la collecte de données , l'intégration d'exigences relatives aux données ventilées par sexe dans les marchés publics , et l'incitation des médecins et des patientes à contribuer des données de santé plus riches et plus nuancées . Caitlin Kraft-Buchman a également suggéré de former des groupes de défense des droits des femmes pour qu'ils deviennent collecteurs et propriétaires de données, leur permettant ainsi de revendre ces données à des marchés qui en sont actuellement dépourvus . Une participante de Tanzanie a souligné que des recherches ancrées dans des populations féminines diversifiées pourraient ouvrir la voie à des découvertes importantes, comme comprendre pourquoi certaines communautés présentent moins de symptômes ménopausiques . La discussion s'est conclue sur un sentiment partagé d'urgence et de frustration face au fait que la prise de conscience n'a pas débouché sur des actions suffisantes . Les intervenantes ont appelé à une responsabilité politique, à des normes internationales plus claires et à l'utilisation de cadres existants tels que les indicateurs e-santé du SMSI pour promouvoir des pratiques de données représentatives du genre . Le message central était que la correction de ces biais profondément enracinés nécessite un engagement coordonné des gouvernements, du secteur privé et des organisations internationales, avant que de nouvelles infrastructures numériques ne soient construites sur des fondations déjà inégales .

Points clés

Principaux points de discussion

- Les biais de genre dans la recherche médicale et les modèles d'IA en santé : Les intervenantes ont mis en évidence que la recherche médicale a historiquement été menée presque exclusivement sur des sujets masculins, entraînant des lacunes systémiques dans les soins de santé des femmes. Ce biais est ensuite amplifié lorsque les modèles d'IA sont entraînés sur ces données défaillantes. Par exemple, les seuils de troponine calibrés sur des hommes passent à côté de 42 % des crises cardiaques féminines , les femmes reçoivent un diagnostic en moyenne quatre ans plus tard que les hommes pour 770 maladies , et l'IA entraînée sur des données non représentatives réduit la précision diagnostique de 11,3 % . La « cascade de distorsion » - de la recherche préclinique aux recommandations cliniques jusqu'aux modèles d'IA - a été décrite en détail , les animaux mâles étant cinq fois et demie plus nombreux que les femelles dans les études . - La nécessité d'une collecte de données plus riche, ventilée par sexe et par étape de vie : Un thème central était l'absence de données capturant les expériences biologiques proprement féminines. Il n'existe actuellement aucun formulaire standard dans les systèmes de santé ou de ressources humaines pour recueillir des informations sur la ménopause, le cycle menstruel ou les expériences post-partum . Les intervenantes ont soutenu que les étapes de vie féminines interagissent avec chaque organe du corps et que la collecte de ces informations - y compris auprès des patientes elles-mêmes via des objets connectés et des journaux de cycle - pourrait être transformatrice . Les notes des médecins femmes se sont révélées deux fois plus détaillées que celles de leurs homologues masculins, ce qui suggère que l'incitation à une collecte de données nuancée constitue une mesure pratique . - L'inscription des biais de genre dans l'infrastructure publique numérique : Yu Ping Chan a soulevé la préoccupation que la construction d'infrastructures publiques numériques risque d'encoder les biais sociétaux existants. L'Indice des normes sociales de genre du PNUD a révélé que près de 90 % des hommes et 87 % des femmes entretiennent au moins un préjugé à l'égard des femmes, et que les progrès dans ce domaine ont stagné depuis une décennie . Combiné à un écart numérique de genre de 10 à 15 % en matière de connectivité à l'échelle mondiale - atteignant 30 à 40 % dans les pays en développement - le panel a averti que la poursuite de la transformation numérique sans remédier à ces biais ne fera qu'aggraver les inégalités . - Traduire les engagements en actions concrètes - normes, marchés publics et responsabilité : La discussion a porté sur la manière de dépasser les déclarations et les documents pour parvenir à un changement réel. La Déclaration de Hambourg sur l'IA responsable pour les ODD a été citée comme exemple d'engagement multipartite intégrant le genre comme domaine prioritaire , mais les intervenantes ont reconnu la difficulté de transformer les engagements sur le papier en actions concrètes . Les mécanismes proposés comprenaient l'établissement de la représentativité comme norme scientifique , l'intégration d'exigences relatives aux données ventilées par sexe dans les marchés publics , et l'utilisation des indicateurs e-santé du SMSI pour encourager des données d'entraînement représentatives du genre . - Autonomiser les femmes en tant que productrices de données et actrices du changement : Plutôt que de présenter les femmes uniquement comme bénéficiaires de la technologie, le panel a plaidé pour les positionner comme collectrices, propriétaires et productrices de données . Les intervenantes ont suggéré de former des groupes de défense des droits des femmes à la collecte et à la gestion des données, et ont noté que les femmes dans les contextes de la femtech sont déjà prêtes à partager leurs données pour faire avancer la santé féminine . Le panel a également appelé à une responsabilité politique, affirmant que la santé des femmes devrait figurer dans les programmes électoraux et que les efforts de lobbying et le plaidoyer de personnalités publiques ont déjà commencé à faire évoluer la sensibilisation du public . ---

Objectif général

La discussion visait à mettre en lumière l'exclusion systémique des femmes de la recherche médicale et des données d'entraînement de l'IA, à illustrer les conséquences concrètes de cette exclusion sur la santé, et à explorer des voies pratiques, notamment des normes de données, des politiques d'achat public, le plaidoyer politique et la collecte de données à la base, pour corriger ces déséquilibres avant qu'ils ne s'ancrent davantage dans les infrastructures numériques et d'IA émergentes. ---

Ton général

Les intervenantes se sont montrées assez préoccupées, reflétant la gravité des disparités de santé décrites. La présentation détaillée par Oriana Kraft de la « cascade de distorsion » était fondée sur des données probantes. Yu Ping Chan a fait preuve de franchise quant à ses incertitudes concernant les solutions , conférant à la discussion une qualité honnête et qui la distinguait des panels de politique plus policés. Cependant, le ton a évolué vers un optimisme prudent vers la fin de la discussion, notamment lorsque les intervenantes ont évoqué le potentiel d'innovation par bonds dans le Sud global , l'opportunité scientifique que représentent les données inexploitées sur la santé des femmes , et l'idée d'une « course aux données de santé » entre nations . Les contributions du public - en particulier de femmes partageant des expériences personnelles et culturelles de Tanzanie et de Chine - ont ajouté une dimension humaine à la discussion et ont mis en avant les efforts collectifs.

Intervenantes et intervenants

- Caitlin Kraft-Buchman - Rôle : Modératrice/Animatrice du panel - Affiliation : Women at the Table (fondatrice/représentante) - Domaine d'expertise : Égalité de genre, IA et représentation des données, droits des femmes dans la technologie - Oriana Kraft - Rôle : Fondatrice - Affiliation : FemmeTechnology.org - Domaine d'expertise : Santé des femmes, données médicales ventilées par sexe, biais de l'IA dans les soins de santé, femtech, lacunes de la recherche clinique affectant les femmes - Yu Ping Chan - Rôle : Représentante/Intervenante - Affiliation : PNUD (Programme des Nations Unies pour le développement) - Domaine d'expertise : Infrastructure publique numérique, genre et équité numérique, IA responsable pour le développement durable, la Déclaration de Hambourg sur l'IA pour les ODD - Participante 1 - Rôle : Participante du public - Contexte : De Chine ; a soulevé la question des biais de genre de l'IA dans ses réponses et la pratique consistant à dissimuler son genre lors des interactions avec l'IA - Participante 2 - Rôle : Participante du public - Contexte : De Tanzanie, Afrique ; a partagé des observations sur les différences générationnelles dans les expériences de santé des femmes et l'importance de recherches fondées sur des populations féminines diversifiées - Participante 3 - Rôle : Participante du public - A soulevé des questions sur le rythme du plaidoyer en faveur de la santé des femmes et de l'inclusion technologique, ainsi que sur les mesures concrètes pouvant être prises - Participante 4 - Rôle : Participante du public - Contexte : De Chine ; a posé des questions sur la sensibilisation à l'endométriose et aux défaillances systémiques dans les soins de santé des femmes, notamment dans le contexte chinois - Yipeng - Rôle : Intervenante/Oratrice (semble être la même personne que Yu Ping Chan, désignée par une version abrégée ou alternative du nom) - Affiliation : PNUD - Domaine d'expertise : Équité numérique, genre et technologie, IA responsable Intervenantes et intervenants supplémentaires : - Intervenant·e 1 - Rôle : Panéliste ou participant·e non identifié·e ; d'après le contexte et le contenu, cet·te intervenant·e semble recouper significativement Oriana Kraft, abordant les données de santé des femmes, les lacunes de la recherche clinique et la femtech - Domaine d'expertise : Santé des femmes, biais dans la recherche médicale, IA et données de santé - Intervenant·e 2 - Rôle : Intervenant·e non identifié·e ; d'après le contexte, cet·te intervenant·e semble recouper Caitlin Kraft-Buchman, jouant un rôle de modération et de synthèse tout au long de la discussion - Domaine d'expertise : Équité de genre, normes de données IA, autonomisation numérique des femmes

