Brown researchers: deepfake laws miss the abuse ecosystem

A new analysis from Brown University finds that the federal law against deepfakes, and the state laws around it, go after the people who make abusive images while leaving most of the machinery that produces them untouched. The TAKE IT DOWN Act, passed in May 2025, was the first federal law to address AI-generated nonconsensual intimate images; the Brown team argues it and its state counterparts aim at the wrong layer.

The study comes from computer scientists at Brown's Center for Technological Responsibility, Reimagination and Redesign. The team presented it on Nov 7, 2026, at the 2026 ACM Conference on Equity and Access in Algorithms, Mechanisms and Optimization, and built an interactive companion site that walks readers through the same map.

What the Brown researchers found

Over 92% of the 64 state and federal laws the researchers analyzed target individual creators and distributors of abuse material, not the technology ecosystem. That focus, they write, sets up an endless game of "whack-a-mole" that chases one image or one app at a time while the systems behind them keep running.

The paper says its legal survey covers the federal TAKE IT DOWN Act plus 63 "intimate deepfake" state laws as tracked by Public Citizen. Brown's release counts 64 state and federal laws in the analysis. The finding holds across both tallies: the statutes converge on the person making or sharing the image, not the platforms and services in between.

How the 11-layer ecosystem works

The study sorts the technologies that enable AI-generated nonconsensual intimate images into 11 categories across five roles. They are creation (the AI models, user interfaces and training datasets), distribution (social media and private messaging apps), proliferation and discovery (search engines, app stores and deepfake communities), infrastructure and support (developer platforms like GitHub and Civitai), and monetization (credit card companies and payment processors).

The layers feed each other, and the researchers use the 2025 Grok episode to show how. Grok's X-based interface let users reach the Grok AI model (creation) and make nonconsensual sexual images by tagging the bot with prompts like "put a bikini on her" or "remove her clothes." Grok posted the results straight into X comment threads (distribution). A community formed on Telegram (proliferation and discovery) to work out how to make more explicit images and video by jailbreaking a Grok interface sold through app stores (monetization).

The open-model side runs on the same wiring. The paper cites research counting almost 35,000 fine-tuned "deepfake" variants on the model hubs Hugging Face and Civitai, a majority of them sexual. After Civitai banned models depicting real people's likenesses in April 2025, 404 Media found users downloaded more than 5,000 models and reuploaded them to Hugging Face, the paper notes.

None of this is new. The paper traces the early proliferation to 2017, when people on Reddit used face-swap technology to make nude images of celebrities. In 2025, users on X used Grok to make thousands of sexual images of mostly women and girls. AI "nudifier" apps that turn a photo into explicit images or video in seconds are now commonplace, often in mainstream smartphone app stores.

Why the laws fall short

Most statutes name an actor; almost none name a technology. Tennessee's H0769 (2025) goes after people who make "technology, software, or digital tools" for creating the material. Arkansas's HB1529 (2025) reaches the "provider or developer of image generation technology" used to make deepfake material, which touches the upstream model and interface. The researchers present those two as exceptions rather than the pattern.

The gap shows up in what the ecosystem keeps doing while the law watches individuals. Independent researcher Genevieve Oh ran 6,000 Google searches across 100 public figures and found that 99.69% of "NAME+DEEPFAKE" queries returned a deepfake pornography site on the first page, according to the paper. In 2025, 47 state attorneys general wrote to Google, Yahoo and Microsoft asking them to block the material. A separate audit of 85 AI nudifier sites found Amazon and Cloudflare provided hosting or content delivery for 62 of them, and Google enabled sign-in for 53.

For creators who build with generative video, the practical point is where the pressure points sit. The researchers argue that takedown-only mitigations are reactive, and that the upstream model, the developer platform hosting a fine-tuned variant, or the payment processor can be a more durable place to intervene.

What the researchers want to change

Shift the focus from the image to the infrastructure that produces it. "Grok was generating over 6,000 images every hour last December," said Michelle Ding, a computer science Ph.D. student at Brown and the study's lead author. "We need to shift our focus to the technical infrastructure enabling and profiting from this abuse."

Ding frames the whole project around that idea. "The core issue is that deepfake abuse is not just about one image, model, website or app, but an entire ecosystem," she said. "In order to protect vulnerable communities online from the various types of harms and emotional distress that can come from this form of abuse, we must first see this ecosystem clearly."

Harini Suresh, an assistant professor of computer science at Brown and a study co-author, said the goal is to push the creation side onto the research agenda. "Current mitigations are all very reactive — taking images down once they're out there," she said. "There's a gap on the creation side of things. Can we motivate AI researchers, for example, to think about how to prevent a model from creating these images in the first place?"

Where this goes next

The team wants the map to become a working tool. The companion site pairs an interactive story for general readers with a searchable database built for researchers, journalists and policymakers, including a map of U.S. intimate deepfake laws that lets users compare state rules and spot gaps. The researchers say they hope to form partnerships across journalism, civil society, policy and industry to keep the database current as the ecosystem changes.

The authors put the next step as a research question: whether models can be built or steered to refuse this output at the source. Brown's map is a starting point for that work.

The paper is at ACM, and the interactive map is free to explore at ncii-ecosystem.org.