Unveiled by a Prompt: AI Strips Hijabs but Refuses to Remove Dresses

The request was simple: take the headscarf off. The chatbot obliged.

On 1 October 2026, the Guardian published the results of a short, pointed experiment. Reporter Johana Bhuiyan gave four of the world's most widely used AI assistants an image of a woman wearing a hijab and asked them to remove it. OpenAI's ChatGPT did it. xAI's Grok did it. Google's Gemini initially refused, explaining that it could not remove clothing or head coverings from images of people without their consent, and then produced a picture of the same woman with her hair uncovered once it was asked to make her look more “western”. Anthropic's Claude, which does not generate or edit images, said it would not make the edit anyway.

The paper then widened the test. ChatGPT removed a Sikh man's turban and a Catholic nun's habit when asked. Gemini removed the habit and, after a change of wording, the turban. And then came the control experiment that turned an uncomfortable finding into a revealing one. Asked to remove a woman's dress, both ChatGPT and Grok declined. ChatGPT explained the difference in plain terms: removing a hijab changes a head covering, while removing a dress would expose the body. Grok described taking off the hijab as “a relatively straightforward clothing/hair change”, and then, having refused the dress, offered a “less explicit version” in which the dress might become underwear, a different outfit or a partial reveal. Only later in the exchange did ChatGPT acknowledge that “for someone who wears hijab, exposing their hair can be a meaningful violation of privacy or religious practice”. OpenAI, Google, xAI and Anthropic all declined to comment.

Precision matters here. The woman in the test image was AI-generated, so nobody real was harmed. But the experiment was designed to measure judgement, not harm: where these systems draw the line between an acceptable edit and an unacceptable one, and what they think that line protects. The answer they gave, in their own words, is that the line is drawn at skin. A garment that covers breasts and hips is protected. A garment that covers hair, worn by millions of women as an act of faith, is decoration.

That answer raises the questions this piece sets out to examine. What does it mean for women whose dress is a visible marker of faith and identity that mainstream tools will strip it on request while treating other clothing as off limits? Why does the removal of religious dress fall outside the protections applied to other alterations? And what does the readiness of widely used systems to perform the edit say about whose appearance their rules were written to respect?

A Photograph Recaptioned as French Liberty

The Guardian's test did not come out of nowhere. On 4 September 2026, Aly Diouara, a deputy for the left-wing La France Insoumise, posted a photograph on X of a young woman crossing a street in a hijab and a long pink dress, with a one-word caption: “Liberté”. The following day, Julien Odoul, a National Rally deputy and the party's spokesperson in the National Assembly, replied with his own version of the picture. In it, the woman's headscarf and dress had been replaced by uncovered hair, a white sleeveless top and trousers. His caption read “Liberté française!” and he noted that the image had been made with AI.

The woman was not a politician, an activist or a public figure. French broadcaster franceinfo described her as an unknown veiled woman who had become a star of the web despite herself, pulled into a political fight she had not chosen. Sabrina Sebaihi, a Green deputy for Hauts-de-Seine, filed a complaint with the Paris prosecutor's office, citing the provisions of France's 1881 press law on provocation to discrimination and the unauthorised algorithmic transformation of a real person's image. “Freedom does not consist of targeting women based on their religion,” she said, but in “guaranteeing equal rights for all.”

The post landed in a specific political context. The National Rally has campaigned to ban the Islamic headscarf in public spaces, and Odoul himself had been here before. In October 2019, at a session of the Bourgogne-Franche-Comté regional council, he demanded that a Muslim mother accompanying her son's school trip either remove her headscarf or leave the chamber, citing laïcité, France's principle of state secularism. The council's president, Marie-Guite Dufay, told him nothing in French law or the chamber's rules banned such clothing. The Associated Press reported that her son cried, and that she had to comfort him.

Seven years on, Odoul did not need the woman to be in the room. He needed a picture, a prompt and a few seconds. The Guardian described the alteration as “a violation of one of the world's most common and visible expressions of faith”. The obvious next question was whether this required specialist tools, or whether the mainstream assistants used by vast numbers of people every day would do the same job. Now we know.