Intervenants
OK
Oriana Kraft
184 wpm · 14 min
CK
Caitlin Kraft-Buchman
139 wpm · 8 min
YP
Yu Ping Chan
196 wpm · 7 min
S1
Speaker 1
191 wpm · 10 min
AM
Audience Member 1
157 wpm · 42 s
AM
Audience Member 2
117 wpm · 2 min
Y
Yipeng
195 wpm · 1 min
S2
Speaker 2
153 wpm · 6 min
AM
Audience Member 3
157 wpm · 54 s
AM
Audience Member 4
96 wpm · 1 min

Résumé élargi : Biais de genre dans la recherche médicale, dans les modèles d'IA en santé et dans les infrastructures numériques publiques

#

Introduction et contexte

La session était animée par Caitlin Kraft-Buchman de Women at the Table et a réuni Oriana Kraft, fondatrice de FemmeTechnology.org, et Yu Ping Chan du PNUD . La discussion s'est tenue dans un contexte de préoccupation croissante concernant le déploiement de l'intelligence artificielle dans des environnements à risque élevé, notamment les hôpitaux, les tribunaux et les systèmes financiers, où les décisions relatives à l'attribution d'un diagnostic, d'un prêt ou d'une voie de recours juridique sont de plus en plus façonnées par des systèmes algorithmiques . Caitlin Kraft-Buchman a ouvert la session en observant que l'apprentissage automatique des systèmes d'IA fonctionne sur la base de moyennes et d'un auto-apprentissage récursif, ce qui signifie que les populations les plus éloignées des lieux bénéficiant de cette technologie, notamment les femmes, les communautés rurales et les personnes handicapées,sont progressivement effacées des résultats des modèles à mesure que la technologie évolue . La problématique principale de la séance était donc la suivante : le biais de genre n'est pas simplement un problème social, mais une caractéristique structurelle de la façon dont les systèmes d'IA sont construits, avec des conséquences profondes dans le monde réel.

#

Le biais de genre dans la recherche médicale

Oriana Kraft décrit un problème qui va de la recherche préclinique aux recommandations cliniques, jusqu'à l'entraînement des modèles d'IA . Elle a proposé une première observation scientifique : il existe des différences biologiques dans chaque cellule du corps humain, ce qui signifie que le fonctionnement de chaque organe diffère entre les hommes et les femmes . Malgré cela, la recherche médicale a historiquement été menée presque exclusivement sur des sujets masculins, avec des cellules mâles, des animaux mâles et des corps masculins . Les animaux mâles sont surreprésentés dans les études de recherche par rapport aux femelles dans un rapport de 5,5 pour un, une disparité enracinée dans une hypothèse manifestement erronée selon laquelle les cycles hormonaux féminins rendraient les sujets femelles trop imprévisibles . En réalité, Oriana Kraft a noté que les les fluctuations de testostérone des souris mâles sont moins prévisibles que les hormones des souris femelles, dont les cycles suivent au moins un schéma connu . L'exclusion des sujets féminins n'était donc pas simplement un jugement de valeur, mais une erreur scientifique.

Les conséquences de cette exclusion sont graves et bien documentées. Les femmes reçoivent un diagnostic en moyenne quatre ans plus tard que les hommes pour 770 maladies , ont 50 % plus de risques de mourir à la suite d'une crise cardiaque , et sont 50 à 75 % plus susceptibles de présenter des effets indésirables médicamenteux, en grande partie parce qu'elles ont été effectivement exclues des essais cliniques jusqu'en 1993 . Les outils diagnostiques ont été calibrés sur la physiologie masculine : les seuils de troponine utilisés pour détecter les maladies cardiovasculaires sont fixés à des niveaux adaptés au corps masculin et sont trop élevés pour les femmes, ce qui entraîne la non-détection de 42 % des crises cardiaques féminines . Les technologies d'imagerie sont également conçues autour des vaisseaux le plus souvent touchés chez les hommes, ce qui signifie que les manifestations cardiovasculaires féminines sont systématiquement négligées à plusieurs étapes du processus diagnostique . Une fois diagnostiquées, les femmes ont moins de chances de se voir prescrire un traitement de référence, et même ces traitements se sont révélés moins efficaces chez les femmes, comme Oriana Kraft l'a noté à propos de l'essai REBOOT, publié environ deux ans auparavant .

Le problème des erreurs de diagnostic est aggravé par l'accumulation de dossiers inexacts dans les bases de données médicales. Des affections telles que l'endométriose prennent en moyenne sept ans à être diagnostiquées, et les femmes doivent consulter en moyenne cinq médecins avant de recevoir un diagnostic correct . Chaque médecin consulté avant le diagnostic correct enregistre un diagnostic inexact, et à moins qu'une patiente ne se trouve dans un pays comme le Danemark, qui a investi dans de grands registres de santé interconnectés, ces dossiers inexacts ne sont jamais mis à jour . Il en résulte un corpus de données médicales qui n'est pas seulement incomplet, mais activement trompeur, et c'est sur cette base que les modèles d'IA en santé sont entraînés .

#

L'effet cumulatif de l'entraînement de l'IA sur des données biaisées

Oriana Kraft a alors donné son avis sur les affirmations qui prétendent que l'IA explicable, à savoir des systèmes conçus pour articuler leur raisonnement, peut compenser des données d'entraînement biaisées. Les données probantes n'approuvent pas cette hypothèse. Oriana Kraft a cité des chiffres indiquant que l'IA entraînée sur des données médicales non représentatives réduit la précision diagnostique de 11,3 %, et que cette réduction persiste même lorsque les médecins utilisent des systèmes d'IA explicable . La seule véritable solution réside dans des données précises et représentatives au moment de l'entraînement du système. Le problème est encore aggravé par le recours croissant aux données synthétiques, vers lesquelles les grandes entreprises d'IA se sont tournées en déclarant manquer de nouvelles données humaines . Chaque fois qu'un modèle est réentraîné sur des données synthétiques, les signaux provenant des extrémités des distributions - là où les femmes, en tant que groupe sous-représenté, se trouvent de manière disproportionnée - sont progressivement effacés . Les femmes ne sont pas rares dans la réalité, mais elles le sont dans la représentation des données, et cet écart se creuse à chaque mise à jour du modèle .

Caitlin Kraft-Buchman a approuvé ces observations et a avancé que le débiaisage des ensembles de données existants n'est pas une véritable solution : il peut atténuer les préjudices, mais ne peut pas résoudre le problème sous-jacent . Cela implique que le domaine doit s'orienter vers la construction de nouveaux ensembles de données représentatifs et de modèles d'IA sectoriels à partir de zéro, plutôt que de tenter de corriger des grands modèles de langage ayant déjà absorbé des décennies de données biaisées . Elle a également observé une volonté croissante des États à créer des IA souveraines et des modèles de langage plus petits et sectoriels, ce qui pourrait créer une opportunité de construire une IA en santé sur une base plus représentative .

