AIrbrush: a transdiciplinary value-sensitive study of biases and stereotypes in AI-generated Global Health images, and their significance for science and society

Informations projet

AIrbrush

N° de convention de subvention: 101150547

DOI

10.3030/101150547

Projet clôturé

Date de signature de la CE 19 Avril 2024

Date de début 1 Juin 2024

Date de fin 31 Mai 2026

Financé au titre de

• Marie Skłodowska-Curie Actions (MSCA)

Coût total

Aucune donnée

Contribution de l’UE

€ 191 760,00

Coordonné par

INSTITUUT VOOR TROPISCHE GENEESKUNDE
Belgium

Periodic Reporting for period 1 - AIrbrush (AIrbrush: a transdiciplinary value-sensitive study of biases and stereotypes in AI-generated Global Health images, and their significance for science and society)

Période du rapport: 2024-06-01 au 2026-05-31

Résumé du contexte et des objectifs généraux du projet

Global health communication relies heavily on images. Photographs of poverty, disease, suffering, care and empowerment have long shaped how global health problems and interventions are understood. These images influence who is seen as vulnerable, who is imagined as capable of expertise or care, and whose dignity, consent and agency may be put at risk.

Generative AI has made these questions more urgent. Images of people, places and health crises can now be produced in seconds, without direct contact with the communities being represented. The overall objective of AIrbrush was to understand how visual representations of global health are produced, justified and transformed in the AI era, and what this means for ethics, policy and society.

The project examined both human and institutional practices behind global health imagery and the automated outputs of AI systems. It asked how stereotypes are created and repeated, how visual professionals navigate ethical dilemmas, and how AI may reproduce, modify or intensify existing problems of representation.

AIrbrush was a social sciences and humanities project, drawing on anthropology, visual culture studies, science and technology studies, AI ethics and critical global health. These approaches were essential because the project did not treat AI bias only as a technical problem. Instead, it examined the social histories of representation that AI systems inherit, and the institutional incentives that shape visual communication.

The expected impact was to help global health institutions, humanitarian organisations, funders, communicators and policymakers develop more responsible approaches to visual communication in the AI era.

Travail effectué depuis le début du projet jusqu’à la fin de la période considérée dans le rapport et principaux résultats atteints jusqu’à présent

The project carried out its empirical and analytical work while adapting to rapid technological change in generative AI. The research combined visual analysis, stakeholder interviews, live platform auditing, media monitoring and interdisciplinary exchange.

First, the project identified and analysed more than 100 images published or circulated by non-governmental, humanitarian and global health organisations on their social media. This analysis examined recurring tropes, including poverty, suffering, ethnicity, childhood, expertise, care, empowerment and technological intervention. It showed how familiar patterns continue to structure global health communication, including where AI tools are beginning to be used.

Second, the project conducted 29 semi-structured interviews with visual and technical stakeholders, including AI artists, technology specialists and communication professionals working at NGOs. These interviews explored image decision-making, ethical risks and responses to generative AI. A key finding was that many dilemmas now associated with AI were already present before AI: images were already shaped by staging, editing, donor expectations, organisational branding and pressure to produce emotionally effective stories. AI can accelerate and automate these existing tensions.

Third, the project audited commercial AI tools, stock-image environments and media platforms, documenting how AI-generated global health and humanitarian-style images can be produced, sold and circulated at scale. This showed that commercial image infrastructures play an important role in making synthetic stereotypes visible, searchable and reusable.

Fourth, the project co-organised an international workshop on artificial images in global health before and in the era of AI. The workshop brought together 27 experts from five continents and strengthened the project’s historical perspective by situating AI-generated imagery within a longer history of image manipulation.

The project produced scientific and public-facing outputs on the ethics of global health communication in the AI era, racialised visual stereotypes, and the need to understand AI bias as a social, historical and institutional issue as well as a computational one.

Methodologically, the project adapted from an initially planned emphasis on systematic generation of synthetic images towards a broader value-sensitive audit of real-world AI imagery and live commercial platforms. This increased the relevance of the research while reducing the need for resource-intensive image generation.

Overall, AIrbrush achieved its core aims by generating new empirical evidence, developing an interdisciplinary conceptual framework, and producing outputs relevant to debates on AI, ethics and global health communication.

Progrès au-delà de l’état des connaissances et impact potentiel prévu (y compris l’impact socio-économique et les conséquences sociétales plus larges du projet jusqu’à présent)

The project advanced the state of the art in four main ways.

First, it moved beyond a narrow understanding of AI bias in global health as a technical problem located only in datasets, prompts or algorithms. It showed that AI-generated global health imagery must also be understood in relation to older visual conventions, institutional incentives and histories of colonial and humanitarian representation. This matters because many proposed solutions focus on technical accuracy or dataset diversity, while paying less attention to wider social processes of image-making.

Second, the project theorised global health representation as a longer “text-to-image” paradigm. Long before generative AI, photographers working on global health assignments were already turning text into images through commissioning briefs that specified the scenes, emotions and subjects to be captured. The AI prompt can therefore be understood as a new iteration of an older practice.

Third, the project developed the concept of AI-generated “poverty porn 2.0”. This concept describes the ethical risk of producing synthetic images of suffering, poverty, disease or humanitarian need without direct contact, consent or accountability to the people and communities being represented. It names a specific harm created by the combination of older humanitarian stereotypes and new generative AI tools.

Fourth, the project generated evidence that commercial AI systems, media libraries and stock-image platforms can scale representational harm. During the fellowship, research evidence contributed to the removal of specific dehumanising AI-generated stock assets from a commercial marketplace. The project also identified risks related to AI systems generating humanitarian-style scenes with racialised stereotypes or recognisable organisational branding. These findings show that harmful or misleading AI-generated visual content can become easy to generate, search, buy and circulate.

The project’s potential impacts include stronger ethical guidance for global health communication, better governance of AI-generated humanitarian imagery, and better integration of social science and humanities evidence into AI policy.