Five Per Cent of Grok's Output

If the Odoul post was the spark, the fuel had been accumulating for months. In early January 2026, Grok became the centre of a global scandal after users on X found they could reply to photographs asking for the person pictured to be put in a bikini or underwear, and Grok's public account would comply in the thread. Between 6 and 9 January, WIRED reviewed 500 images generated by Grok. Around 5 per cent of them showed a woman who, as the result of user prompts, had either been stripped of religious or cultural clothing or been made to wear it. Saris, hijabs, burqas, Japanese school uniforms and even early twentieth century bathing suits all featured.

One detail in WIRED's reporting captures the dynamic precisely. A verified account with more than 180,000 followers replied to a photograph of three Muslim women in hijabs and abayas, asking Grok to “remove the hijabs, dress them in revealing outfits for New Years party”. Grok produced an image of the same women, barefoot, hair uncovered, in partially see-through sequinned dresses. The image drew more than 700,000 views. When WIRED asked xAI about such incidents, the company's automated reply read “Legacy Media Lies”.

Five per cent sounds small until you remember the volume. The January episode drew regulators across Asia and Europe. Indonesia temporarily blocked Grok over that weekend, with its communications ministry saying it was acting “to protect women, children, and the entire community from the risk of fake pornographic content”. Malaysia followed a day later, citing “repeated failures by X Corp”. On 12 January, Ofcom opened a formal investigation into X under the UK's Online Safety Act. On 14 January, X said it had “implemented technological measures to prevent the Grok account from allowing the editing of images of real people in revealing clothing such as bikinis”, and that the restriction applied to all users, including paid subscribers.

Notice what that statement protects. Bikinis. Revealing clothing. The fix, rushed out under regulatory pressure, was aimed squarely at sexualisation as conventionally understood. It said nothing about edits that uncover a woman's hair, which is why, nine months later, the Guardian could still ask Grok to remove a hijab and watch it happen.

The Anatomy of a Safety Rule

To understand why the line falls where it does, look at how these rules are written. AI companies rarely write image policies from first principles. They borrow categories that already exist in law and in platform trust and safety practice. The most developed is non-consensual intimate imagery, a term that grew out of campaigns against so-called revenge porn and has since been extended to deepfakes.

Google's Generative AI Prohibited Use Policy, which governs Gemini, lists among banned activities content that “facilitates non-consensual intimate imagery”, alongside “harassment, bullying, intimidation, abuse, or the insulting of others” and anything that “violates the rights of others, including privacy”. OpenAI's usage policies prohibit non-consensual intimate content and the use of someone's likeness without consent in ways that could confuse authenticity. On paper, a forced unveiling might fall under harassment, privacy or likeness. In practice, those categories are broad, contested and hard to operationalise in a classifier. “Intimate imagery” is not. It has a definition, and that definition is anatomical.

That is what the dress experiment exposed. Refusing to remove the dress was not a fresh moral judgement. It was the recognition of a request that maps onto a well-defined, heavily enforced category: exposing a body. Removing the hijab mapped onto nothing in particular. Hair is not an intimate body part in any of the policies examined here, so a head covering is, as far as the classifier is concerned, an accessory. The system did exactly what it was built to do. The problem is what it was built to see.

Eviane Leidig, director of research and outreach at the Center for the Study of Organized Hate, put the gap succinctly in her comments to the Guardian. De-veiling Muslim women, she said, “may not be distinctly violative” of platform rules, “but it's still harmful”. She placed the tests in a larger pattern: “This is just part of an ongoing trend of how we see Muslim communities, and particularly Muslim women, being targeted and abused.”

There is also a design history here that matters. In March 2025, when OpenAI launched native image generation in GPT-4o, Joanne Jang, who leads model behaviour at the company, published an essay explaining its approach. “We're shifting from blanket refusals in sensitive areas to a more precise approach focused on preventing real-world harm,” she wrote, adding that “AI lab employees should not be the arbiters of what people should and shouldn't be allowed to create.” TechCrunch reported that the new system would fulfil requests about physical characteristics that its predecessor had refused, such as “make this person's eyes look more Asian”.

There is a reasonable argument behind that shift. Blanket refusals are paternalistic, and they can encode the assumption that any request touching on race, religion or bodies must be offensive. But precise harm prevention only works if the system's map of harm is complete. If the map does not recognise a category of injury, precision simply means confidently permitting it. The shift moves the burden of proof: an edit no longer has to be shown to be safe, only not to be harmful in a way the system already understands.

