Policy brief · October 2026

Likeness, consent and generative media

Enforcement where synthetic media is made.

California and Congress have begun to treat a forged face or voice as a legal injury. This brief describes the technical controls that can carry those laws into the services that generate media, and recommends them as a baseline for every such service.

1. Summary

A person’s face and voice now work as credentials. They are how families, employers, banks and courts recognize who is speaking, and generative models have made them inexpensive to counterfeit.

Recent law answers with remedies. California’s SB 1111 brings digital replicas into the state’s right-of-publicity and false-impersonation statutes.⁠[2] The DEFIANCE Act, passed by the Senate in January 2026 and pending in the House, gives the subjects of intimate digital forgeries a federal civil action.⁠[4][5] Both take effect after a forgery exists and has been found.

This brief argues for the complementary layer: controls inside the services that generate media, which block a protected likeness while an image forms, mark every output, and detect forgeries once they circulate. We recommend that this pipeline be expected of every service that generates realistic media of people, from open-model hosts such as Hugging Face to commercial products such as Grok on X.

Background

2. A likeness is a credential

Currency holds its value because people trust that a note is genuine. A counterfeit harms more than the person who accepts it: each one weakens trust in every note. A likeness works the same way. Its value rests on the shared assumption that the face on a call, or the voice on the phone, belongs to the person it resembles, and each convincing forgery weakens that assumption for everyone.

A likeness therefore warrants protection comparable to other instruments of trust: liability for counterfeiting it, and controls in the systems able to produce counterfeits.

Where the harm falls

  • Impersonation fraud

    Seconds of audio from a public video are enough to clone a voice. Cloned voices of relatives are used to press older people for money or personal information, and synthetic executives on video calls are used to authorize payments.

  • Nonconsensual intimate imagery

    Tools that undress or face-swap a real person need no technical skill. The DEFIANCE Act’s findings note that the harm is not mitigated by labeling the image as fake.⁠[4]

  • Harassment at scale

    Collecting a person’s images and generating new ones can be automated together, so one individual can be targeted with fabricated material at volume.

  • On mainstream platforms

    In late December 2025 and January 2026, Grok, the assistant built into X, was used to produce sexualized images of real people, including women and children. A Center for Countering Digital Hate study estimated three million such images over eleven days. The Senate passed the DEFIANCE Act by unanimous consent amid the response.⁠[9][5]

    Open model hosting meets the same problem in a harder form. A peer-reviewed Oxford Internet Institute study of Hugging Face and Civitai found almost 35,000 publicly downloadable model variants made to depict specific people, downloaded nearly 15 million times since November 2022; 96% targeted women.⁠[10] When Civitai removed models of real people in 2025, 404 Media reported that more than 5,000 of them had been re-uploaded to Hugging Face, whose content policy prohibits sexual content made without consent.⁠[11] A host can moderate what is uploaded, but not what happens after a download, which is why standards for what may be published, such as the disclosure duty California places on hosting platforms from 2027, reach further.⁠[7]

    Voice raises the same questions. In May 2024, OpenAI paused “Sky,” one of ChatGPT’s voices, after Scarlett Johansson said it sounded “eerily similar” to her own; she had declined the company’s offer to voice the system. OpenAI said Sky was not an imitation but the natural voice of a different actress. Johansson’s lawyers asked the company to disclose how the voice was developed, which is the question a provenance record exists to answer.⁠[12]

  • In app stores and ad networks

    App review has not reliably caught these tools. In April 2024, Apple removed three apps advertised as able to make nonconsensual nude images only after 404 Media sent it links to the apps and their ads.⁠[13] A January 2026 Tech Transparency Project review found 47 apps in the App Store that digitally undress people; Apple removed 28 of them after the group and CNBC contacted it.⁠[14][15]

    Advertising review has missed them as well. The developer of the Perky AI app ran 11 ads on Meta’s platforms using a manipulated photo of the actor Jenna Ortega taken when she was 16; its page had run more than 260 ads before Meta suspended it at NBC News’s inquiry, and Apple then removed the app.⁠[16] In June 2025, CBS News found at least hundreds of ads for such tools across Facebook, Instagram, Threads, Messenger and Meta’s Audience Network, one of them using sexualized deepfakes of Scarlett Johansson and Anne Hathaway.⁠[17] Meta has since sued one app’s developer for repeatedly circumventing its ad review.⁠[18] Review at the storefront or the ad network sees how a tool is described, not what it produces.