#

L'ancrage du biais de genre dans les infrastructures numériques publiques

Yu Ping Chan a abordé le défi plus large de la construction d'infrastructures numériques publiques sur un substrat biaisé. Elle a cité l'Indice des normes sociales de genre du PNUD, qui a sondé 80 pays représentant 85 % de la population mondiale, et a constaté que près de 9 personnes sur 10 - 90 % des hommes et 87 % des femmes - entretiennent au moins un préjugé à l'égard des femmes, avec une décennie de stagnation dans les progrès sur cette mesure . Cela signifie que les données alimentant les modèles d'IA ne sont pas seulement techniquement incomplètes, mais socialement biaisées, encodant des hypothèses inconscientes sur les rôles et les capacités des femmes qui sont restées stables pendant une génération.

Yu Ping Chan a averti qu'un écart mondial de 10 à 15 % dans la connectivité et l'utilisation d'internet entre les genres - atteignant 30 à 40 % dans les pays en développement - signifie que les femmes sont déjà sous-représentées dans les données qui alimentent les systèmes numériques . À mesure que les infrastructures numériques publiques continuent de se développer sans remédier à ces biais, les inégalités ne sont pas simplement préservées, mais ancrées et aggravées . Elle a reconnu que le secteur technologique a historiquement résisté aux quotas et aux réglementations strictes, et qu'elle ne savait pas vraiment quelle était la solution .

#

Données manquantes : étapes de vie féminines et informations générées par les patientes

Un thème central de la discussion était l'absence quasi totale de données capturant les expériences biologiques propres aux femmes. Oriana Kraft a noté qu'il n'existe aucun formulaire standard dans les systèmes de santé ou de ressources humaines pour enregistrer si une femme traverse la ménopause, connaît des changements de cycle menstruel ou se trouve en période post-partum . Il ne s'agit pas d'une omission mineure : les étapes de vie féminines interagissent avec chaque organe du corps, et des affections telles que la prééclampsie et le diabète gestationnel augmentent le risque de maladie cardiovasculaire à vie de deux à quatre fois . Sans collecte systématique de ces informations, il est impossible de surveiller les femmes de manière appropriée ou de procéder à une détection précoce susceptible de réduire les comorbidités ultérieures et de sauver des vies .

Le cycle menstruel lui-même contient des signaux cliniquement significatifs que les instruments actuels ne sont pas conçus pour capturer. Le profil immunitaire d'une femme évolue tout au long du cycle, affectant les taux de réponse à la vaccination, les effets indésirables médicamenteux et même les résultats de la chimiothérapie . Il ne s'agit pas d'effets marginaux, mais potentiellement transformateurs, pourtant l'infrastructure de données permettant de les capturer n'existe sous aucune forme systématique . Pendant ce temps, les femmes génèrent déjà des données de santé riches grâce aux objets connectés, aux journaux de cycle et à des notes personnelles détaillées, et apportent ces informations lors de leurs consultations médicales, où elles sont largement ignorées . Les femtech ont constaté que les femmes sont prêtes à donner leurs données simplement pour bénéficier de meilleurs soins et contribuer à l'avancement de la santé féminine dans son ensemble . Le paradoxe, comme l'a observé Oriana Kraft, est que cette volonté existe précisément au moment où les grandes entreprises d'IA déclarent manquer de nouvelles données humaines .

Une étude non précisée citée par Oriana Kraft a révélé que les notes des médecins femmes étaient environ deux fois plus détaillées que celles de leurs homologues masculins, ce qui suggère qu'inciter à une collecte de données nuancées - notamment en récompensant les cliniciens pour une documentation plus riche - pourrait constituer une étape pratique vers l'amélioration de la qualité des données d'entraînement . Le point plus large est que le problème ne réside pas dans un manque d'informations disponibles, mais dans l'incapacité à les capturer, à les valoriser et à les intégrer dans les systèmes qui comptent.

#

La dimension intersectionnelle : les femmes de couleur et le Sud global

La discussion a reconnu, sans toutefois pleinement résoudre, les couches supplémentaires de désavantage auxquelles font face les femmes de couleur et les femmes du Sud global. Même aux États-Unis, les données sur la mortalité maternelle sont insuffisamment nuancées, les femmes afro-américaines étant confrontées à des taux disproportionnellement élevés qui ne sont pas adéquatement reflétés dans la recherche ou les recommandations cliniques . Les facteurs de risque spécifiques à certains groupes ethniques, tels que le risque accru de certaines formes d'anémie chez les femmes d'origine sud-asiatique, ne sont pas systématiquement intégrés dans les protocoles de dépistage . Le modèle universel, prévu pour un jeune homme blanc en bonne santé, ne fonctionne tout simplement pas pour la majorité de la population mondiale .

Une participante de Tanzanie a suggéré que des recherches ancrées dans les expériences des femmes africaines pourraient produire des avancées scientifiques transformatrices, notamment concernant la ménopause, y compris pour les femmes d'autres régions du monde qui connaissent aujourd'hui ces symptômes . Elle a également noté que si ses enfants de neuf et dix ans portaient déjà des lunettes, elle-même n'avait aucun problème de vue à près de cinquante ans, une illustration supplémentaire de la valeur potentielle de recherches ancrées dans des populations spécifiques. Cette contribution a renforcé l'argument selon lequel les populations les plus exclues de la recherche peuvent contenir les informations scientifiques les plus précieuses. Yu Ping Chan a soulevé la préoccupation supplémentaire : les données sur les femmes du Sud global doivent tenir compte non seulement de la race et de l'ethnicité, mais aussi de la langue, du contexte culturel et des conditions de vie très différentes dans lesquelles vivent les femmes des pays en développement .

#

Solutions proposées : normes, marchés publics et modèles réglementaires

La discussion a généré plusieurs propositions concrètes pour remédier à ces défaillances systémiques. Caitlin Kraft-Buchman a proposé d'établir la représentativité comme norme scientifique formelle, en s'appuyant sur le principe d'égalité substantielle inscrit dans la CEDAW, le traité international sur les droits des femmes, qui prévoit que l'égalité signifie garantir que chacun reçoit ce qui est véritablement adapté à ses besoins, plutôt qu'une solution universelle . Dans un contexte de santé, cela signifierait s'assurer que les personnes d'une population donnée soient sincèrement représentées dans les données utilisées pour entraîner les modèles destinés à ce cas d'usage .

Yu Ping Chan a proposé des exigences en matière de marchés publics comme mécanisme plus immédiatement actionnable, arguant qu'il est plus facile d'exiger des données désagrégées par sexe et une transparence algorithmique avant de signer un contrat gouvernemental que de vérifier la conformité après coup . Cette approche placerait la charge de la preuve sur les fournisseurs de technologie et les concepteurs de modèles avant qu'ils ne se voient attribuer des contrats publics, plutôt que de s'appuyer sur des engagements volontaires ou une responsabilisation a posteriori. Les pays nordiques ont été cités comme modèle réglementaire : plusieurs ont rendu obligatoire la collecte de données désagrégées par sexe dans les registres nationaux, permettant aux chercheurs de suivre l'impact sur l'ensemble du cycle de vie des problèmes de santé chez les femmes et de constituer la base de données probantes nécessaire pour impulser un changement systémique .

Oriana Kraft a également décrit une approche de concertation multipartite qui réunirait des ministres de la santé, des entreprises pharmaceutiques, des systèmes de santé et des institutions financières pour remédier à la nature cloisonnée du problème . Chaque acteur du système de santé tend à pointer du doigt un autre : les cliniciens veulent de meilleurs médicaments, les entreprises pharmaceutiques veulent des incitations gouvernementales, et les gouvernements veulent des données cliniques. Ce phénomène rend la responsabilité collective essentielle . Une enquête auprès d'environ un millier de médecins dans six pays a révélé qu'environ 80 % reconnaissaient des différences biologiques entre les sexes chez leurs patients, mais estimaient manquer des outils, des ressources et des recommandations cliniques pour dispenser des soins adéquats . Cela suggère que le problème n'est pas principalement lié aux attitudes des cliniciens, mais à l'infrastructure systémique.

