Full title: Recommended Amendments to Legislative Proposals on Deepfakes

NO FAKES Act

The NO FAKES Act of 2023 was introduced by Sens. Coons, Blackburn, Klobuchar, and Tillis.

• Creates a property right over one’s own likeness in relation to digital replicas which are nearly indistinguishable from the actual image, voice, or visual likeness of that individual.
• Imposes civil liability for the non-consensual production, publication, distribution, or transmission of a digital replica.
• States that it would not be a defense for the defendant to have included a disclaimer claiming that the digital replica was unauthorized or that they did not participate in its creation, development, distribution, or dissemination.

Analysis: This proposal would enable victims to bring civil liability claims against persons who produce, publish, transmit, or otherwise distribute deepfakes of them. While this might capture model developers or providers whose tools enable the production of such deepfakes, this interpretation should be made clearer by the bill. Moreover, civil liability for breaching a property right will not be as strong of a deterrent as criminal liability, which better reflects the level of harm deepfakes pose to the public.

NO AI FRAUD Act

The NO AI FRAUD Act of 2024 was introduced by Rep. Salazar along with other House lawmakers.

• Creates a property right in each person’s rights and likeness.

• Imposes civil liability for distributing, transmitting, or otherwise making available to the public a “personalized cloning service”, with a remedy of at least $50,000.
• Defines personalized cloning services as algorithms, software, tools, or other technologies, services, or devices the primary purpose or function of which is to produce digital voice replicas or digital depictions of particular, identified individuals.
• Also imposes civil liability for publishing, performing, distributing, transmitting, or otherwise making available to the public a deepfake.

Analysis: This proposal is strong in that it specifically incorporates model developers and providers through a civil liability provision for persons distributing, transmitting, or otherwise publishing a “personalized cloning service”. However, the definition of “personalized cloning service” is limited by requiring the “primary purpose” of the AI system to be creating digital replicas. Because generative AI systems are often multi-purpose, this definition would therefore fail to capture systems which were not specifically designed to generate harmful deepfakes but which are actively being used for this purpose. Furthermore, the proposal requires actual knowledge that the individual or rights holder did not consent to the conduct in order to hold the creator, distributor, or other contributor liable. This standard is remarkably hard to meet and may discourage moderation of content produced by these services to ensure that there is plausible deniability of knowledge of what would otherwise be a violation of the bill. The remedy for providing such a service would also start at only $50,000 per violation. While the generation of many deepfakes would constitute multiple violations for those creating the deepfakes themselves, the publication of a service used to create them, i.e. the source of the deepfake supply chain, would seemingly only constitute a single violation regardless of the number of deepfakes produced by the service. This penalty may therefore be insufficient for incentivizing the incorporation of robust safeguards by model developers and providers.

Preventing Deepfakes of Intimate Images Act

The Preventing Deepfakes of Intimate Images Act was introduced by Rep. Morelle and Rep. Kean.

• Prohibits the disclosure of intimate digital depictions, defining such depictions broadly to include sexually explicit content or altered images.
• Outlines both civil and criminal penalties for unauthorized disclosure, provides relief options for victims, and includes exceptions for certain disclosures, such as those made in good faith to law enforcement or in matters of public concern.

Analysis: The bill successfully creates a two-pronged approach to combating non-consensually generated, sexually explicit material by introducing both civil and criminal penalties. The bill draws on the definition of ‘disclose’ outlined in the Violence Against Women Act Reauthorization Act of 2022 which is to “transfer, publish, distribute, or make accessible.” While this definition might encompass the actions performed by platforms which facilitate the distribution of deepfakes and developers who enable the creation of deepfakes, the language of the bill must be clarified in order to clearly capture these key entities in the supply chain.

AI Labeling Act

The AI Labeling Act of 2023 was proposed by Sens. Schatz and Kennedy.

• Requires that each AI system that produces text, image, video, audio, or multimedia AI-generated content include on such AI-generated content a clear and conspicuous notice.
• Requires that the notice include an identification that the content is AI-generated, the identity of the tool used to create the content, and the date and time the content was created.
• Requires that, to the greatest extent possible, the disclosure must also be permanent or difficult to remove.
• Imposes a responsibility on developers and third-party licensees to “implement reasonable procedures to prevent downstream use of such system without the disclosures required”. There is also a duty to include in any licenses a clause prohibiting the removal of the disclosure notice.

