Lawfare podcast questions whether AI consent tools can stop sexualized likeness misuse

Georgetown computer scientist Elissa Redmiles says product safeguards must distinguish between proving media is synthetic and preventing searches or generation involving a specific person.

A new Lawfare podcast on AI-generated sexual content argues that widely discussed technical safeguards may offer limited protection against the non-consensual use of a person’s likeness.

Lawfare Senior Editor Renée DiResta discussed the issue with Senior Editor Kate Klonick and Elissa Redmiles, an assistant professor of computer science at Georgetown University. Their conversation focused on the people creating AI-generated sexual content and on whether technical proposals can prevent AI systems from producing exploitative material.

Redmiles framed the central challenge as more than a question of detecting manipulated media. “Is this media fake?” differs from “Is this media of me?” she said. Systems that let users report or attest that content is synthetic may help establish that an image was generated, but they do not necessarily stop someone from searching for, creating, or distributing sexual content depicting a particular person.

That distinction matters for compliance teams designing consent controls. A product may be able to label generated content, restrict some prompts, or provide a reporting channel without reliably preventing a user from recreating a real person’s likeness through another model, service, workflow, or account.

Redmiles cited claims by DALL-E about using advanced techniques to prevent photorealistic generations of real people. She also referenced Civitai, a platform that hosts model-related objects, and its claims that people could opt in to blocking the sharing of models capable of generating their likenesses. Those examples prompted a practical question: whether automated protections can work consistently once a person’s likeness may be represented across many images, models, prompts, and platforms.

The discussion also touched on Meta AI in the context of opt-in and encouragement decisions, though the available transcript does not establish the precise notification, policy, or product change at issue.

For product leaders, the immediate takeaway is that consent features should not be marketed as complete prevention measures unless they have been tested against realistic circumvention paths. A likeness-protection setting on one platform may reduce misuse within that platform’s own model-sharing or generation environment. It does not necessarily prevent users from obtaining similar outputs elsewhere.

Redmiles cautioned that developers and policymakers need to define what counts as a sufficient safeguard and set expectations about what a control protects. In her framing, a mitigation that limits a narrow product function may still be useful, but it should not be presented as a guarantee that an individual cannot be depicted in non-consensual intimate imagery.

The enforcement problem follows from that limitation. Complaint and takedown systems can act after content appears, while generation restrictions seek to prevent creation in the first place. Neither approach alone resolves the broader search and replication problem when likeness-based content can move between services.

Uncertainty remains about whether the Lawfare discussion reflects new research findings, a specific policy proposal, or broader commentary. Lawfare also labels its transcript as auto-generated and potentially inaccurate. But the core compliance challenge is clear: technical safeguards may reduce certain forms of abuse, yet organizations need to state their scope, document known limitations, and avoid implying that consent tools can fully block misuse of someone’s likeness.