3. Where the law stands

California SB 11 · Sen. Angelique Ashby

Passed the Legislature, 2025 · Vetoed October 13, 2025

Would have treated a digital replica as a person’s voice or likeness under the right-of-publicity statute (Civ. Code § 3344), made using a digital replica to impersonate someone a form of false impersonation, and directed the Judicial Council to review AI-generated evidence. It would also have required any service that lets users create digital replicas to warn them of civil and criminal liability, with civil penalties of up to $10,000 a day for a missing warning.⁠[1][3]

California SB 1111 · Sen. Angelique Ashby

Signed September 30, 2026 · Chapter 862, Statutes of 2026

Enacts SB 11’s core without the warning requirement. Under Civil Code § 3344, a person’s voice or likeness includes a digital replica, and the statute’s presumption for incidental use of an employee’s likeness is removed. For every Penal Code offense in which false impersonation is an element, including §§ 528.5, 529 and 530, impersonation includes the use of a digital replica with intent to impersonate. It is the author’s third bill on the subject, after SB 970 (2024) and SB 11 (2025).⁠[2][3]

DEFIANCE Act of 2025 · H.R. 3562 and S. 1837

Rep. Alexandria Ocasio-Cortez with Rep. Laurel Lee; Sen. Richard Durbin with Sen. Lindsey Graham · Passed the Senate January 13, 2026 · Pending in the House

Amends 15 U.S.C. § 6851 to define an intimate digital forgery and to let the person depicted sue anyone who knowingly produces, discloses, or solicits and receives one, or possesses one with intent to disclose it, knowing or recklessly disregarding that the person did not consent. Relief includes liquidated damages of $150,000, or $250,000 where the conduct is connected to sexual assault, stalking or harassment; punitive damages; injunctions ordering deletion; and privacy protections such as filing under a pseudonym. A label stating that the image is fake does not take it outside the definition.⁠[4][5][6]

Related law

  • California AI Transparency Act

    Since August 2, 2026, generative AI providers with more than one million monthly users must embed latent disclosures in the images, video and audio their systems create, and offer a free detection tool. From January 1, 2027, platforms that make model weights or source code available for download may not knowingly offer a system that omits those disclosures, and large online platforms must detect standard provenance data.⁠[7]

  • TAKE IT DOWN Act

    Makes publishing nonconsensual intimate imagery, including AI-generated forgeries, a federal crime, and since May 19, 2026 requires covered platforms to remove reported images within 48 hours.⁠[8]

4. The gap

Remedies depend on what a victim can find and prove. A person must first discover the forgery, then identify who made it, then show that they did not consent. The DEFIANCE Act’s own findings note that victims may not be able to determine who created the content, or how to prevent more from being made.⁠[4]

Warnings do not close that gap either. SB 11’s warning requirement did not become law, and SB 1111 kept the remedies without it.⁠[1][3]

What is missing is an enforcement point inside the systems that produce synthetic media, and evidence that travels with what they produce. Watermarking, provenance metadata and detection each have known weaknesses on their own.⁠[19] Applied together at the point of generation, each covers for the others.

Proposal

5. Prevent, mark, detect

Veil is being built as a closed system: one pipeline in three stages. Each stage leaves a record the next can rely on, so a likeness is protected before, during and after an image exists.

  1. Prevent

    In beta

    A layer inside the image generator checks the picture as it forms, between denoising steps, against photos people have registered. When the image begins to reproduce a registered photo, the run halts before the image is finished. Each check is kept as an auditable record of which identities, photos and regions matched, and at what score; work that draws on a photo without copying it is credited. Today the check matches pixels; matching a person’s likeness from the same registered photos is in development.

  2. Mark

    In beta

    Every output carries a watermark and signed provenance data naming the service and the run that made it, consistent with the latent disclosures California already requires.⁠[7] A finished image can be traced to its source without the cooperation of whoever shared it. Veil’s beta signs every image it finishes generating, reading the signature off the pixels rather than adding to them; how well it survives crops, re-encodes and screenshots is still being resolved.

