I've watched AI image editing go from novelty to something that can change a photograph with a single sentence.
That convenience creates a harder question than “Does the edit look realistic?”
What happens when the thing being changed is part of someone's identity, faith or privacy?
A new Guardian investigation tested ChatGPT, Grok, Gemini and Claude using images of Muslim women wearing hijabs and found that the systems did not respond in the same way.
ChatGPT and Grok performed the requested head-covering removal in the reported tests. Gemini initially refused, then produced an image without the hijab after a prompt asking for a more “Western” appearance.
Claude did not perform the edit, and Anthropic separately states that Claude does not currently generate images.
The bigger story isn't simply which chatbot said yes or no.
It is how AI systems define harm when an edit does not involve nudity but can still alter a meaningful religious or cultural signal.
Editorial illustration of AI image editing and the question of where privacy and cultural boundaries should apply.
What the Guardian Found
The investigation was prompted by a recent incident in which a French politician posted an altered photograph of a Muslim woman without her hijab.
The Guardian then tested several major AI systems to see how they handled similar requests with AI-generated subjects.
The Reported Responses
- ChatGPT: Performed the hijab-removal edit in the reported test.
- Grok: Also performed the edit.
- Gemini: Initially refused the request but later generated a version without the head covering after a more “Western” framing.
- Claude: Did not perform the edit and said it lacked image-editing capability; Anthropic says Claude is text-based and can analyze uploaded images.
These results should be understood as a journalistic test of model behavior at one point in time, not a permanent description of every version, account or safety setting.
Why a Hijab Edit Is Different From a Normal Clothing Change
A head covering can be ordinary clothing.
It can also be a visible expression of religious practice.
That distinction is important for AI safety because a moderation system focused only on nudity may miss other forms of unwanted representation.
The central issue is not that every hijab edit is inherently harmful.
The issue is whether a system should recognize that changing a culturally or religiously significant feature can change how a person is represented.
The “Western” Prompt Is Especially Interesting
The Gemini portion of the test raises a different technical question.
The system reportedly refused the straightforward request but later produced an altered image after the request was reframed around making the woman appear more “Western.”
That suggests a potential weakness in a rule-based safety approach: changing the wording can change the classification of the same underlying request.
This is a classic challenge for generative AI.
Users do not always describe their intent directly.
They can disguise an action inside a broader request about style, culture, fashion or appearance.
This Is Bigger Than Muslim Women
The Guardian also tested other religious head coverings.
The report says ChatGPT removed a Sikh turban and a Catholic religious habit from AI-generated figures, while Gemini showed similar behavior after broader reframing.
That matters because it suggests the issue is not limited to one particular religion.
| Type of Edit | Potential AI Safety Question | Why It Matters |
|---|---|---|
| Religious clothing | Does the edit alter an important faith marker? | Appearance can communicate identity and religious practice. |
| Cultural clothing | Does the edit erase or stereotype cultural identity? | Context can be lost even without explicit content. |
| Real person's likeness | Was the person allowed to consent to the manipulation? | Can raise privacy, impersonation and reputational concerns. |
| Sexualized transformation | Does the edit create nonconsensual intimate imagery? | Additional safety and legal consequences can apply. |
The same technical capability can therefore have very different consequences depending on what is being changed and who is depicted.
AI Image Editing Has Already Been Used for Harassment
This is not the first warning that image-editing tools can be used against women in religious or cultural clothing.
WIRED reported in January 2026 that Grok was being used to alter images of women wearing hijabs, saris and other religious or cultural clothing.
The Center for the Study of Organized Hate has separately documented AI-generated anti-Muslim imagery circulating across X, Instagram and Facebook.
Its 2025 study analyzed 1,326 AI-generated posts from 297 public accounts and found that sexualized depictions of Muslim women received more than 6.7 million interactions across the dataset.
That research concerns broader harmful imagery, not simply hijab removal, so the two issues should not be treated as identical.
The Problem Isn't Just the Image Generator
Even a perfect moderation system cannot solve everything after an image is created.
The next stages are distribution and amplification.
A manipulated image can be copied, reposted, screen-recorded and moved between platforms.
That means AI companies, social networks, search engines and users all become part of the safety chain.
The Full Abuse Pipeline
- Creation: An AI tool generates or edits the image.
- Distribution: Someone uploads it to a public platform.
- Amplification: Algorithms recommend, repost or surface it.
- Persistence: Copies continue circulating after the original is removed.
- Harm: The target may face harassment, impersonation or reputational damage.
Stopping the problem at only one stage leaves the other stages intact.
What U.S. Users Should Know About the TAKE IT DOWN Act
There is now an important federal framework for a specific category of AI abuse.
The TAKE IT DOWN Act became law in May 2025, and the FTC began enforcing its platform notice-and-removal requirements in May 2026.
The law covers nonconsensual intimate real imagery and digital forgeries, including images created or altered with artificial intelligence.
Important Legal Boundary
The law is specifically about nonconsensual intimate imagery. A non-intimate image showing a woman without a hijab is not automatically covered simply because a hijab was removed. The legal analysis depends on the content and circumstances.
For covered intimate imagery, platforms must provide a reporting mechanism and remove qualifying content and known identical copies within 48 hours of a valid request.
What to Do If an AI Image Is Used Against You
The FTC advises people to report nonconsensual intimate imagery directly to the platform first.
It also points people toward StopNCII.org, which creates a digital fingerprint of an image and shares the fingerprint with participating platforms rather than uploading the original image itself.
For covered platforms that fail to follow the federal removal requirements, the FTC provides a reporting process through TakeItDown.ftc.gov.
Practical Safety Tip
Preserve evidence before reporting: save the URL, capture screenshots, record the account name and note when you discovered the content. Avoid repeatedly reposting the harmful image just to document it, because that can increase its spread.