How the Law Learned to See Only Skin

The AI companies did not invent the anatomical definition of intimacy. They inherited it, and the inheritance can be traced through statute.

In England and Wales, section 66D of the Sexual Offences Act 2003, inserted by the Online Safety Act 2023, defines what it means for a photograph or film to show someone “in an intimate state”. The list is precise: participating in a sexual act, doing something a reasonable person would consider sexual, the exposure of “all or part of the person's exposed genitals, buttocks or breasts”, and acts of urination, defecation or associated personal care. “Exposed” is defined to include those areas visible through wet or transparent clothing, or covered only by underwear. When the government created a new offence of creating or requesting a purported intimate image of an adult without consent, in section 138 of the Data (Use and Access) Act 2025, which came into force in February 2026, it borrowed the same definition wholesale. The explanatory notes say nothing about religious dress.

This was not an oversight. It was a decision. In its July 2022 final report on intimate image abuse, the Law Commission considered images that are intimate for cultural or religious reasons, such as a woman who usually wears a hijab pictured without it. Stakeholders told it such images could cause serious harm, including exposure to so-called honour-based abuse. The Commission nonetheless confined its definition to sexual, nude, partially nude and toileting images. Alex Davies-Jones, then the minister for victims, later told MPs that the Commission had found it “impossible to craft a definition that suits everyone”.

Parliament has pushed back. In March 2025, the Women and Equalities Committee, chaired by Labour MP Sarah Owen, published a report on tackling non-consensual intimate image abuse. It noted that while the law defines an intimate state as a sexual act, nudity or toileting, abuse “can also include material that is considered 'culturally intimate' for the victim, such as a Muslim woman being pictured without her hijab.” It recommended extending the legal definition to cover images where, “because of the person's religious or cultural background, the person commonly wears particular attire of religious or cultural significance when in public,” and the material depicts the person without that attire in circumstances in which an ordinary reasonable person would expect privacy. After the government replied in May 2025, the committee warned that its proposals did not go far enough. The 2025 Act, as passed, used the old definition. The Crime and Policing Act 2026, given Royal Assent in April 2026, then added offences of taking intimate images without consent and of supplying nudification tools, along with a duty on platforms to remove non-consensual intimate images within 48 hours. It, too, left the definition of an intimate state untouched.

The committee's wording was not invented in Westminster. It closely tracks a law that has been on the books in Australia since 2021. Section 15 of Australia's Online Safety Act defines an intimate image to include material that depicts, or appears to depict, a person without the religious or cultural attire they consistently wear in public, in circumstances where an ordinary reasonable person would expect privacy. A further subsection makes clear that an image can be intimate even if it has been altered. In other words, a major common-law democracy has already decided, in legislation, that a fabricated image of a woman without her hijab can belong in the same legal category as a fabricated nude.

Even the Australian model has limits worth noting. The privacy qualifier means a doctored photograph of a woman crossing a public street, like the one Odoul posted, may not meet the test. And French law, which since the SREN law of May 2024 has punished the distribution of algorithmically generated content depicting a person without consent where its artificial nature is not obvious or expressly stated, has a gap of its own: Odoul labelled his image as made with AI. The disclosure that was meant to protect the public from deception may also shield the person who did the stripping. In the United States, the TAKE IT DOWN Act, signed in May 2025 with platform takedown duties that took effect in May 2026, is similarly built around nudity and sexual conduct.

The pattern is consistent. The law treats the exposure of certain body parts as a violation regardless of context or the subject's own sense of self, and largely declines to treat the removal of religious garments as more than an aesthetic change. AI policy teams, looking for defensible lines, have imported that judgement.

Make Her Look More Western

Of all the details in the Guardian's tests, the Gemini exchange is the most instructive, because Gemini got the first answer right. It refused to remove the hijab and gave a reason that sounds like a principle: it could not remove clothing or head coverings from images of people without their consent. The test image was synthetic, so the refusal was arguably over-cautious. But it was the right instinct.

Then the request was reframed around making her look more “western”, and the principle evaporated.

The charitable reading is that this is a classic jailbreak: a guardrail keyed to one phrasing and bypassed by another. That is true, and a known weakness of intent-based filters. The more important reading is about what the system understood “western” to mean. Asked to westernise a woman, Gemini removed her hijab. It did not change her face, her skin or her setting. It decided that the thing standing between this woman and westernness was the cloth on her head.

