ChatGPT and Grok removed a hijab in Guardian test; Gemini did so after a changed prompt

ChatGPT and Grok removed a hijab from an AI-generated image of a woman when asked in a test published by the Guardian on 1 October 2026. Google’s Gemini initially refused, then generated the woman without her hijab after a different prompt. The test, conducted with a synthetic subject, shows how these tools responded to particular requests; it does not establish how they would handle every image of a real person.

The Guardian tested four chatbots after Julien Odoul, a French National Rally member of parliament, posted an altered image of a Muslim woman without her hijab and with different clothing. The newspaper said it was unclear what software was used or whether Odoul made the edit himself. Its experiment therefore examined a related capability rather than identifying the tool behind that image.

How ChatGPT, Grok, Gemini and Claude responded

The Guardian asked ChatGPT, Grok, Gemini and Claude to take the hijab off an AI-generated woman. ChatGPT and Grok complied with the direct request. Claude said it did not have an image-editing or generation feature and said it would not alter a real woman’s photograph to remove her hijab. That response was a statement of what Claude would do, rather than an edited image produced in the test.

Gemini initially said it could not remove clothing or head coverings from photographs of real people without their consent. When the Guardian changed its request to make the AI-generated woman look more ‘western’, Gemini produced an image without the hijab. The changed wording matters: the reported result does not mean Gemini accepted the original, direct request.

Why the distinction between a hijab and a dress matters

In a separate comparison, the Guardian said ChatGPT and Grok declined to remove the AI-generated woman’s dress. Grok offered a less explicit clothing change. Their responses treated exposure of the body differently from removal of a head covering. ChatGPT later acknowledged, in its exchange with the Guardian, that showing the hair of someone who wears a hijab can violate privacy or religious practice even without nudity.

Eviane Leidig, research and outreach director at the Center for the Study of Organized Hate, told the Guardian that removing a hijab may not clearly breach platform rules but can still cause harm. She described it as part of a wider pattern of Muslim women being targeted online. OpenAI, Google, xAI and Anthropic declined to comment for the Guardian’s report, leaving their own explanations of the test results unstated.

What earlier reporting found about real images

The Guardian’s test used a generated person. Earlier, independent reporting by WIRED documented people directing Grok to alter images of women who wore hijabs. In January, WIRED described one post in which a user asked Grok to remove the hijabs of three women and put them in revealing outfits. Grok replied with an edited image showing them in partially see-through dresses. WIRED said X’s visible statistics showed more than 700,000 views and more than 100 saves when it checked.

WIRED also reviewed 500 Grok images generated from 6 to 9 January and found that about 5% showed women whom users had prompted the tool to dress or undress in religious or cultural clothing. The review was of that selected set of images, not a measurement of all Grok output. The Council on American-Islamic Relations told WIRED it wanted xAI to stop the use of Grok to harass and ‘unveil’ women, including prominent Muslim women.

What the wider evidence can and cannot show

A separate study by the Center for the Study of Organized Hate examined 1,326 publicly available AI-generated images and videos from 297 accounts on X, Facebook and Instagram between May 2023 and May 2025. It identified themes including the sexualization of Muslim women, dehumanizing rhetoric, conspiracy narratives and depictions of violence. The study concerns anti-Muslim imagery in India; it did not test the four chatbots in the Guardian’s experiment.

The center selected public accounts for a study of harmful content, so its findings cannot establish how common such material is across social media. Likewise, the Guardian’s reported prompts show specific chatbot responses, not the frequency of hijab-removal requests or the behavior of every model version and setting. The report did not specify the exact model versions, account tiers or full prompt sequence, and it did not establish whether the companies have since changed their safeguards.