#

Donner aux femmes les moyens d'être productrices de données

Une idée est née durant la discussion : les femmes devraient devenir des productrices actives de données, ainsi que des propriétaires et vendeuses potentielles de données de santé. Caitlin Kraft-Buchman a proposé de former des groupes de défense des droits des femmes pour qu'ils deviennent collecteurs, gestionnaires et propriétaires de données, leur permettant de vendre ces données à des marchés qui en manquent actuellement, et aussi de comprendre le fonctionnement de leur propre écosystème d'information, notamment dans le contexte de la mésinformation et de la désinformation . Cela fonctionnerait à plusieurs niveaux : améliorer la qualité des données de santé, renforcer les capacités locales et créer des opportunités économiques pour les organisations de femmes . La vision plus large était celle de femmes productrices d'informations et de solutions numériques, et non simplement bénéficiaires de la technologie. Cela signifierait que les femmes n'auraient plus simplement accès au système mais également la capacité de déposer des brevets, de créer de nouvelles technologies et de détenir les moyens de production numérique .

De plus, le Sud global devrait devenir une source potentielle d'innovation plutôt qu'un simple récepteur. Les technologies développées sous contraintes de ressources peuvent dépasser les infrastructures traditionnelles et être ensuite appliquées dans le Nord global, y compris dans les zones rurales reculées des États-Unis . Caitlin Kraft-Buchman a prolongé cette réflexion en suggérant que les pays africains, avec leurs populations diversifiées, pourraient collecter des données de santé nuancées et les revendre au Nord global, qui manque de données sur ses propres populations racialement diverses . Ce renversement de la direction habituelle du transfert de technologie a été identifié comme une opportunité à la fois scientifique et commerciale.

Une participante de Chine a demandé spécifiquement ce que les individus peuvent faire au quotidien pour attirer l'attention sur l'endométriose et l'incapacité du système médical à servir les femmes, en particulier dans des contextes où les manifestations publiques sont difficiles. Oriana Kraft a répondu en soulignant le rôle de certaines célébrités dans la sensibilisation aux problèmes de santé des femmes, et les réseaux sociaux comme moteur important du progrès dans ce domaine, notant que les femmes partageant publiquement leurs expériences avaient contribué à faire évoluer la conversation d'une manière que la mobilisation classique n'avait pas toujours réussi à accomplir.

#

Le biais de genre de l'IA dans les interactions quotidiennes

Une participante de Chine a introduit une illustration frappante de la façon dont le biais de genre de l'IA se manifeste dans l'usage quotidien, décrivant des publications sur les réseaux sociaux chinois conseillant aux femmes de dissimuler leur genre lorsqu'elles interagissent avec l'IA, car les systèmes d'IA répondent de manière plus logique aux utilisateurs se présentant comme masculins et de manière plus émotionnelle à ceux se présentant comme féminins . Cette intervention a suscité une réponse de Yu Ping Chan, qui s'est adressée à l'intervenante en tant que femme de couleur d'origine est-asiatique, reconnaissant les pressions culturelles particulières autour de l'affirmation de soi et de la non-confrontation qui aggravent le problème . Son conseil pratique était que les femmes rédigent des instructions explicites dans leurs requêtes adressées à l'IA, en précisant qu'elles attendent des réponses logiques plutôt qu'émotionnelles, et qu'elles n'acceptent pas les résultats biaisés selon le genre comme inévitables .

Oriana Kraft a apporté un contexte important, expliquant que ces réponses biaisées émergent parce que les systèmes d'IA ont été entraînés sur des données internet qui encodent des générations de stéréotypes de genre et de normes sociales . Elle a également mis en lumière une forme connexe de biais algorithmique : le mot « vagin » est effectivement le mot le plus censuré sur internet, et des affections de santé telles que l'endométriose et le post-partum font l'objet d'un shadow-ban sur les plateformes de réseaux sociaux, tandis que les termes équivalents relatifs à la santé masculine ne le sont pas . Cela signifie que le problème ne concerne pas seulement les données d'entraînement, mais aussi les choix algorithmiques intégrés dans les systèmes de modération de contenu, qui suppriment activement les informations sur la santé des femmes et réduisent encore davantage leur représentation dans les données utilisées pour entraîner les grands modèles de langage.

#

Traduire les engagements en actions

Un thème récurrent était la frustration partagée par toutes les intervenantes face à l'écart entre les engagements internationaux et les résultats concrets. Yu Ping Chan a décrit la Déclaration de Hambourg sur l'IA responsable pour les ODD, un document multipartite qui inclut le genre comme domaine prioritaire spécifique et élaboré dans le cadre de la collaboration du PNUD avec le gouvernement allemand, comme un exemple d'engagement qui existe sur le papier mais fait face au même défi de mise en œuvre que d'innombrables autres cadres internationaux . Elle a reconnu que la communauté internationale est habile à produire des documents, notamment le SMSI, le Pacte numérique mondial et la Déclaration de Hambourg, mais échoue systématiquement à les transformer en actions mesurables et responsables . Une participante a exprimé directement la frustration commune, notant que malgré cinq années de discussions soutenues sur l'intégration des perspectives des femmes dans les technologies et les produits pharmaceutiques, les progrès n'ont pas été à la hauteur du volume des discussions .

Caitlin Kraft-Buchman a proposé que les indicateurs e-santé du SMSI puissent être mis à jour pour exiger des données d'entraînement représentatives du genre, avec l'OMS et l'UIT comme facilitateurs, créant une cascade de responsabilisation à travers la normalisation internationale . Oriana Kraft a introduit le concept d'une « course aux données de santé », dans laquelle les pays se disputent la primauté de la collecte de données de santé complètes, diversifiées et nuancées, comme une incitation compétitive positive, en particulier pour les nations cherchant une reconnaissance scientifique et un avantage commercial . Le message de clôture était celui d'un optimisme prudent : l'opportunité scientifique représentée par les données inexploitées sur la santé des femmes est immense, la volonté des femmes de contribuer leurs données est déjà démontrée, et les outils politiques, tels que des exigences en matière de marchés publics, des normes scientifiques et des mandats réglementaires, sont déjà disponibles . Ce qui reste, c'est la volonté politique et la coordination institutionnelle nécessaires pour les déployer avant que de nouvelles infrastructures numériques ne soient construites sur une base déjà inégale .

