Women's Privacy in the Age of AI-Generated Images

In early 2025, Syeda Rizwana Hasan, an advisor to Bangladesh's Ministry of Environment at that time, opened Facebook to find her face placed onto a body lifted from an adult website. The image had been circulated by an account called "Chemical Ali," described by researchers as a repeat offender known for targeting prominent Bangladeshi women. She had not posed for it. She had not consented to it. She could not fully remove it. Hasan is not an isolated case; she is one data point in a pattern accelerating faster than the law, the platforms, or the culture around her can respond to it.

The technology behind this is no longer exotic. Generative AI tools can produce a convincing fake intimate image from a single photograph, a wedding post, a university ID, or a LinkedIn headshot within minutes at almost no cost. Enforcement is only beginning to catch up. Just this month, US federal agencies seized two of the largest deepfake distribution sites, CFake.com and SOCFake.com, under the country's new TAKE IT DOWN Act. Last month, the same law was used to arrest two men, Cornelius Shannon and Arturo Hernandez, for publishing thousands of deepfake images of women, both prominent figures and private individuals with no public profile at all. These are among the first real enforcement actions against a problem that has existed, largely unpunished, for nearly a decade.

The numbers behind that decade are stark. Research compiled by the Office for the Prevention of Domestic Violence found that the overwhelming majority of AI-generated deepfakes circulating online are sexual in nature and that nearly all of them depict women. A 2023 study by the cybersecurity firm Home Security Heroes documented a 550 percent rise in deepfake videos since 2019. A UN-linked report published this past April described coordinated campaigns engineered specifically to humiliate and discredit women in public life, politicians, journalists, and activists until they withdraw from it.

Bangladesh's legal response has not kept pace. The Cyber Security Ordinance 2025, which replaced the widely criticized Digital Security Act, was the first South Asian legal instrument to explicitly acknowledge AI-enabled cybercrime, a genuine first. But legal commentators have noted that its language on deepfakes remains vague enough to leave victims dependent on older statutes never designed for synthetic media and an under-resourced Police Cyber Support Centre for Women that cannot move at the speed at which the content spreads. The National Parliament does not need to invent this framework from scratch. It can adapt the US TAKE IT DOWN Act's 48-hour platform takedown mandate or the EU AI Act's disclosure requirements for AI-manipulated content. What it cannot do is keep treating this as a subset of ordinary harassment law.

A 2026 academic study based on interviews with Bangladeshi women found something that distinguishes this harm from older forms of online abuse: the fear is rarely about whether people believe an image is fake. It is about how quickly a fabricated image becomes accepted as evidence, regardless. The researchers documented the case of a 23-year-old student in Rajshahi, identified by the pseudonym Riya, whose face was manipulated into explicit images and circulated through her university network after she became visible through student organizing. She was pressured to resign from every group she belonged to. Her mother told her to leave the campus. She considered reporting it to the police and decided against it, certain they could not trace the perpetrators and afraid a complaint would draw more attention to the images, not less.

That fear is not irrational. Bangladeshi press has also reported at least one case in which a woman took her own life after AI-manipulated content reached her family, a reminder of what is actually at stake when institutions move slowly. Organizations like Bangladesh Mahila Parishad and Digitally Right have stepped into that gap, offering victims legal guidance and a place to report when police capacity falls short. But the system still asks the targeted woman to act as her own investigator, screenshotting and logging URLs and timestamps, often within hours, before evidence disappears along with the post. That burden should sit with platforms and law enforcement, not with the woman who has already been violated once.

A predictable objection to fast-takedown laws is that they risk overreach, that platforms pressured to remove content within hours will err toward deletion, sweeping up legitimate speech alongside genuine abuse. It's a real concern, and one reason laws like TAKE IT DOWN build in penalties for bad-faith reporting. But the asymmetry of harm here is not close. A wrongly removed post can be appealed and restored. A circulated intimate image cannot be uncirculated. Erring toward swift removal, with safeguards against misuse, is the correct tradeoff.

None of this is an argument against AI itself. It is an argument that apps and sites built specifically to strip clothing from a photograph, the CFakes and SOCFakes of the world, are not a neutral byproduct of useful technology, and that the app stores and payment processors that host or fund them should be treated as complicit, not neutral. It is an argument that Bangladesh's parliament should amend the Cyber Security Ordinance to name AI-generated nonconsensual imagery explicitly, rather than leaving it to be inferred. It is an argument that Meta and other platforms operating in Bangladesh should be held to the same 48-hour removal standard their US operations are now legally bound to honor. And it is an argument that no woman should have to choose, as Riya did, between her safety and her own case file.

The account that targeted Hasan is still described by researchers as a repeat offender, meaning it has done this before to other women, and nothing so far has stopped it from doing it again. Until the law treats the manufacture of a woman's image with the same seriousness it gives to the theft of her money, that pattern will keep repeating: one woman, one image, one platform delay at a time.