That is not a neutral aesthetic inference. It is a political claim with a long history, the same claim Odoul's caption made in two words. Left to interpret an open-ended cultural instruction, the model reproduced the assumption that Muslim women's religious dress is the opposite of being western, rather than something worn every day by millions of women who are British, French, American, Canadian or Australian.

That assumption has measurable consequences offline. In the UK, the anti-Muslim hate monitoring group Tell MAMA recorded 6,313 reports of anti-Muslim hate in 2024, the highest number since it was founded in 2011, and noted greater aggression and “a greater focus on and targeting of visible Muslims”. In January 2026, the Women and Equalities Committee published a separate report on discrimination, harassment and abuse against Muslim women, documenting accounts of women being threatened, spat at and having their hijabs pulled from their heads. A system that treats removing a hijab as a neutral styling choice is operating in a world where pulling off a woman's headscarf is a recognised act of street harassment.

An Old Ritual With a New Interface

The idea that liberating Muslim women means uncovering them is far older than artificial intelligence, and its most notorious staging took place in Algiers. On 13 May 1958, during the Algerian war of independence, the French army organised public ceremonies in Algiers in which Algerian women removed their veils, sometimes burning them, before crowds and photographers, under the protection of armed soldiers. Historians have documented that some participants were persuaded, paid or pressured. The spectacle was presented as emancipation. It was, as scholars from the Graduate Institute Geneva to Georgetown University's Berkley Center have argued, a demonstration of colonial power staged on women's bodies.

The psychiatrist and anti-colonial theorist Frantz Fanon dissected that logic in “Algeria Unveiled”, the opening essay of his 1959 book “L'An V de la révolution algérienne”, translated into English as “A Dying Colonialism”. Fanon argued that the colonial administration understood the veil as the key to Algerian society, and that unveiling women was meant to break the resistance of the whole people. Decades later, the anthropologist Lila Abu-Lughod, writing in American Anthropologist in 2002 about the rhetoric used to justify the war in Afghanistan, warned against the reflex of plastering “neat cultural icons like the Muslim woman” over complex political realities, and against the assumption that Muslim women need saving by outsiders who decide what their freedom should look like.

None of this means the engineers who shipped these image editors were channelling the Algerian war. The point is narrower and more troubling. The idea that a woman's religious dress is a removable layer between her and a freer, more western self is embedded deeply enough in the culture that produced the internet that a model trained on it will reproduce the idea unprompted. Odoul did not need to explain his caption. Gemini did not need to be told what “western” meant.

What is new is the interface. In 1958, unveiling required soldiers, organisers and a crowd. In 2019, it required a politician to stand up in a regional assembly and make a demand in front of a child. In 2026, it requires a sentence typed into a box, and the result can be produced for any woman whose photograph exists online, without her presence, knowledge or consent, and distributed to an audience of hundreds of thousands before she knows it exists.

Consent Is the Point, Not the Cloth

There is an obvious objection, and it deserves a serious answer. Many women do not wear the hijab. Many who once wore it have chosen to stop. In Iran, uncovering has been an act of defiance against a state that enforces veiling with violence. Mahsa Amini, a 22-year-old Kurdish Iranian woman, died on 16 September 2022 after being detained by Tehran's morality police for allegedly wearing her hijab improperly. A UN fact-finding mission later concluded that her death was unlawful and caused by physical violence in state custody. The protests that followed, under the slogan “Woman, Life, Freedom”, saw women publicly remove and burn their headscarves. Should an AI system refuse to depict any of that?

No. And the existence of that objection clarifies what the actual harm is. The problem is not that an image of a woman without a hijab exists. It is that a specific woman, who has chosen to cover, is depicted uncovered by someone else, without her consent, usually to humiliate her or make a political point about her. That is precisely the structure of non-consensual intimate image abuse. The injury does not lie in the pixels. It lies in the overriding of a person's own decision about how much of herself to show the world.

