AI poverty porn erases the people it claims to protect
NGOs are now using artificial intelligence to illustrate sensitive subjects. But the result reduces reality to clichés – and risks undermining the organisations’ credibility at the same time, writes Antti Yrjönen.
The girl in the picture has a black eye.
She looks straight into the camera. Her face is framed by an unnaturally folded, dull grey headscarf. Yet another nameless victim in an NGO advertisement. One thing, however, is new: the girl in the picture does not exist.
In recent years, many NGOs have quietly begun using AI-generated imagery in their communications. Plan International’s Netherlands office used synthetic images of abused girls in a campaign against child marriage. Amnesty International illustrated a report on police violence in Colombia with AI-generated images of protests. A UN peacekeeping mission produced an awareness video depicting sexual violence in conflict zones with the help of artificial intelligence.
In all three cases, the AI-generated content prompted a public backlash. Plan withdrew the materials and updated its internal guidelines, Amnesty removed the images and apologised, and the UN peacekeeping mission shelved its video.
At the same time, image libraries are filling up with AI-generated slums, refugee camps and malnourished children. Commercial companies are making money by selling generated victims that aid organisations can buy for their advertising in place of real people and real lives.
It is tempting to see AI imagery as a rupture: aid organisations suddenly abandoning documentary photography and moving into synthetic fantasy. In reality, however, it is part of a continuum. Generative AI is not reinventing the visual grammar of humanitarian work. It is reproducing and intensifying the old white-saviour gaze on which it was trained.
Visual culture researcher Arsenii Alenichev recently described AI-generated images of suffering in The Lancet Global Health as a new version of poverty porn. They feature exaggerated scenes of malnourished children, girls in filthy wedding dresses and disaster victims dramatically lit from the side.
Many of the most disturbing images are not the work of isolated individuals. They have been produced for campaigns by respected organisations or distributed through major image libraries such as Adobe and Freepik.
The logic behind this is simple. Photography is expensive, slow and ethically complicated. Someone has to be sent to the location, permissions have to be negotiated and, on top of everything else, the people being photographed have to be encountered as people.
At the same time, the entire aid sector is operating on an increasingly uncertain financial footing. In Finland too, major organisations have recently been searching urgently for savings and conducting restructuring negotiations as public funding tightens and international funding becomes less predictable.
Talk to communications staff at large organisations and a familiar demand keeps surfacing: do more with less. As travel and freelancer budgets are cut, a growing share of content is produced without anyone leaving the office.
Generative AI slips neatly into this world of desk-based humanitarianism. Why wrestle with time zones, language barriers and logistics when a ‘refugee camp’ can be conjured onto a screen in thirty seconds? Isn’t a sufficiently convincing simulation of suffering enough to evoke empathy and raise money?
But empathy built on illusion is fragile. Every time a high-profile organisation is caught illustrating genuine human suffering with fakes, it chips away at public trust. When supporters realise that the heartbreaking child bride was never real at all, but merely the poverty-imagery equivalent of Enkeli-Elisa – someone who could have existed – they may begin to wonder which parts of the story are true in the first place. For communities living amid conflict or poverty, the message is harsher still: your experience is not worth documenting.
NGOs have always used – and still use – people as fundraising mannequins: anonymous ‘beneficiaries’; paid models dressed in traditional clothing; stock photographs bought because the child in them looks precisely vulnerable or grateful enough.
Even when photographs are technically authentic, the stories they tell are controlled by strict guidelines. Photographers are instructed to look for smiles, for something hopeful, to avoid showing too much misery or – depending on the fashion and the target audience – to pursue maximum shock value.
In other words, the Global North decides what other people’s lives elsewhere in the world are supposed to look like. The AI-generated image is the logical end point of that way of thinking. If the aim is not to encounter people but to fit them into a pre-written narrative, why bother meeting them at all?
When Plan International defended its AI campaign to The Guardian, the organisation said it was seeking to protect the privacy and dignity of real girls.
The justification exposes a perverse logic: people are being protected by erasing them.
Dignity has become a virtue to be safeguarded by removing real people from the picture. They are replaced by fictional stand-ins assembled from a vast dataset of past suffering. The suffering body remains at the centre of the image – perhaps more centrally than ever – but no real person is granted the right to be seen as themselves.
Synthetic images feel safe only if we forget how the machine learnt to hallucinate them: by consuming millions of photographs of real children, real pregnant teenagers and real victims of war. Many of those images were taken without genuine consent. Making money from them all over again merely adds another layer of exploitation.
Generative AI does not understand poverty, race or gender. Instead, it reproduces patterns found in its training data – billions of images scraped from the internet, news archives, advertising and social media. That archive is not neutral. It reflects who has held the camera, which images have been selected for publication and whose lives have been considered worth depicting.
Of course, those same choices shaped our view of the world long before algorithms. Roy E. Stryker, who ran the programme that documented the Great Depression in the United States in the 1930s, became notorious for rejecting thousands of negatives by punching holes through them. The images destroyed were those that did not fit the official story – people who looked too prosperous, politically awkward scenes, idleness or resistance.
The photographs we now regard as the visual memory of the Depression – such as Dorothea Lange’s Migrant Mother – are those that survived that selection process. Today’s AI systems have been trained on archives that were already filtered in this way.
In humanitarian imagery, the archive is heavily skewed. There are countless photographs of Black African children as malnourished patients, refugees or orphans; far fewer of them as surgeons or bored teenagers scrolling through their phones. There are endless images of white doctors, volunteers and donors; far fewer of white people receiving aid.
When researchers and journalists have tested generative AI with prompts intended to reverse these roles – asking, for example, for images of Black doctors treating white children – the results have made the biases painfully visible.
In one experiment using the Midjourney AI image generator, researchers tried to create an image of a Black African doctor treating white children. Despite more than 300 attempts, the AI did not produce a single image in which the patients were light-skinned. In some images, white doctors treated Black patients; in others, the doctor’s Africanness was emphasised by adding a giraffe to the scene.
The stereotype of the white saviour and the Black victim is therefore not an occasional error but the most probable outcome.
The problem is not merely that AI copies existing attitudes. When generative systems are repeatedly trained on their own outputs – on synthetic images rather than new photographs – they collapse towards the most common patterns in the data. In a field where the dominant pattern is already a narrow set of racialised clichés, this spiral threatens to make the imagery more one-dimensional still.
The bruise on the face of a girl who does not exist is a story about distance, because synthetic images and videos remove not only the photographer from the equation but the encounter itself. They allow institutions to continue doing what they have long wanted to do: decide from afar how people’s lives are represented and what kind of suffering is useful to show.
At the same time, this is also a calculation of risk that may backfire. Research suggests that audiences are already wary of anything that feels automated or mass-produced. Jasper David Brüns and Martin Meißner found in their research on social-media content creation that brands’ use of AI can trigger negative reactions because it diminishes perceived authenticity. The backlash is stronger when AI is seen as replacing human effort rather than assisting it.
The important question, then, is not whether a good-enough image can be made without a photographer, but what the image is being made for in the first place. If the purpose is to build understanding between people, that cannot be done without encounters. It requires time on the ground, local authorship and a willingness to tolerate the discomfort of complexity.
It begins with looking someone in the eye and giving them room to disagree with the frame you arrived with.
This piece was originally published in Maailman Kuvalehti on 26 January 2026 under the title ‘Tekoälyn tuottama kurjuuskuvasto on köyhyyspornon uusi versio’. Read the original →