Caitlin Kraft-Buchman
Thank you. Thank you. Thank you. And I'm so grateful that you're here, that you found your way here. And it's early in the morning and there's chaos and so much else that's interesting and fascinating to do. I'm Caitlin Craft Buckman from Women at the Table. And I'm here with this amazing panel. Yeah. OK. People connecting and doing. And I think our. Another point is that high -risk AI, as you all probably already know, because I'm sure that you are all experts in your own sectors, if not this one at least. So we know that in hospitals, in courtrooms, in systems, you know, who's deciding who gets a loan, who gets diagnosed. All of these are being judged on the averages, as you know, from the technology, right? So there's the privileged center of people. I would be somebody in the privileged center. And the farther away you are from the center of privilege, the less you are in the data, the more rural, the more poor, the more disabled. It just goes out. And that way machine learning works is that, first of all, there's averages. So there's a little bit of the line becomes very flat. And then there's recursive learning, learning that learns on itself, comes in. That line sort of gets shorter and shorter. There's scientists in the room. You may or may not like my. particular metaphor, but to this idea that getting smaller and smaller until even that privileged center probably doesn't have as much of a prominent place. So this is a problem writ large for women all over the world, compounded and super urgent for women with more intersections, but it is a problem, I think, for us all. And I think that one of the ways, um, anyway, so we, we just think that it's completely urgent. We're going to have, we're, um, I'm a big admirer of both people on this panel. So, um, I'm going to wind up having more of a conversation maybe then, which is great instead of a canned panel. Um, we have, first of all, the regrets from Ambassador Kha, very sadly was called to the, because he's very popular, to the global dialogue, um, where he has moderating a panel that seems to be sooner rather than later. And his, uh, So his actually wonderful and brilliant staff is at the Human Rights Council because there's something happening. They're negotiating a settlement. So here we are with us without them. But we're going to carry a lot of what Ambassador Carr has brought forth, which is about representativeness in the data and how that works and what that means here, too. So I'm here with Yuping Chen from UDP, who is a mover and shaker. And we will hear more about her in a moment. And Oriana Kraft, who is the founder of FemmeTechnology .org. So we've asked Oriana just to sort of start us off to do a little bit of a landscape and to sort of walk us through some of the work that FemmeTechnology did with also women at the table on AI. So would you share your screen and take us
Oriana Kraft
through, please? Yes. And this is the microphone. Can everybody hear me? Okay, awesome. Yeah, so maybe just as a primer, because some people are aware and some people are not, and it shocks them to find out, there are sex differences in every single cell in the human body, which means the way every organ functions between men and women is functionally different. The way diseases present is functionally different. But we've only really studied men and through every layer of kind of the research and clinical stack. So we study male cells. We study male animals like mice. We don't even use female rodents. And we've only really functionally studied men and usually only kind of healthy 70 kilogram white men. And what does that mean? That results in kind of shocking stats like the fact that women are diagnosed on average four years later for the same disease as men across 770 diseases. Women are 50 percent more likely to die following a heart attack, despite it being the leading cause of death in both men. Women because. while physicians are only really trained on the way cardiovascular disease presents in men, but also because the tools we use to diagnose cardiovascular disease are based on the way it presents in men. So basically troponin levels, which is a biomarker for cardiovascular disease, the cutoff rate is based on the male body and it's lower in women. And so a lot of cardiovascular disease gets dismissed that way. Also, the imaging technology that we use is based on the vessels that are more impacted by men. Basically, men and women have different vessels that are more likely to cause a heart attack. And so that's also missed. So kind of every layer of the stack, you're not catching it. And then also, once women have had a heart attack and are diagnosed, they're not as likely to be prescribed gold standard treatment. So that's why they're also more likely to die. And then even following that, there was just a big study that came out two years ago, I think, which was called the REBOOT trial, which found that the treatments that are considered gold standard are not as effective in women. So at every layer, right? Yeah. And so maybe to show why this is a problem when it comes to models, we kind of created this cascade of distortion showing like what that stack is. So on the first kind of echelon, you have the actual clinical research that's done. That's kind of what I explained at even the preclinical level that's studying cells and studying animals. And we don't even use kind of female models there. And there are, as I mentioned, since there are sex differences in every single cell in the human body, it matters whether you use male or female cells. But we don't. From from C, as you can see, male animals outnumber female animals five point five to one. And this is actually based on a false assumption. Basically, they decided women or even female mice were too complicated because of the menstrual cycle, because of fluctuating hormones. But it actually turns out that male male mice are more unpredictable than female mice. Because with the female mice, you at least have a predictable cycle. So you kind of know what hormones are going to fluctuate throughout. But with male mice, with testosterone, it's a much more dramatic increase. So it's less predictable. So actually, it's totally based on a completely false assumption. Then at the next echelon, because you've only studied. male cells, male animals, male bodies, which results in, by the way, women are 50 to 75 % more likely to have adverse drug reactions because we've not included women in equal numbers in clinical trials. They were effectively banned until 1993 in clinical trials, which means basically all the drugs on the market were never tested in women's. You're kind of doing like a live experiment. But you have that stack with the clinical research. But because you only study them there, we create our clinical guidelines on the basis of that research, which means that they're also all based on men. So the kinds of examples that I gave with troponin, with the cutoff being too high for women, with the way cardiovascular disease presents what you should watch out for. I mean, I was in medical school myself not that long ago, only a couple of years, and we're still only trained in the way it presents in men. And then you kind of have an asterisk. And they're like, but remember, it presents differently in women. And women are 50 % of the population. And it's still kind of not the standard of care. in training. And so, yeah, as you can see, troponin thresholds calibrated on men miss 42 % of female heart attacks. I mean, that's a huge problem. Then what that means is that endometriosis, it takes on average seven years to get diagnosed. And, you know, I think PMDD, which is post -menstrual dysphoric disorder, takes something like 15 years, right? There's, I mean, you're just not trained to recognize the way it presents in women. What does that then mean? If you don't have the clinical guidelines, if the clinical guidance are not reflecting reality, women are more likely to be misdiagnosed, kind of like the example that I gave with endometriosis. Women with endometriosis have to see on average five physicians before they get an accurate diagnosis. What does that mean? That means that every physician the person has seen before is recording an inaccurate diagnosis. And unless you live in somewhere like a Nordic country like Denmark, actually, which has done a lot of really interesting large scale research, you're never getting that updated diagnosis. Right. Because you don't you don't have it linked to the patients. You have all these discordant records and kind of with a condition like endometriosis, a ratio of of four times as much misdiagnosis as an accurate diagnosis. And nobody's going back and essentially cleaning that up or updating it. And, you know, a lot of the large language models. Right. Because the theme of this is all around AI are trained on that inaccurate data and nobody's going back and seeing what is the actual reality. And then so then, as we kind of said, the models are then trained on that biased history and then AI trained on misrepresentative data reduces diagnostic accuracy by eleven point three percent. You know, there's a lot of people who think that you can just kind of have explainable AI. So if you have the model, explain the reasoning that it will improve it. But it it turns out, no, like physicians will it still reduces accuracy by eleven point three percent. So you really need the accurate data. to have accurate algorithms. And then, yeah, I've kind of alluded to sort of some of those negative outcomes that it have. Women are twice as likely to have adverse drug reactions as men. And they also spend there's like this common misconception that because women live longer, that means that they're healthier. But women live longer, but spend five more years in poor health than men. So their quality of life is not as high. And it actually doesn't come at the end of life. People think it's OK. They live longer. And so it's stacked at the end. It comes in their prime working years. So it has a real outcome on their ability to earn financially. I don't know. I'm happy to go into more
Caitlin Kraft-Buchman
detail, but I think that that was a big, big overview. I don't give it to the moderator. That's fabulous. And we want to hear more. And I'm sure everybody's got questions that we'll go through, too. So what does that so that's the. That's the reality that we're living with. How would, you know, now you're building all this extraordinary DPI, all this digital public infrastructure. We're building out. We want to have countries have access to this. But what do we do when we're building on top of something that actually in
Yu Ping Chan