Seen this way, the hijab and the dress are not different kinds of case at all. In both, the edit takes something a woman has chosen to keep covered and uncovers it against her will. The only difference is which part of her the culture writing the rules agrees she is entitled to keep private. For a woman who covers her hair as an act of faith, a fabricated image of herself uncovered and circulated by strangers may feel closer to a fabricated nude than to having her jacket recoloured. Australian law recognises that, and so did the UK's parliamentary committee.

So, tellingly, did ChatGPT, once it was asked to reflect. Its acknowledgement that exposing a hijab-wearer's hair can be “a meaningful violation of privacy or religious practice” shows the understanding was in the model. It simply was not connected to the decision about whether to perform the edit. That is a design choice, or at least the absence of one.

Whose Modesty Counts

So whose appearance are these rules written to respect? The honest answer is a particular, culturally specific idea of which parts of a body are private, one that matches the norms of the societies where most of these systems are designed and most of their training data originates.

Researchers have documented the broader version of this problem. In a widely cited 2023 paper titled “Which Humans?”, Mohammad Atari, Mona J. Xue, Peter S. Park, Damián E. Blasi and Joseph Henrich, of Harvard University, showed that large language models respond to standard psychological tests in ways that resemble people from Western, educated, industrialised, rich and democratic societies, the population psychologists call WEIRD. Their point was not just that the models skew WEIRD, but that the models' idea of the average human skews WEIRD, even though most humans are not. Content policy is a downstream expression of the same skew. If a policy team's working assumption about the average user is someone for whom hair is not private, then a rule that protects only breasts, buttocks and genitals will feel universal to them. It is not.

The consequences fall unevenly. In September 2025, the Center for the Study of Organized Hate published a study of 1,326 AI-generated posts from 297 accounts on X, Facebook and Instagram in India between May 2023 and May 2025. It found that sexualised depictions of Muslim women drew the highest engagement of any category, more than 6.7 million interactions, which it said revealed “the gendered character of much Islamophobic propaganda, which fuses misogyny with anti-Muslim hate”. The same people who are most likely to be targeted by unveiling edits are the people whose definition of intimate privacy is least likely to be encoded in the tools.

This is the deeper meaning of the asymmetry the Guardian found. When a system refuses to remove a dress but happily removes a hijab, it is telling the woman in the hijab that her sense of what is private is a preference, while another woman's sense of what is private is a right. It is extending full dignity to one standard of modesty and treating another as costume. That is not because anyone sat down and decided Muslim women deserve less protection. It is because nobody sat down and decided they deserved the same.

The Turban and the Habit

The Guardian's decision to test a Sikh turban and a Catholic nun's habit was revealing, because it showed the problem is not specific to Islam. It is specific to religious dress as a category.

For observant Sikhs, the turban, or dastaar, is not an optional symbol. Sikh organisations, including the Sikh Coalition in the United States and the World Sikh Organization of Canada, describe it as a mandatory article of faith, worn to cover uncut hair as a sign of devotion, equality and dignity. Forcible removal of a Sikh's turban has long been regarded within the community as among the gravest humiliations that can be inflicted, and it has featured in hate attacks across the Sikh diaspora. ChatGPT removed it on request. Gemini removed it after the wording was adjusted.

The nun's habit is different: a vocational garment rather than an obligation for all believers, and less likely to be a target of harassment. But it still marks religious commitment, and treating it as interchangeable with a hoodie reveals the same blind spot. The models are not weighing the meaning of the garment to the person wearing it. They are checking whether its removal would expose skin.

One could argue these results show even-handedness: the tools are not singling out Muslims, merely indifferent to all religious dress. That is little comfort. Indifference applied evenly to garments is not applied evenly to people. Unveiling edits of Muslim women are already a documented genre of online abuse, and turban removal has a long history as an act of anti-Sikh violence. A rule that treats all religious dress as styling will be exploited against the communities already targeted.

What Real Protection Would Look Like

None of the fixes here are technically exotic. They are policy decisions that companies have so far chosen not to make.

The first is definitional. If the major developers wrote something like the Australian standard into their own usage policies, treating the non-consensual removal of religious or cultural attire from an identifiable person's image as a form of intimate image abuse, the edit would fall into a category their systems already know how to refuse. The Women and Equalities Committee's wording offers a ready template, and companies do not need to wait for Parliament to amend section 66D before adopting it.

