Beyond Takedowns: Decode-Tattle Report Maps Ways To Tackle AI Deepfakes
The report by Decode Internet and Tattle Civic Technologies builds on their investigation into the cross-border infrastructure powering AI tools used to create non-consensual intimate images of women.
The report, Deconstruct the Deepfake Stack: An Action Plan to Tackle Non-Consensual Synthetic Imagery in India, calls for stronger safeguards in AI models, clearer legal provisions and action against services and payment networks that enable non-consensual intimate imagery.
Published on October 12, the report argues that removing harmful content after it appears online is not enough to tackle an ecosystem of AI models, image-generation apps, messaging bots, hosting services and payment networks.
It builds on Desi Faces, Foreign Servers: Inside the AI Economy Undressing Women, a joint investigation by Decode Internet and Tattle Civic Technologies, published in June 2026 with support from the Pulitzer Center. The investigation traced the cross-border infrastructure and commercial networks powering AI tools used to manipulate women's images without their consent.
A network that goes beyond individual perpetrators
The report maps six layers of this ecosystem, from training datasets and AI models to applications, search engines, hosting providers and payment networks.
The investigation found that services operating across countries were accessible to Indian users. Some Telegram bots offered image-generation tools in Indian languages, accepted UPI payments and used local agents to convert payments into credits. Seemingly separate services also relied on shared backend systems, allowing them to reappear under new names.
An analysis of 31 image and video generation websites found that around two-thirds used Cloudflare as their content delivery network, complicating efforts to identify their operators.
The report argues that removing a website, bot or image may not dismantle the infrastructure enabling the abuse.
AI tools are easy to find
The investigation found that websites and apps capable of generating nude or intimate imagery appeared on the first page of Google, DuckDuckGo and Brave search results. Chatbots, including ChatGPT and Claude, also suggested apps and Telegram bots offering these services.
The report also found that 67% of the top 50 trending models on Hugging Face that enabled video generation from images using text prompts were designed to create sexual scenes. Nearly all were fine-tuned versions of models from the Wan series.
These findings highlight how users can discover and access such tools, while the underlying services remain available even when individual content is removed.
The investigation was followed by a discussion at Method Art Gallery in Delhi.
Ways to tackle the problem
The report proposes four broad interventions.
Legal clarity: Regulators should clarify how existing laws, including the Digital Personal Data Protection Act, 2023, and copyright law, apply when publicly available photographs are used to train AI models capable of generating non-consensual imagery.
Action against enabling services: Enforcement should extend beyond individual images to apps, websites, bots and model repositories that facilitate their creation and distribution, including services operating across borders.
Stronger AI safeguards: Developers should strengthen safety evaluations, address explicit material in training datasets where appropriate and restrict models that enable non-consensual imagery. The report acknowledges that such measures can be technically difficult and that prompt restrictions can be circumvented.
Cutting off financial incentives: Payment networks and advertising platforms should be held accountable for enabling or profiting from these services. The report recommends measures to disrupt payments to services that facilitate abuse.
Beyond removing harmful content
The report draws on the Decode-Tattle investigation into 'Babydoll Archi', an AI-generated persona created using a real woman's face without her consent. The persona attracted a following and was monetised, while the woman whose likeness was used withdrew from public life.
The investigation showed how the tools and services enabling the abuse extended beyond an individual perpetrator.
India has rules requiring platforms to remove certain harmful content within set timelines. But the report says takedowns alone cannot solve the problem, as images can be reuploaded, AI models can continue circulating elsewhere, and action against services operating abroad can be slow. Faster takedowns may also lead to legitimate content being wrongly removed.
The report argues that addressing non-consensual synthetic imagery will require action across the entire chain, from AI development and discovery to hosting and payments, alongside safeguards against disproportionate enforcement.