its very core is not serving all the population? So it's your piece. Thank you so much, Caitlin. And it's really great to be here. Like Caitlin knows, I'm quite a big fan about really pushing on the gender aspect when it comes to digital, because it's so often in some ways forgotten. And there are too many men, frankly, in tech. I was just actually on a panel where it's all women. So it was actually remarked upon that this is one of the few times that we actually have an all women panel. And I'm really proud to be on one with you as well, because really. it's really kind of frankly in tech circles and if you look at AI for good and WIS is all too often women really are disproportionately underrepresented in a lot of this so it's amazing to have this conversation I really applaud also the men that are in the room really thank you for being allies in this and we really need to keep pushing because if not really gets overlooked on that note I completely agree that like as we're building up digital public infrastructure and digital systems right we have to recognize the biases that already exist in the data and the substrate on which we're building these types of foundations so for instance UNDP did a gender social norms index in 2023 where we surveyed 80 countries 85 % of the global population and we found that close to 9 in 10 people 90 % of men 87 % of women hold at least one bias against women and there's been a decade of stagnation in terms of progress on this so the data itself is based on a substrate of already inherent assumptions about men and women and so the models the data already based on that so the data that feeds the models is built already on unconscious biases or certain types of norms and expectations and then you're right Caitlin as we build this sort of digital public infrastructure that draws on those types of of biases and so forth, we're in some ways hardwiring these kinds of assumptions into how we build digital public infrastructure. And at the point that already is a digital gender gap, right? I think the global surveys are something like 10 to 15 % gap between women and men in connectivity and use of the internet that grows to something like 30, 40 % in these developed countries. It becomes sort of a worsening problem where the more we keep digitally transforming without actually addressing these types of questions, the worse the biases will get and the models will basically be built in an asymmetrical, disproportionate type of way. I'm not very frankly, very sure how to fix this besides just sort of continuously calling attention to this fact. Because I do think that unless there is a very clear commitment from the model builders, from the big tech companies, from the people that are actually coding and building it, that they're going to address these types of issues. proactively fix for it, I don't really know what is the way to do it. So I know there have been, like, in the general gender discourse, right, like pushback against having quotas or very firm types of regulations that say you need to address this type of issues. And in the tax circle, I know we've shunned away from this type of stuff, but I'm not actually very honestly and being very candid here, sure what is the fix for it? Because as you pointed out, these things exist. We know these things exist, but yet we've not really taken that proactive action to address it. Can there be sort of guidelines or expectations that we set that we say we expect these types of things out of international organizations, national governments, and private tech companies? Maybe that's a start to call for that kind of gender accountability when it comes to algorithmic transparency, data, a commitment to update, like, data sources when this happens, the models and so forth. Maybe that's the kind of thing that we should be calling for, because if not, we talk about the problem, we clearly recognize that there
Caitlin Kraft-Buchman
Yeah, I don't think, I mean, I think everybody now agrees that there's a fix. We all agree that actually de -biasing the data is not possible, really. You can sort of make it less awful, less bad. You can mitigate a bit, but it doesn't, isn't solving the problem. I do think that as everybody is also turning somehow to sovereign AI for different geopolitical reasons, that we're all going to arrive at the fact that smaller language models for sector -specific reasons are going to work, right? So you could have like a health model that works if you built it from scratch, which will be my next question to Oriana. But one of the ways to get to that also is that we're starting to look at, could we call for, representativeness? as a scientific standard. So when we say standard, it's also a word that, you know, we're all using what does standard mean, but like really a scientific standard that we agreed to so that you have in a use case, a specific use case, then you have the people in that population that are really sincerely represented. In a general population, it would be half women, but in some populations, if it was a sign language thing, it would be the people really generationally and demographically within the deaf community. And that may be a way. We would also be looking towards CEDAW. We have a panel with CEDAW later about like where does substantive equality, right, because CEDAW, the treaty that everybody signed and agreed to, is that there's a difference between equal opportunity or general equality. There's a lawyer here who will tell me what it is. But at the end of the day, substantive equality means is that everybody gets the sort of the bicycle. the size bicycle that would fit them as opposed to everybody gets a one size fits all that you really get things that are adapted to. Anyway, we can go into those means, too. But if Oriana, if we were going to build something from scratch, what are the kinds of things that you need to see in the capture the data differently and what and what are the opportunities for that?
Oriana Kraft
Yeah, maybe just kind of building off of what you said about the tech sector not being a big fan of quotas and mandates and kind of insights. We so we did an event recently in March with a lot of health ministers in New York where kind of the the idea behind it was to bring people from government, from the pharmaceutical industry, from health systems and from the financial industry, because really with health, part of the issue we've seen is if you only take a siloed approach, everybody is going to point the finger at someone else. Right. The clinician is going to say. you know give me better pharmaceuticals that work for women that don't cause twice as many adverse drug events and you know like I'll do a better job or give me the fundamental research and I can have the guidelines right the clinicians don't have the tools we actually also did a survey where we surveyed I think like a thousand or two hundred physicians across six countries which showed that something like 80 percent of them you know they see sex differences in their patients but they don't feel that they have the tools the resources the clinical guidelines to be actually be able to deliver adequate care so it's not that clinicians don't see the problem and then you have the stat that 80 percent of women feel dismissed by their physicians so you have kind of this imbalance on both sides but the point of the event was to bring all these stakeholders together because everybody points a finger at someone else right like then the pharmaceutical industry will say like government has to incentivize me right to run these clinical trials or to innovate for diseases that impact women disproportionately and maybe just to say there are some countries I think like the Nordics are really the most important countries in the world and I think that's a really good point I think that's a really good example I've already pointed to them that mandated sex disaggregated data collection across national registries, because with an issue like health, you kind of have to see the full life cycle effect. You have to see how it impacts like a woman's like lifetime earning potential, what it's costing your health system to not innovate for it. So I think that that could be an incentive to kind of because I think everybody's just very siloed and how they're acting to your question, right, about what is the kind of data that we would need to collect? Well, one is we need to collect sort of a lot. It might be surprising to you to know that all women go through menopause, but there is no basically button or form in any HR to capture that a woman is going through menopause. Right. So we have forms for capturing kind of blood pressure, you know, various other body parts, but there's nothing around that. There's nothing around the menstrual cycle. There's nothing around menopause. There's nothing really around postpartum. And so an example here, we would need to capture female life stages because they really they interact with every organ in the female body. So pregnancy. especially things like preeclampsia or gestational diabetes cause a two to four times lifetime increased risk of cardiovascular disease. Right. And so you would want to be capturing that information to then monitor the woman and be able to do early detection, which could at the very least it could potentially save her life, but also just kind of reduce downstream comorbidities. So female life stage factors. I mean, I think like in medicine as a whole, you would just want much more. Nuanced data collection. I think the EHR kind of served its purpose, but it reduces things to very like flat binaries and health is much more multifactorial. And part of the problem is also that you have all these siloed organ systems that are not talking to each other. So you're not capturing kind of like nuanced signals. Or is that was that was that your question around what kind of things would you want to increase? I mean, specialists, doctors, specialists who don't speak to each other or the organs don't speak to each other. I mean, right now, the way our. medical system is designed is right. You have an ophthalmologist, you have a cardiologist, you have an endocrinologist. So you're having all these things in silos. So that's one issue, right, just with medicine as a whole. Then the other issue, particularly as it relates to women, is there's no capture of the life stage. There's also no capture of the menstrual cycle, which is incredibly important because a woman's immune profile changes throughout the menstrual cycle. So you have better kind of vaccination response rates and less adverse drug reactions, depending on when you vaccinate a woman in the menstrual cycle. Right. And like that could be transformative for a woman's life. You also have better, interestingly enough, responses to chemotherapy. Right. Like there are there's a lot of signals in the female body that we haven't like designed the instruments to collect. Part of the reason we don't have the data is we have like no easy way to capture that information and nobody is incentivized to capture that information currently. But we had, you know, it might it might seem kind of controversial, but one of our recommendations at the end of this very long paper that is very interesting to read, I promise, was... It's online, so you can all have access to it. So don't worry if you're not capturing the slides. Okay. One of the recommendations we had was incentivize female physicians to capture the information because they found that female physician notes were, like, twice as detailed as male physicians, right? And, I mean, if male physicians also have detailed notes, that's great. But, like, they might all, you know, you want richer information, right? And right now people are just training on the data, but there's not, nobody's really incentivizing nuanced data capture. Another way you could do it is incentivize patients to capture their own information, right? Women are coming to physicians with aura rings, with, like, cycle logs, with their own handwritten notes, and, like, where's all that information going? Maybe if you have a physician who believes you, it's being captured in the HR, but it's not being captured, and the patient. It's actually the richest source of information possible. And what we find with a lot of femtech startups is women are willing, you know, I think they should be financially incentivized, but women are willing to donate their data just to have better care and to contribute to advancing women's health care as a whole. So it's not that people are not willing to participate. They are. It's just that we're dismissing this incredibly rich source of data, which is somewhat ironic when, you know, all the big AI companies and model labs are saying they've run out of fresh human data. And so they're just going to continue training on synthetic data. And part of the problem with synthetic data as it relates to women's health is each time you train a new model, it kind of loses the signals from the tail ends of the cycle. So if it kind of goes like this at the tail end of the distribution, women are not rare in reality, but they're rare in the representation of data. So each time it gets trained, you lose that signal. And so you have all these signals as it relates to women's health that are being erased that already exist. And then you have also the problem of them not being capitalized. They're not being captured in the first place, right,