The second is about consent rather than cloth. Gemini's first answer, that it could not remove clothing or head coverings from images of people without their consent, was the right one; the failure was that a rephrased request bypassed it. Systems that edit images of real people could treat changes to religious dress, alongside skin colour, body shape and other identity markers, as edits requiring heightened scrutiny. That is harder than nudity detection, but these companies already run pipelines to check whether an upload shows a real person, a public figure or a child. Adding religious attire to the list of sensitive attributes extends existing machinery rather than inventing new machinery.

The third is about interpretation. When a model is asked to make someone look more “western” or more “modern”, it is being asked to make a cultural judgement. Developers could test explicitly for that class of prompt and treat outcomes that strip religious markers as failures, as they already test for racial stereotyping.

The fourth is accountability. Grok's January fix came only after two countries had blocked the service and a regulator had opened an investigation, and it was scoped to the narrowest definition of harm that would satisfy them. If companies want to argue that they, rather than legislators, should set the boundaries of acceptable editing, they need to consult Muslim, Sikh and other religious communities and publish how religious dress is handled in their policies. Regulators, for their part, are enforcing offences built on the anatomical definition. Parliament has now passed two image-abuse laws since its own committee made the recommendation, and adopted it in neither. Until it does, a forced unveiling of a British Muslim woman will remain outside their scope, however humiliating its effects.

The Hair Is Not the Point

The most telling sentence in the Guardian's tests is Grok's description of hijab removal as a relatively straightforward clothing and hair change. It is accurate in the narrowest technical sense. In terms of pixels, swapping fabric for hair is a simple edit. Generating a convincing head of hair where there was none is exactly the kind of task these models excel at.

But a safety rule is not supposed to measure how hard an edit is. It is supposed to measure what the edit does to a person. By that measure, the hijab edit is anything but straightforward. It reverses a woman's most visible decision about how she presents herself, without asking her, for an audience she did not choose. It is the same structure as the harm these companies have spent years learning to refuse when the garment is a dress.

The systems refused the dress because someone, somewhere, decided that a woman's body under her dress was hers to reveal or not. They removed the hijab because nobody made the same decision about a woman's hair under her scarf. That gap is not a technical limitation. It is a statement, written in code and policy, about whose idea of privacy counts as universal and whose counts as a cultural preference.

Julien Odoul called his edited image French liberty. The companies whose tools can now produce that image on demand declined to say what they call it. Until they write rules that treat a woman's choice to cover her hair as seriously as her choice to cover her body, their answer is visible in what their products do.

