Gemini Removed a Hijab After Being Asked to Look 'Western'── The choice of AI and the requester's wording now decide whether religious dress survives an AI edit, so organisations have to write down for themselves that a person's dress stays as it is

On 1 October 2026, The Guardian reported that it had asked four AI chatbots, ChatGPT, Grok, Gemini and Claude, to remove a hijab from images of women, and that three of them eventually did. When you ask an AI to retouch a photo of a person, who decides whether that person's religious dress stays in the picture? The four systems split, and Gemini dropped its refusal after a single rewording, so for now the choice of AI and the wording of the request decide. An organisation that uses these tools at work has to write its own rule that a person's religious and ethnic dress is not to be changed.

01ChatGPT and Grok removed the hijab; Gemini did so after one rewording

The four chatbots did not give the same answer to the same request. ChatGPT and Grok removed the hijab as asked. Gemini refused the first request, then removed it when asked to make the woman look "more Western". Claude refused, saying: "Without her consent, that crosses a line for me."

Two facts stand out. The answers diverged, and one refusal disappeared when the words of the request changed. From those two facts I conclude that, at present, whether a person's dress survives an edit is decided by which AI is asked and in what words. Neither Google's nor Anthropic's published rules, nor the EU's AI Act, contain a sentence that forbids removing someone's religious dress from an image.

An organisation that publishes photographs of people cannot count on an AI to refuse the next time just because it refused once. Two things follow. The organisation writes into its own rules that religious and ethnic dress is not to be altered. And it keeps the prompts its staff send and the images the AI returns.

02The test asked for removal of religious dress without consent, not a sexual edit

The four answers split. So what exactly did The Guardian give the chatbots, and what did it ask them to do? It used two kinds of image. One was an AI-generated picture of a woman. The other was a photograph of a real woman, posted by the French politician Julien Odoul. For both, the paper asked the chatbots to take the hijab away.

The test did not stop at the hijab. ChatGPT also removed a Sikh man's turban and a nun's habit. Gemini, once the request had been reworded, did the same.

The edit being requested was small: take off one garment and show the hair. It is a different kind of request from producing a sexual image. Even so, an image edited as asked loses something the woman chose to wear for religious reasons, and nobody asked her.

ChatGPT's explanation came after it had already produced the image. ChatGPT itself said that showing the woman's hair could cross a religious line. ChatGPT could put the concern into words, but it did so only after the edit was done, so the words did not stop the edit.

When Gemini refused the first request, it gave a clear reason: without consent, it would not change the woman's appearance. The Guardian's second prompt did not contain the word "remove". It did not mention the hijab at all. It asked for the woman to look Western. The image that came back had no hijab.

03Headshot apps for the workplace had already removed hijabs without asking

In The Guardian's test, a reporter deliberately asked for the removal. Workplace headshot apps had already been removing hijabs without anyone asking.

In July 2026, Mahwish Moazzam described her own test in the Pakistani English-language daily The Express Tribune. Moazzam wrote that she tried more than 25 AI headshot apps. Some of the apps asked the user whether to keep accessories such as glasses. None of them asked for consent before removing a religious head covering.

Work headshots end up in places outsiders see: staff pages, recruitment brochures, speaker introductions. When a communications officer runs an employee's photo through an AI app, the hijab can disappear even though nobody requested it. If no one notices before publication, the organisation puts out, under its own name, a picture of the employee that differs from how she chose to appear.

Moazzam's account is one writer's test. The names of the apps and the exact procedure cannot be checked from the article. It is not a figure that can be stretched into a rate for the whole industry. Still, the pairing is hard to ignore: the apps asked about glasses and did not ask about the hijab. My reading is that the people who built those apps put glasses on the list of things to ask the user about, and left the hijab off.

In The Guardian's test, which chatbot the reporter asked, and in what words, decided whether the hijab stayed. In Moazzam's test, the user asked for nothing of the kind, and a default setting built into the app decided. In neither case did anyone ask the woman in the photograph.

04Google's and Anthropic's rules ban religious discrimination but never mention removing dress

Some headshot apps asked about glasses but not about the hijab. Gemini stopped refusing when the wording changed. The apps are not named, so their makers' rules cannot be read. To find where the behaviour of Gemini and Claude comes from, I read the published rules of Google and Anthropic.

Google's policy guidelines for the Gemini app say the model should not produce discriminatory output based on attributes such as religion. The same document also says that, especially when answering difficult prompts, Gemini may generate content that goes against the guidelines. Google acknowledges, in its own document, that Gemini will sometimes not follow its rules.