Melvin Kranzberg Had the Right Framework
“Technology is neither good nor bad; nor is it neutral.”
Kranzberg wrote that principle in the 1980s, decades before generative AI.
But it fits this problem unusually well.
An image editor does not have to be malicious to produce harmful outcomes.
The same feature that lets a photographer change an outfit for a legitimate creative project can be abused to alter a person's identity or create misleading material.
The technology is the same.
The context is different.
What AI Companies Should Be Measuring
Traditional safety tests often ask whether a model refuses obviously prohibited prompts.
Image editing suggests a more difficult test.
A Better Safety Evaluation
- Intent: What is the user really trying to change?
- Identity: Does the edit alter an important personal characteristic?
- Consent: Is a real person's likeness involved?
- Context: Is the edit likely to misrepresent the subject?
- Reframing: Does a harmless-sounding alternative prompt bypass the original safeguard?
The fifth question is especially important.
A safety system that only blocks certain words can be much easier to bypass than one that understands what the requested transformation actually does.
Pros and Cons of Today's AI Image Editing Approach
Useful Capabilities
- Fast creative experimentation.
- Powerful photo and design workflows.
- Accessibility for users without advanced editing skills.
- Rapid professional prototyping.
- Useful transformations when the subject has given permission.
Safety Challenges
- Intent can be difficult to infer from text alone.
- Prompt reframing can expose inconsistencies.
- Religious or cultural harm may fall outside traditional nudity filters.
- Real-person edits create additional privacy concerns.
- Copies can survive after the original is removed.
Amazon: Simple Privacy Tools Worth Considering
Webcam Privacy Cover
A physical webcam cover is a simple way to reduce accidental camera exposure on laptops and external webcams. It does not stop AI image abuse, but it adds a useful layer of physical privacy.
Browse Webcam Privacy Covers2FA Security Key
Strong account security matters because attackers may combine stolen photographs or accounts with AI manipulation. A hardware security key can provide stronger two-factor authentication for supported services.
Browse Security KeysPrivacy Screen for Laptop
A physical privacy filter can reduce the ability of people nearby to photograph sensitive work displayed on your screen, which is useful when handling private images or documents in public spaces.
Browse Laptop Privacy ScreensWatch the FTC's Take It Down Act Guide
This official English-language FTC video explains how to report platforms that fail to remove qualifying nonconsensual intimate images under the federal law.
The Bottom Line on AI Chatbots and Hijab Edits
The latest chatbot tests expose a weakness in the way AI image safety is often discussed.
Safety is not only about whether a model produces nudity, violence or other obviously prohibited content.
Context matters.
Changing a person's religious or cultural appearance can carry meaning even when the resulting image contains nothing sexually explicit.
The Guardian's test does not prove that every AI image edit is harmful, nor does it establish that every chatbot will behave the same way after future model or policy updates.
What it does show is that major systems can interpret the same type of transformation differently.
Earlier research from CSOH and WIRED shows that religious and cultural imagery has already been pulled into broader AI-enabled harassment campaigns.
That makes robust safeguards increasingly important.
For developers, the lesson is straightforward: do not build an image-safety system around prohibited words alone.
For users, the practical lesson is equally important: treat AI-generated images as editable representations, not automatically harmless digital toys.
And for people targeted by qualifying intimate deepfakes in the United States, there are now formal federal reporting and removal mechanisms.
AI makes visual manipulation easier.
It should also make responsible editing, consent and protection easier — not optional.
The Real Threats of AI Misuse
Exploiting image-editing guardrails is just one piece of a much larger puzzle. From autonomous agent vulnerabilities to hyper-targeted digital manipulation, the landscape of AI abuse is shifting rapidly. Read our complete guide to uncover the real AI threats nobody is talking about.
Read the AI Misuse Guide →Sources checked for this article:
The Guardian — AI chatbots remove hijabs from images of Muslim women when prompted
WIRED — Grok and AI manipulation of women in religious and cultural clothing
CSOH / Al Jazeera — AI-enabled abuse targeting Muslim women
Anthropic — Can Claude Produce Images?
xAI — Reporting Non-Consensual Intimate Content
Federal Trade Commission — Complying With the TAKE IT DOWN Act
FTC Consumer Advice — What to Do About Nonconsensual AI Images
Federal Trade Commission — How To Report Platforms That Violate the Take It Down Act
Frequently Asked Questions
Can AI chatbots remove a hijab from an image?
According to The Guardian's October 1, 2026 tests, ChatGPT and Grok performed a hijab-removal edit, while Gemini initially refused and later generated an image without the head covering after a more “Western” framing. Claude did not perform the edit and Anthropic says Claude does not generate images.
Does removing a hijab from an AI image violate U.S. law?
Not automatically. A non-intimate image showing a woman without a hijab is not automatically covered by the federal TAKE IT DOWN Act. That law addresses nonconsensual intimate imagery, including qualifying AI-generated or altered intimate images.
What is the TAKE IT DOWN Act?
The TAKE IT DOWN Act is a U.S. federal law covering the intentional disclosure of nonconsensual intimate visual depictions and requiring covered platforms to provide a removal process. Qualifying real and AI-generated intimate imagery must generally be removed within 48 hours of a valid request.
What should I do if someone posts an AI-generated intimate image of me?
Report the content to the platform, preserve evidence and use available removal resources. The FTC directs users to TakeItDown.ftc.gov, and StopNCII.org can create a digital fingerprint of qualifying images for participating platforms.
Why is religious clothing an AI safety issue?
Religious clothing can be an important part of how a person expresses identity and faith. An AI system that treats every clothing or appearance change as an ordinary visual edit may fail to account for consent, dignity, cultural context and the possibility of misrepresentation.
No comments:
Post a Comment