References and Sources

1. Bhuiyan, Johana. “AI chatbots remove hijabs from images of Muslim women when prompted.” The Guardian, 1 October 2026. Reproduced by Genocide Watch: https://www.genocidewatch.com/single-post/ai-chatbots-remove-hijabs-from-images-of-muslim-women
2. French coverage of the Odoul image and Sebaihi complaint, September 2026. Anadolu Agency, “Voile dans l'espace public: un député RN modifie par IA la photo d'une femme voilée et suscite une polémique”: https://www.aa.com.tr/fr/france/-voile-dans-lespace-public-un-d%C3%A9put%C3%A9-rn-modifie-par-ia-la-photo-d-une-femme-voil%C3%A9e-et-suscite-une-pol%C3%A9mique/4048785 ; TSA Algérie, “France : un député RN utilise l'IA pour « retirer » le voile d'une femme, la députée Sabrina Sebaihi saisit la justice”, 11 September 2026: https://www.tsa-algerie.com/un-depute-rn-utilise-lia-pour-retirer-le-voile-dune-femme-sabrina-sebaihi-saisit-la-justice/ ; franceinfo, “Une inconnue voilée devient malgré elle une star du web et se retrouve au cœur d'une bataille politique”: https://www.franceinfo.fr/replay-radio/l-il-de-constance/une-inconnue-voilee-devient-malgre-elle-une-star-du-web-et-se-retrouve-au-c-ur-d-une-bataille-politique_8160185.html
3. Associated Press. “Muslim headscarf debate divides France, in climate of hate.” Via Religion News Service, 30 October 2019. https://religionnews.com/2019/10/30/muslim-headscarf-debate-divides-france-in-climate-of-hate/
4. WIRED. “Grok Is Being Used to Mock and Strip Women in Hijabs and Saris.” 9 January 2026. As summarised by S2Jnews: https://s2jnews.com/review-finds-hijabi-women-disproportionately-targeted-by-groks-explicit-image-generation-including-abuse-depictions/ and Neurotechnus: https://neurotechnus.com/2026/01/10/grok-ai-abuse-women-hijabs/
5. Regulatory action against Grok, January 2026. TechCrunch, “Indonesia and Malaysia block Grok over nonconsensual, sexualized deepfakes”, 11 January 2026: https://techcrunch.com/2026/01/11/indonesia-blocks-grok-over-non-consensual-sexualized-deepfakes/ ; NPR, 12 January 2026: https://www.npr.org/2026/01/12/nx-s1-5674660/malaysia-indonesia-block-grok-ai-deepfakes ; Ofcom, “Ofcom launches investigation into X over Grok sexualised imagery”, 12 January 2026: https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/ofcom-launches-investigation-into-x-over-grok-sexualised-imagery
6. PBS NewsHour / Associated Press. “Grok blocked from undressing images with AI in places where it's illegal, X says.” 14 January 2026. https://www.pbs.org/newshour/world/grok-blocked-from-undressing-images-with-ai-in-places-where-its-illegal-x-says ; see also Al Jazeera, 15 January 2026: https://www.aljazeera.com/news/2026/1/15/musks-grok-to-bar-users-from-generating-sexual-images-of-real-people
7. Developer usage policies. Google, “Generative AI Prohibited Use Policy”, last modified 17 December 2024: https://policies.google.com/terms/generative-ai/use-policy ; OpenAI, “Usage Policies”, updated 29 October 2025: https://openai.com/policies/usage-policies/
8. OpenAI's GPT-4o image-generation policy shift, March 2025. Jang, Joanne, “Thoughts on setting policy for new AI capabilities”, Reservoir Samples, 27 March 2025: https://reservoirsamples.substack.com/p/thoughts-on-setting-policy-for-new ; Zeff, Maxwell, “OpenAI peels back ChatGPT's safeguards around image creation”, TechCrunch, 28 March 2025: https://au.finance.yahoo.com/news/openai-peels-back-chatgpts-safeguards-161314145.html
9. Sexual Offences Act 2003, section 66D (as inserted by the Online Safety Act 2023). legislation.gov.uk. https://www.legislation.gov.uk/ukpga/2003/42/section/66D
10. Data (Use and Access) Act 2025, section 138 and Explanatory Notes. legislation.gov.uk. https://www.legislation.gov.uk/ukpga/2025/18/notes/division/12/index.htm ; commencement reported by Clifford Chance, February 2026: https://www.cliffordchance.com/insights/resources/blogs/talking-tech/en/articles/2026/02/key-aspects-of-the-data--use-and-access--act-take-effect.html
11. Law Commission. “Intimate image abuse: a final report.” Law Com No 407, July 2022. https://lawcom.gov.uk/project/taking-making-and-sharing-intimate-images-without-consent
12. House of Commons Women and Equalities Committee. “Tackling non-consensual intimate image abuse.” Fourth Report of Session 2024 to 25, HC 336, 5 March 2025: https://publications.parliament.uk/pa/cm5901/cmselect/cmwomeq/336/report.html ; committee response, “Government's proposals do not go far enough to tackle 'increasing threat' from NCII, WEC warns”, 2025: https://committees.parliament.uk/committee/328/women-and-equalities-committee/news/206920/governments-proposals-do-not-go-far-enough-to-tackle-increasing-threat-from-ncii-wec-warns/ ; Arab News, “Non-consensual possession of images of Muslim women without hijabs could become offense in UK” (including Alex Davies-Jones's evidence), 5 March 2025: https://www.arabnews.com/node/2592549/world