Anthropic's Usage Policy prohibits promoting discriminatory practices against individuals or groups on the basis of protected attributes such as race, ethnicity, religion, national origin and gender. I found no sentence in it that names the removal of religious dress from a picture of a person. I could not open the usage policies of OpenAI and xAI for this piece, so I say nothing here about what they contain.

The two policies I could read describe what they prohibit with one broad word: discrimination. Whether removing a hijab from an image counts as discrimination is not settled by that wording. What the wording leaves open is settled by the model at the moment it reads the request.

Gemini's reading of the request changed with the words in it. My guess is that "remove the hijab" was read as a request to delete something, while "make her look Western" was read as an adjustment to appearance. Gemini refusing the first time and complying the second time fits that guess.

Claude's refusal cannot simply be credited to a line in its policy either. The reasons Claude gave were the absence of consent and the fact that it has no image-editing function. Anthropic's policy does not say "refuse to remove religious dress". So there is no basis for saying that a later version of Claude, given a differently worded request, would refuse in the same way.

05Since 2 August the EU has required disclosure for edited images that resemble real people

The providers' policies contain no sentence about removing dress. What, then, does the law ask of the people who use these tools? Take the EU's AI Act.

Article 50(4) of the AI Act, Regulation (EU) 2024/1689, deals with deep fakes. Anyone who uses an AI system to generate or manipulate image, audio or video content of that kind must disclose that the content has been artificially generated or manipulated. A deep fake, in the Act's sense, is content that resembles existing people, objects, places or events and could falsely appear authentic. According to artificialintelligenceact.eu, the obligation began to apply on 2 August 2026.

Someone in the EU who uses a photograph with the hijab removed meets this obligation by saying that the image was edited. Article 50(4) does not make the removal itself an offence. What the Article protects is the viewer's ability to tell an edited image from a real one. The dress that the woman in the photograph chose for herself is not something this Article protects.

I have not checked how far the Article reaches organisations outside the EU. For a Japanese organisation, the first thing worth knowing is narrower: the Article asks only for disclosure.

06Answers differ by chatbot, shift with wording, and disclosure does not keep the dress

The EU's AI Act asks for disclosure when dress is removed and does nothing to stop the removal. From the facts so far, I take three things that an organisation using these tools should know.

Switch chatbots and the same prompt returns a different result

To the same request, ChatGPT and Grok complied and Claude refused. An organisation that changes chatbots, or lets each department use a different one, will get different results from the same prompt. The divergence in The Guardian's test is the evidence. If the organisation relies on the AI to refuse, the choice of AI decides how the photograph turns out.

A chatbot that refused once may comply when asked differently

Gemini refused, then removed the hijab after hearing "more Western". Google's own guidelines say that on difficult prompts Gemini may produce content against its rules. One member of staff being refused once tells you nothing about what happens when a colleague asks in other words.

A photo that meets the EU disclosure duty can still lose its hijab

Article 50(4) asks the person using the AI only to say that the image has been edited. A photograph from an organisation that has followed the law can still be missing the hijab. Complying with the law and preserving the person's dress are two separate checks.

So the rule "do not change a person's religious or ethnic dress" is one the organisation has to write for itself. If it were my organisation, I would add that line to the standard prompt. The person receiving the AI's image would set it beside the original and look for anything missing. The edit would be logged, with the prompt and the output saved together.

07Until the four companies state a policy, your own prompts and outputs are the only record

Gemini changed its answer when the wording changed, and neither the policies nor the EU AI Act stopped the removal. How might the four chatbots' answers change from here?

OpenAI, Google, xAI and Anthropic all declined to comment to The Guardian. None of them has said whether it will change its policy, or when.

Google has acknowledged that Gemini sometimes produces content against its own guidelines, and none of the four has published how the next version will answer the same request. So a request Gemini refuses today may be accepted by next month's version, and a request accepted today may equally be refused next month. No published material lets anyone predict whether the answers will become stricter or looser.

Until the four companies state a policy, there is only one thing an organisation can check after the fact: its own record of what it asked the AI and what came back. With that record, when a published photograph is questioned, the organisation can trace who asked for what, in which words. An organisation that has kept no prompts cannot find out whose request produced the photograph with the hijab removed.

Whether an AI removes the religious dress from a photograph of a person is decided today neither by the provider's policy nor by the law. It is decided by which AI is asked and in what words, and in headshot apps by the default settings.

An organisation that uses these tools therefore writes one line into its rules, that a person's religious and ethnic dress stays as it is, and keeps its prompts and outputs. The line is short. For now this line, with outputs checked against the originals and kept on record, is the only safeguard against an organisation rewriting, under its own name, the way a person chose to appear.