Caitlin Kraft-Buchman
with things like pregnancy, postpartum, menopause. okay so how do we turn this interesting uh semi -tragic state of affairs into into something that we can do something about and maybe one of one of those ways is through commitments to something like this maybe not that maybe not but maybe on the the hamburg declaration which you've had a really very pivotal part in in in um in crafting and driving forward and it names women and girls as its commitment so is there a way to sort of turn a commitment into action and to so caitlin's referring to undp's role with the german government in
Yu Ping Chan
drafting the hamburg declaration and the responsible use of ai for the sdg so this has been a product of the hamburg sustainability conference that's been running in hamburg you for the last three years. And actually, it was just a week ago that I was actually in Hamburg for the third Hamburg Sustainability Conference, where we were actually holding a session on how the Hamburg Declaration that started the year before has actually translated into impact in terms of the endorsing organizations and the areas of work that they've done. So we reported on, for instance, people being trained, and I think there was a lot of statistics around, for instance, the use of AI -trained systems resulting in better healthcare, not specifically on gender, but more in general. So the idea is that this was a multi -stakeholder declaration, which includes gender as one of the specific areas of action and priority for where private sector companies, international development organizations, and national governments could commit to seeing through these types of practices, responsible AI in these particular areas as well. And so that question of, like, how do you translate a commitment that's on paper into actual action? it's frankly very challenging and I think that this is the same conundrum that we're facing across the entire UN right we have countless documents we have WSIS we have the global digital compact and what do we do with these that then become meaningful in concrete ways that changes people's lives because frankly like we're very good at coming up with documents but like where the international community then falls short is turning this into actual action I literally was just on another panel where it was with the private sector community and the business community and it is exactly the same point right we all have committed to these things but what is the next step to actual actionable outcomes and specific aspects that you hold people accountable to I think you mentioned this idea around standards I am actually quite a big fan of this idea around procurement as well because frankly I do think it's a lot easier to require certain things before you actually sign a contract than to audit for it after so to some extent if you can have that kind of requirement or expectation that's already brought into the particular contract before you do it, that could be a way that we actually hardwire in these fixes around gender and the requirement around gender disaggregated stratified data. So for instance, if it's a health contract, right, is this something that we could actually consider building in? And perhaps you pointed to the fact that Nordic governments have actually been requiring this type of data disclosure and so forth. I don't know whether we can go further to require this in terms of like government contracts and procurement, but procurement seems to me of, well, not an easy, but at least a direct way of asking companies, builders and tech providers to make sure that this is at least part of what we do. I actually had a question for Ariana because, and this is something that I struggle with as UNDP sometimes, because our focus is, again, on developing countries and the global self. The global north, to some extent, is taken care of. But when you look at the global self, how do you see, for instance, the issue of the disaggregated data and the gender data, for instance, what is the extent to which it takes into account women of color? And then not just like women of color, but also the question of women of color that come from different countries of the world. and the fact that in some ways the data sets, the languages that we in the global south operate in, are vastly different from
Speaker 1
anywhere. I think in terms of, you know, how nuanced is the data? I think it's sadly not nuanced enough, right? If you look at a country like the U .S., which has just shameful maternal mortality, and especially as it comes to African -American women, which it's much higher, it's not nuanced enough. I think sex should just be the first step, right? There's this thing called like sex as a biological variable. Because there are sex differences in every single cell in the human body, it should be the first step of precision medicine. And then you continually stratify, right? A woman who's gone through menopause at 60 is completely different from a woman at 20, right? Women of different ethnicities are completely different, right? We know you're more likely, if you're of Southeast Asian descent, I believe, to have a certain form of anemia, right? Like that should be a risk factor in your screening. You should continuously stratify it. It's just not. It's not nuanced. They've kind of taken this model of like a young, just honestly, a young, healthy white man. Because we don't even have studies on older men, right? And extrapolated it out to everybody. And it doesn't work for everybody. It's this one size fits all. It's kind of like... which I think is the real opportunity with AI, before we couldn't study large -scale populations at scale. So we would take, I don't know, a sample of really sometimes 20 people, 30 people, and extrapolate it out to a population of 7 billion people. Instead of kind of doing it in the reverse, like collecting signals from as many people as possible continuously and trying to create these cohorts. And if you're here, you might have some of this cohort, you might have some of that cohort, you might have some of that cohort.
Speaker 2
Yeah. I mean, you were kind of saying, I mean, I guess your question was like how could we get more nuanced data or, yeah. What is the state of data now when it comes to, I think you just answered that question, right, women of color?
Speaker 1
Yeah. And my bigger concern is women of color in the global South as well. Well, the one thing that I will say is interesting. So the Gates Foundation does a lot of work when it comes to women's health, particularly kind of in the global South. And one of the things that they said is, is because as they said on. Fortunately, the state of women's health is so bad everywhere. It's worse, obviously, in some places. They're finding technologies being able to kind of leapfrog. So if you take this thing like this remote kind of ultrasound that you can do with a mobile phone, right, to like so you don't you're not requiring a physician. You can do this remote monitoring that was actually developed for the global south, but could then be applicable in the US or in other remote rural areas. So in some ways, kind of the constraints that are when you're in a lower resource setting that are forcing you to innovate are making there be able to like be innovations that could then be applicable also to the north. You could actually innovate for the south and then extrapolate it to the north, which is usually not the direction it
Speaker 2
goes in. Yeah. You know, you know, I think actually that's that's an opportunity for small states. It could be a Singapore thing. So that's not the global south. But do you know what I mean? In a small tech enabled to really to collect the data in a real. fabulous nuanced way understand it and then export the model i think that there's actual real financial sort of at the end of the day i've been saying to some of our african colleagues that if we really got the data right that you would be able to sell it back to the north because a they don't have data a and b they don't have it on their own populations um of their sort of racially diverse populations at all so um and it would be probably from a scientific point of view you'd want something well simple like super diverse but also super um also sort of mono focused anyway it would it would be like a scientifically great use case i think um we one of the things that we've been talking about besides for this representativeness um ness in the uh as a scientific standard and the data is also to maybe uh we would like to be able to implement this. So it's only an idea now, but it's to train women in essentially very, in any context, women's rights groups to be able to go out and collect data and to collect the data and to become the data collectors and the data owners and the data organizers and to be able to also sell that data back and manage that data. And that would work in a couple of levels, one for health, but also would work in sort of a misinformation, disinformation context, where they would also be able to sort of understand the forensics of their information ecosystem. Sort of happier about because it's like less syllables. But I think that there are a lot of opportunities for us to change. This notion of development of sort of protecting women. and going more to women as the producers of information, of solutions, that we stop. Only access is, of course, incredibly important, although we often say, as people focus on the political economy, access to what? But really, we're only measuring access because we want to understand what the destination is. We want to know how many people are accessing technology. This is a larger point, but going to how many women are filing patents, how are they enabled to express their creativity in terms of creating new technologies, in terms of creating and owning sort of the means of digital production. So that's sort of like destination. And this is a destination play, I think, this health data part. So we're sort of also looking to, incentivize and I hope inspire. Thank you. partners. People are able to go back to their governments to sort of say, you know, why don't we do this? Because I think that there's a real, there's a place for people to innovate and to really sort of change You had, I may open it up for questions or it may, does anybody have a burning question? We'll keep talking to ourselves.
Audience Member 1
Yes. Hi, ma 'am. Hi. And first of all, I learned a lot from your speaking and it reminds me of a post I have seen on China's internet. It is teaching girls to hide your gender when you are talking to AI, because it is found that if you make a mistake, you're going to be in trouble. So I think that's a really good point. If you act like a man, the AI will be more logical, but if you are a woman, the AI will be more emotional. So I want to know how this can be changed or how we can deal with it. Yeah, this is my
Speaker 2
I'll take a little step, but I bet you Ping has something really smart to say.
Yipeng
I don't have anything smart to say except to say that you should not live with it.
Speaker 1