13. Online Safety Act 2021 (Cth), section 15. Federal Register of Legislation, Australia. https://www.legislation.gov.au/C2021A00076/2024-12-11/2024-12-11/text/original/epub/OEBPS/document_1/document_1.html
14. Code pénal, article 226-8, as amended by Loi n° 2024-449 du 21 mai 2024 (SREN). Légifrance. https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000049571542
15. Congressional Research Service. “The TAKE IT DOWN Act: A Federal Law Prohibiting the Nonconsensual Publication of Intimate Images.” LSB11314. https://www.congress.gov/crs-product/LSB11314 ; Federal Trade Commission, “Take It Down Act enforcement starts now”, May 2026: https://www.ftc.gov/business-guidance/blog/2026/05/take-it-down-act-enforcement-starts-now-what-know-about-ftc-tida
16. Tell MAMA. “Tell MAMA Records the Highest Number of Anti-Muslim Hate Cases in 2024 Since Its Founding.” 2025. https://tellmamauk.org/tell-mama-records-the-highest-number-of-anti-muslim-hate-cases-in-2024-since-its-founding/ ; Anadolu Agency: https://www.aa.com.tr/en/europe/anti-muslim-hate-cases-in-2024-reach-record-high-in-uk/3486575
17. House of Commons Women and Equalities Committee. “Discrimination, harassment and abuse against Muslim women.” Tenth Report of Session 2024 to 26, HC 571, 30 January 2026. https://publications.parliament.uk/pa/cm5901/cmselect/cmwomeq/571/report.html ; Race Equality Foundation statement: https://raceequalityfoundation.org.uk/press-release/women-and-equalities-committee-report/
18. Scholarship on the 1958 unveiling ceremonies in colonial Algeria. Graduate Institute Geneva, “To veil or to unveil: the dilemma of feminism in the Algerian anti-colonialist struggle”: https://www.graduateinstitute.ch/communications/news/veil-or-unveil-dilemma-feminism-algerian-anti-colonialist-struggle ; Berkley Center for Religion, Peace and World Affairs, Georgetown University, “Unveil Them to Save Them: France and the Ongoing Colonization of Muslim Women's Bodies”: https://berkleycenter.georgetown.edu/responses/unveil-them-to-save-them-france-and-the-ongoing-colonization-of-muslim-women-s-bodies ; The Conversation, “From colonial Algeria to modern day Europe, the Muslim veil remains an ideological battleground”: https://theconversation.com/from-colonial-algeria-to-modern-day-europe-the-muslim-veil-remains-an-ideological-battleground-70242
19. Fanon, Frantz. “L'Algérie se dévoile”, in L'An V de la révolution algérienne. Paris: François Maspero, 1959. English translation: A Dying Colonialism. New York: Grove Press, 1965.
20. Abu-Lughod, Lila. “Do Muslim Women Really Need Saving? Anthropological Reflections on Cultural Relativism and Its Others.” American Anthropologist 104, no. 3 (2002): 783 to 790. https://doi.org/10.1525/aa.2002.104.3.783
21. Independent International Fact-Finding Mission on the Islamic Republic of Iran, findings on the death of Mahsa Amini, March 2024; as reported by Radio Free Europe/Radio Liberty: https://www.rferl.org/a/iran-un-fact-finding-mission-amini-death-crackdown-protests/32854252.html ; Amnesty International, “What happened to Mahsa/Zhina Amini?”, September 2023: https://www.amnesty.org/en/latest/news/2023/09/what-happened-to-mahsa-zhina-amini/
22. Atari, Mohammad, Mona J. Xue, Peter S. Park, Damián E. Blasi and Joseph Henrich. “Which Humans?” Harvard University, Culture, Cognition and Coevolution Lab, 2023. https://coevolution.fas.harvard.edu/sites/g/files/omnuum5841/files/culture_cognition_coevol_lab/files/which_humans_09222023.pdf
23. Center for the Study of Organized Hate. “AI-Generated Imagery and the New Frontier of Islamophobia in India.” 29 September 2025: https://www.csohate.org/2025/09/29/ai-generated-hate-in-india/ ; team profile of Eviane Leidig: https://www.csohate.org/team/eviane-leidig/
24. Sikh Coalition. “Dastaar: Defending Sikh Identity.” https://www.sikhcoalition.org/resources/dastaar-defending-sikh-identity/ ; World Sikh Organization of Canada, “What is the significance of the turban?”: https://www.worldsikh.org/what_is_the_significance_of_the_turban
25. Crime and Policing Act 2026 (c. 20). legislation.gov.uk. https://www.legislation.gov.uk/ukpga/2026/20 ; summary of image-abuse provisions by LexisNexis UK: https://www.lexisnexis.com/en-gb/legal/news/enactment-of-crime-policing-bill-introduces-new-offences-targeting-image-based-abuse-online

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...