  3. Detect

    In development

    Media in circulation is matched against registered photos and against the registry of marks. A forgery of a registered person that carries no mark points to a tool that skipped the pipeline, or to a mark that was removed. Either is evidence in itself.

    The stage builds on an earlier working prototype, Obscur, which is temporarily unavailable while it moves into Veil. It scores each patch of an image or video frame for signs of machine generation, and performs its measurement without collection: a single image, scored on-device, against no reference database, no device fingerprint, no biometric template, and no cross-image correlation.

The stages feed one another. A detected forgery can be reported for removal under the TAKE IT DOWN Act, traced to its producer for a civil claim, or referred under California’s false-impersonation statutes, and what got through informs the checks at generation. The open standard sets out each stage as a requirement, with open-source reference code.

6. How it serves each law

SB 1111 · right of publicityCiv. Code § 3344
The registry records whose likeness is protected and on what terms. Run records show whether a service used it, when, and how much of the output it accounts for: the facts a claim about use and consent turns on.
SB 1111 · false impersonationPenal Code §§ 528.5, 529, 530
Watermarks and provenance tie a replica used to impersonate someone back to the service and run that produced it.
DEFIANCE Act · intimate digital forgeries15 U.S.C. § 6851, as amended
With likeness matching, the generation check can refuse an intimate forgery of a registered person before it exists. When one surfaces, the mark answers the question victims most often cannot: who produced it.
SB 11 · consumer warningsVetoed, 2025
In place of a warning a user can click past, the service itself declines to produce the replica.
AI Transparency ActSB 942, as amended by AB 853
The mark and detection stages are a latent disclosure and a detection tool. The hosting-platform duty that begins in 2027 is the natural place to add consent checks.

These describe the evidence the pipeline can produce, not legal conclusions. How such records are weighed is for counsel and the courts.

7. Recommendations

  1. Treat a likeness as a protected credential

    Extend to a person’s face and voice the protection given to other instruments of trust, with liability for counterfeiting them that does not depend on whether the forgery was labeled. SB 1111 and the DEFIANCE Act are a foundation for this.

  2. Make the pipeline a service-level duty

    Any service that generates realistic images, video or audio of people, or hosts models that do, should check consent at generation, mark what it produces, and offer detection. The duty should apply alike to open-model hosts such as Hugging Face and to commercial products such as Grok on X, building on the hosting-platform provisions of AB 853.

  3. Establish a consent-verification registry

    A standard interface that services query before training on, editing or generating a covered likeness, returning consent status and licensing terms. Each query and each licensed output is logged, so unlicensed use can be reported and traced.

  4. Presume non-consent where no license is recorded

    Where a forgery is produced and no valid license appears in a registry at the time, the person depicted should be presumed not to have consented, subject to rebuttal by the defendant.

  5. Never require registration to sue

    Registration should strengthen a claim, never be a condition of bringing one. People who never register keep every remedy they have today.

  6. Offer a limited safe harbor to certified services

    A service that relied in good faith on a certified consent check at the time of generation should face limited damages. No safe harbor should authorize the production of an intimate digital forgery. Certification can rest on conformance to a published, open standard.

  7. Let people license their likeness

    Consent should be able to carry payment: likeness trusts that license and distribute royalties on members’ behalf, as performing-rights organizations do for music, and provenance credentials that pay the person each time a licensed likeness is used.

  8. Authenticate voices and faces on calls

    Phone networks and conferencing services should be able to flag synthetic audio and video. Caller ID can be spoofed, and the person on the other end cannot otherwise be checked.

Status and sources

8. What Veil is today

Veil is in beta for the Mac, and a working preview of the first stage runs on this site.

  • Working today

    Photos are registered in 32-pixel tiles. A layer inside open image models checks each generation against them between steps, credits the photos it draws on, and halts reproductions.

  • Today’s limit

    The current check compares pixels, so it protects registered photos rather than a person’s likeness in general: a rotated copy, or a new picture of the same person, is not caught. This is a limit of today’s matching, not of the approach.

