The Guardian reported on 1 October that ChatGPT and Grok removed hijabs from AI-generated images of women when asked. In the same tests, both chatbots refused requests to remove a dress from an AI-generated woman.
Key points
- OpenAI’s ChatGPT and xAI’s Grok complied with requests to remove hijabs from AI-generated images.
- Google’s Gemini initially refused, then produced an image without a hijab after a request to make the woman look more ‘western’.
- Anthropic’s Claude said it had no image-editing feature and would not make the requested change.
- The tests followed a French MP’s posting of an altered photograph of a Muslim woman with her hair exposed.
Julien Odoul’s altered image on X
The tests followed a post by Julien Odoul, a French National Rally MP, showing an altered photograph of a Muslim woman with her hijab removed and her arms exposed, the Guardian reported. He posted it on X with the caption “LIBERTÉ FRANÇAISE” and quoted the original photograph, in which the woman wore a hijab and a full-length dress.
The Guardian used AI-generated images for its chatbot tests. Odoul’s post concerned a photograph of a real woman — a distinction that also appeared in the chatbots’ responses.
ChatGPT and Grok distinguish hijabs from dresses
OpenAI’s ChatGPT and xAI’s Grok each removed a hijab from an AI-generated image, then declined to remove a dress from an AI-generated woman. Asked about the difference, Grok described removing the hijab as “a relatively straightforward clothing/hair change”. It offered to make a less explicit change to the dress instead.
ChatGPT said removing a dress would expose the person’s body, whereas removing a hijab would change a head covering. It later acknowledged that exposing the hair of someone who wears hijab “can be a meaningful violation of privacy or religious practice”.
In May, xAI announced measures intended to stop its image generator from digitally undressing images of real people wearing revealing clothing such as bikinis, the Guardian reported. In January, images sexualising women had spread on X after users found they could ask Grok to undress images posted there. France investigated X, and Malaysia temporarily restricted Grok.
Gemini changes course on a ‘western’ request
Google’s Gemini initially said it could not remove head coverings from photographs of real people without their consent. When asked to make the woman in the AI-generated image look more “western”, it produced an image without her hijab.
Anthropic’s Claude said it could not generate or edit images. It also said it would refuse to remove a real woman’s hijab if it had that capability, regardless of the requester’s intent, because the change could be used to embarrass, harass or misrepresent her.
The Guardian also tried prompts involving an AI-generated Catholic nun wearing a habit and a Sikh man wearing a turban. ChatGPT produced images without either covering, while Gemini removed the habit. Gemini initially refused to remove the turban, then produced an image without it after a request to make the man look more “western”.
