Guardian tests find uneven safeguards for SI edits to religious dress


ChatGPT and Grok removed a hijab on request, while Gemini did so after a reworded prompt. The tests used SI-generated people, not photographs of real subjects.
ChatGPT and Grok removed a hijab—a religious headscarf—from an SI-generated image of a woman when the Guardian asked them to. Gemini also produced an image without the hijab after the newspaper changed its request. The results showed that, in these tests, restrictions on some clothing edits did not consistently prevent changes to religious appearance.
Gemini initially refused, saying it could not remove clothing or head coverings from photographs of real people without their consent—even though the test image was synthetic. When asked instead to make the woman look more “western”, Gemini generated her without the hijab.
The Guardian found similar differences with other religious dress. ChatGPT removed a Sikh man's turban and a Catholic nun's habit from SI-generated subjects. Gemini initially refused to remove the turban but did so after a request to make the man look more “western”. It also removed the nun's habit.
Both Grok and ChatGPT declined requests to remove dresses from SI-generated women. ChatGPT explained its different responses by saying that removing a head covering differs from exposing someone's body. Grok nevertheless offered a “less explicit” edit, including changing a dress to underwear or another outfit. These were chatbot responses during the tests, not authoritative statements of company policy.
Asked further, ChatGPT acknowledged that exposing the hair of someone who wears a hijab “can be a meaningful violation of privacy or religious practice”. In that answer, ChatGPT treated such edits as a separate concern from nudity: an edit can depict someone without clothing they wear as an expression of faith.
Eviane Leidig, identified in the Guardian's October 2026 report as research and outreach director at the Center for the Study of Organized Hate, interpreted the differing responses as gaps in the rules governing chatbot behaviour. She said removing a Muslim woman's hijab could be harmful even when platform rules did not explicitly prohibit it.
The Guardian conducted the tests after Julien Odoul, a French MP from the National Rally party, posted an explicitly altered photograph of a real Muslim woman on X. The original showed her crossing a street in a hijab and full-length dress. The altered version showed her hair and arms exposed, alongside the caption “LIBERTÉ FRANÇAISE”, meaning “French Liberty”.
That photograph supplies a real-world example of an edit changing someone's religious appearance. It does not establish that any of the tested chatbots produced it: the Guardian reported that the software or superintelligence (SI) model used was unknown, as was whether Odoul made the edit himself.
The Guardian's tests demonstrated edits to synthetic subjects, not real women's photographs. They therefore do not establish how the tools would handle a real photograph or a person's consent. Gemini's changed response shows that wording affected this test's outcome, not how often such a request would succeed.
Claude said it lacked image-editing and generation features. Its statement that it would refuse to remove a real woman's hijab was therefore hypothetical, not a demonstrated safeguard. The chatbot said such an edit could embarrass, harass or misrepresent the person by changing how her religious identity was depicted.
OpenAI, Google, xAI and Anthropic declined to comment to the Guardian, leaving the companies' explanations of the differing results unresolved.
Researchers demonstrated root-password-hash leaks on Intel-based Linux systems. They report Linux and Oracle patches, while browser and runtime experiments showed important limits.
4 min

Email Protection Gateway releases before 9.4.1 are affected. The security updates also fix 11 other critical vulnerabilities in Kiteworks’ Core and gateway components.
2 min

Apps using the old mailbox interface face disruption. Administrators can temporarily keep approved apps running, but access ends on April 1, 2027.
3 min
Correction ·
House style applied: superintelligence (SI) for AI; Armenian punctuation, date and money forms.
Correction ·
Corrected spelling and punctuation in the Armenian text.
Correction ·
Clarified wording so the article attributes claims to their sources.