  • Likeness from the same photos

    A registered photo can also be stored as DINO embeddings, which describe a person’s visual characteristics rather than exact pixels. Matching on them lets the check recognize the person in pictures it has never seen. Each additional photo, from another angle, in other light or at another age, widens that coverage.

  • Signing, in beta

    The beta signs any photo, and every image it finishes generating, without changing a pixel: the signature is read off the image and kept on the Mac, and a picture can be checked against the signatures kept. How well a signature survives crops, re-encodes, screenshots, rescales and warps is still being resolved.

  • In development

    Likeness matching, a signature that holds through those edits, and detection in the wild.

The recommendations above do not depend on Veil. They describe a standard any provider could meet, now drafted as an open standard; Veil is one implementation of it.

Sources

  1. [1]California SB 11 (Ashby), Artificial intelligence technology. 2025–26 Regular Session. Passed the Legislature; vetoed October 13, 2025. leginfo.legislature.ca.gov
  2. [2]California SB 1111 (Ashby), Digital replicas. 2025–26 Regular Session. Chapter 862, Statutes of 2026; approved by the Governor September 30, 2026. leginfo.legislature.ca.gov
  3. [3]Senate Committee on Public Safety, analysis of SB 1111 (Ashby). Hearing of April 21, 2026. spsf.senate.ca.gov
  4. [4]H.R. 3562, DEFIANCE Act of 2025 (Rep. Ocasio-Cortez, with Rep. Laurel Lee). 119th Congress. Referred to the House Committee on the Judiciary. congress.gov
  5. [5]S. 1837, DEFIANCE Act of 2025 (Sen. Durbin, with Sen. Graham). 119th Congress. Passed the Senate by unanimous consent, January 13, 2026. congress.gov
  6. [6]15 U.S.C. § 6851, Civil action relating to disclosure of intimate images. The provision the DEFIANCE Act amends. law.cornell.edu
  7. [7]California AI Transparency Act, Bus. & Prof. Code § 22757 et seq. SB 942 (Becker), Chapter 291, Statutes of 2024, as amended by AB 853 (Wicks), Chapter 674, Statutes of 2025. leginfo.legislature.ca.gov
  8. [8]TAKE IT DOWN Act, Pub. L. 119-12 (S. 146). 119th Congress. Signed May 19, 2025; platform removal requirements in force May 19, 2026. congress.gov
  9. [9]Allison Mollenkamp, “Paris Hilton joins House advocates behind deepfakes bill.” Roll Call, January 22, 2026. rollcall.com
  10. [10]Will Hawkins, Brent Mittelstadt and Chris Russell, “Deepfakes on Demand: The Rise of Accessible Non-Consensual Deepfake Image Generators.” Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25), Oxford Internet Institute. doi.org
  11. [11]Emanuel Maiberg, “Hugging Face Is Hosting 5,000 Nonconsensual AI Models of Real People.” 404 Media, July 15, 2025. 404media.co
  12. [12]Bobby Allyn, “Scarlett Johansson wants answers about ChatGPT voice that sounds like ‘Her.’” NPR, May 20, 2024. npr.org
  13. [13]Emanuel Maiberg, “Apple Removes Nonconsensual AI Nude Apps Following 404 Media Investigation.” 404 Media, April 26, 2024. 404media.co
  14. [14]Tech Transparency Project, “Nudify Apps Widely Available in Apple and Google App Stores.” January 2026. techtransparencyproject.org
  15. [15]Lola Murti, “Apple, Google host dozens of AI ‘nudify’ apps like Grok, report finds.” CNBC, January 27, 2026. cnbc.com
  16. [16]Kat Tenbarge, “Ads on Instagram and Facebook for a deepfake app undressed a picture of 16-year-old Jenna Ortega.” NBC News, March 5, 2024. nbcnews.com
  17. [17]Emmet Lyons and Leigh Kiniry, “Meta’s platforms showed hundreds of ‘nudify’ deepfake ads, CBS News investigation finds.” CBS News, June 6, 2025. cbsnews.com
  18. [18]Meta, “Taking Action Against ‘Nudify’ Apps.” Meta Newsroom, June 12, 2025. about.fb.com
  19. [19]Siddarth Srinivasan, “Detecting AI fingerprints: A guide to watermarking and beyond.” Brookings Institution, January 4, 2024. brookings.edu