Yeah, I mean, so just first of all, maybe I'm just going to kind of explain what you already said. It's like, you know, where does that come from, right? It's taken because it's been scraped from the Internet, stereotypes and social norms and all of those roles that we've inherited for generations. And it's absorbed them in terms of its reaction. Different people have different reactions to that. There's some people that say we need to flood. We, every older woman, every young woman. And girl needs to flood the Internet with positive representations of women. So that will overwhelm what's already there and the bias. I think that's what we need to do. I think that's what we need to do. I think that's what we need to do. I think that's what we need to do. that may or may not be the most effective, but that's an idea. At least it makes you feel that you have some agency to it. Some of it may be just like to go back to the very beginning and to work on, I think, these small sector models and not lean so much on the large language models that have these problems that we're never going to solve that. And I think that that's where I think kind of in sovereign AI we're going to go. But it's a horrible situation. I also think, on one hand, it's horrible, but it's already great that they're saying, that they're acknowledging that there's this bias baked in. I think that an opportunity, I think there's really low literacy when it comes to understanding the assumptions a model is making about you. And I think across the board you should have a right as a citizen to know what you're doing and what you're doing. And what assumptions a model makes about you. And I think in particularly high risk settings like the law or like health, you should be entitled to that knowledge. Right. Because right now people are deploying these models that are making assumptions about you and you don't know what it's on the basis of. It could be also your income level. Right. It's like all these kinds of things that are that are being made. And I think the kind of flooding the Internet with positive representations is unfortunately not a solution because it's not only a question of data. It's a question of the algorithm. And unfortunately, the majority of people shaping the algorithm are making assumptions about women. So one example is like vagina is actually the most censored word on the Internet. It's like basically shadow banned on social media, but also things like endometriosis, like postpartum, literally just health sector. Conditions are banned, shadow banned on the Internet. But like semen is not. Right. And
Yipeng
And I also specifically want to answer this because you're clearly from China, woman of color, Asian descent, and we are in our ethnicity particularly seen as, well, less assertive, submissive, tendency not to push back and non -confrontational. And so when you said, do we live with this? I really want to tell you don't, right? You push back. And I know it's hard culturally sometimes to really push back. But I would say that even in the basic use of the AI model, right? Write it in to say that this response that you wrote is overly emotional. I want a more logical response. I'm coded into the actual prompt itself and make it a standing instruction in your AI model. So I have very clear preferences in my AI models for how they interact with me. And this should be one of your standing instructions. And just say it in no uncertain terms that you expect the model to interact with you in this particular manner and not in this other way. And I will say this again to East Asian women. Asian women, because I see quite a number of you in the room, young women, this is not what we should be standing for. And that over time, we need to correct the way we interact with technology on that type of term. Thank you so much, Yipeng.
Speaker 2
It was this madam and then that madam.
Audience Member 2
Thank you very much for a wonderful presentation. It was actually eye -opening and in some of the areas I wanted to share an example from where I come from. I come from Africa, Tanzania. My country is Tanzania. What surprised me is that my grandmom, my mother, and the old ones, they never experienced like, I'm just giving one example, menopause symptoms. It's good if the researchers were based on women. and say of African -American origin, maybe they will find out why people do not experience worse menopausal symptoms and maybe it could have helped even our generation because now some of us are really experiencing it. So I'm just joining hands with what you presenters have said. I've also checked in the area of eyesight. I'm almost 50 and I've never had any problem with my eye, but my kids, only 9 and 10, they've already started wearing glasses. So doing research based on the actual group, let us say women, is very important. It's sad that all these years we've been treated, but we're still doing research based on the actual group. and data based on men. Maybe it's why sometimes even you use the medications and they don't work as to your expectations because of such mess up. I thought I should say that. Thank you very much. And I would love if you would give us the documents and sites so that we can read them all. Thank you.
Speaker 1
We do think absolutely. This is a new frontier. There's scientific opportunity. There's a lot of money to be made, but there's also a lot of great science and discovery to happen. So it's horrible to leave that on the table. We don't have to go to space. There's so much we need to really discover here.
Speaker 2
Hi, I'll give you the last question.
Audience Member 3
Thank you. I have a question about, and thank you for the panel. I think it was a great conversation to have. I have a question about advocacy and what can we do at this stage? Because I feel like for the last five years, there's been a lot of conversation around incorporating women in tech, incorporating women's health perspective. in the pharmaceutical sector but yet it doesn't feel like it's advancing at the pace that it should based on how many conversations we've had and how many organizations so i guess my question to you all is like okay 2026 where are the gaps in terms of advocacy and what can we do to help maybe move the speed a bit faster so that the generation that's coming up is able to not experience what we've been experiencing for now several generations
Speaker 1
having the health ministers and various leaders of different countries i think it's a i think it's a political issue um i think um i i actually think it should be a political issue and in that i think health should be a part of people's platforms when they're running I think we're seeing that increasingly become an issue. Right. So, like, for example, in Switzerland, health insurance has risen consistently every year. Right. In the U .S., it's becoming a cost of living crisis. But in many countries. Right. I mean, if you're not healthy, you don't have any quality of life. And I think, you know, I think now politicians should have to speak to what they're doing when it comes to women's health. Personally, it's not it's not there yet, but we are seeing in the U .S. there are now people who have I'm not a huge fan of lobbying, but there's people who've created basically a pack to lobby for women's health when they say there's all these special interest groups. Right. You know, AI being one of them. And why don't we have a special interest group when it comes to women's health? I think I think there's been kind of a lack of awareness. Right. Like it took my going to medical school and being curious. I wasn't even formally taught about these gaps. So I think when you have half the population being. essentially uneducated just because it's not talked about. You know, we have this assumption. I had an assumption being that I was an equal citizen. Right. And actually, fundamentally and functionally, you're not an equal citizen in health is just one example right across the law. And I think we do, though, have the right to vote. So for me, I think it's a political issue because you see that when the public sector sets incentives or regulation, particularly for the pharmaceutical industry, that that's the core buyer for the pharmaceutical industry is the government. I think that that's what would have the most leverage. But you need politicians to want to do that. And therefore, you need the population to only elect people
Speaker 2
Avenue, I agree with all that. I agree that I think basically we'll just have to keep pushing at it and really just asking for this to be, like you said, more concrete. And I share that frustration, really. I do think that this is the moment. To think about what is actionable in this space. and see what is possible to create coalitions around what is actionable. Yes, please. I just feel bad. I know you had a question. I don't know if you still have a question.
Audience Member 4
Yeah. At least not in the favor of women. And I also know that there is an end of March that happens every March just to promote the to call attention to this disease. But here in China, I'm from China here in China, the promotion of. let's say drawing attention to endometriosis is definitely hard to to be held in such a way like holding a march or holding a parade it is it tends to not work this way so i'm thinking that so my question is as an individual is there anything that we could do in our daily lives to call attention to not just endometriosis but to call on the fact that our medical system currently has not served women well and what can we do to promote to change
Speaker 1
that um well i first of all i think that you kind of are already doing that by by being here and by speaking out i think men have a huge platform um disproportionately right because it doesn't work if it's only women speaking i think in the u .s i can't speak to china but uh in the u .s and the uk part of what had a lot of impact was celebrities going out and sharing their individual stories So like you have Padma Lakshmi, you have like various supermodels have endometriosis and you start to see that it doesn't matter who you are. Right. Like you have Oprah kind of talking about menopause. It doesn't matter how affluent you are, how well known you are. You are impacted by the lack of research on these conditions and the lack of solutions. So I don't know if there's a like probably the equivalence in China. Right. In terms of like celebrities. And then I don't I as an individual kind of speaking out like I think like it's kind of weird to say, but I think social media is what has led to there being some progress in women's health. There's been a lot of conversation and suddenly people started to see that this wasn't an individual problem, but it was a collective problem. And so if you start to get lots of people sharing their stories, I think that that's kind of the first pillar. Yeah. And then also, like, it's a huge problem. It is like a huge financial opportunity. Right. When you have half the population being underserved, like in terms of like the therapies you could develop in the population, you could address, I think, making
Speaker 2
So as we before we close, I've got to say two things. One is one very concrete way, I think, is also for us to use WSIS since we're here and the indicators. So there's an e -health indicator and we all should be helping to encourage those facilitators, which are WHO and ITU, to train on gender representative data. And if we can put that and really have a series of indicators there, I think that that could also maybe start to change things in a cascade way. And another possibility, perhaps, is in a context for any country, is a country that wants to lead and wants to be the first and wants to show how great their science is. The way I think the way to get this point across, perhaps instead of like marching, is to say, we could be the first group ever to do that.
Speaker 1
Yeah. Yeah. And it could also make it work for our population in a way, because also we know in terms of the genome mapping, it all comes from one, basically one white guy. Right. Everything that we know about the human genome, which is super crazy to me. So I think that there's a scientific, you know, gold star to be gotten. And I think that that would be great if there was kind of like, you know, a health data race in that way and that maybe you can start that off and everybody would would follow. So that that's my that's my perhaps positive, not not specific way to end that. I know that people are very, very conscious of time. There are people waiting outside. We're conscious of your time and where you need to go
Oriana Kraft
Thank you so, so much. We'll have the link to this website. This incredible thing that Oriana built. And we're really very grateful for your time and your attention. Thank you.

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.