Analysis: Disclosing the tool used to generate the content in the notice can make it easier to identify and hold developers accountable for the harmful deepfakes their models create. This can also help identify models that are particularly vulnerable to being employed for the creation of deepfakes due to insufficient design safeguards. Though we suggest the authors clarify that identification of the tool used to create the content should include, at a minimum, the model name, model version, and the name of the model’s developer, we applaud the authors’ foresight in including this requirement. We further note that many deepfakes, such as those which are sexually abusive in nature, are harmful in and of themselves, even if it is known that the content has been manipulated. As such, while useful, labeling of AI-generated content alone is likely insufficient to discourage deepfake proliferation and incentivize responsible design of generative AI systems without imposing liability for harms from the deepfakes themselves. That said, this may best be accomplished through complementary legislation.

AI Disclosure Act

The AI Disclosure Act of 2023 was introduced by Rep. Torres in 2023.

• Requires that any output generated by AI includes: “Disclaimer: this output has been generated by artificial intelligence.”
• States that failing to include this disclaimer would violate the Federal Trade Commission Act (15 U.S.C. 57a(a)(1)(B)) regarding unfair or deceptive acts or practices.

Analysis: These types of disclaimers can be easily removed from the outputs of generative AI, such as videos, images, and audio, meaning malicious actors could continue to utilize deepfakes for deceptive purposes, which would be made more deceptive by absence of a disclaimer. It is also not clear which party would be responsible for the lack of disclaimer in the output. It would be most reasonable for the model developer to bear this responsibility given that they can designs models to ensure all outputs incorporate this disclaimer, but without any requirement to include information identifying the generative AI model used to create the content, enforcement would be extremely challenging.

DEEPFAKES Accountability Act

The DEEPFAKES Accountability Act was introduced by Rep. Clarke.

• Requires that deepfakes include a contents provenance disclosure, such as a notice that the content was created using artificial intelligence.
• Requires that audiovisual content includes no less than one verbal statement to this effect and a written statement at the bottom of any visual component.
• Imposes a criminal penalty for a failure to comply with disclosure requirements where the noncompliance is:
  • intended to cause violence or physical harm;
  • intended to incite armed or diplomatic conflict;
  • intended to interfere in an official proceeding;
  • in the course of criminal conduct related to fraud;
  • intended to humiliate or harass a person featured in the deepfake in a sexual manner;
  • by a foreign power intending to influence a domestic public policy debate, interfere in an election, or engage in other unlawful acts.
• Also applies the criminal penalty where a person removes or meaninfully obscures the disclosures, with the intent of distribution.
• Grants a private right of action against a person who fails to compy with the disclosure requirements, even if there is only a “tangible risk” of suffering the enumerated harms.

Analysis: While the inclusion of criminal penalties may increase compliance and reflects the range of harms posed by deepfakes, this penalty would only apply for failure to include a contents provenance disclosure. Therefore, producing and sharing harmful deepfakes would still be permitted. Moreover, the bill includes an exception to the disclosure requirements for deepfakes created by an officer or employee of the United States in furtherance of public safety or national security, which could create a wide loophole. Finally, the contents provenance provision does not appear to require disclosure of the specific model that was used to generate the deepfake content; as such it would be difficult to identify the upstream model provider responsible for a harmful deepfake.

Protecting Consumers from Deceptive AI Act

The Protecting Consumers from Deceptive AI Act was introduced by Reps. Eshoo and Dunn.

• Directs the National Institute of Standards and Technology to establish task forces to develop watermarking guidelines for identifying content created by AI.
• Requires providers of generative AI applications to ensure that content created or modified by their application includes machine-readable disclosures acknowledging its AI origin.
• Requires providers of generative AI applications to make available to users the ability to incorporate, within the metadata of content created or modified by the application, information including the AI application used to create or modify the content.
• Requires providers of covered online platforms to prominently display disclosures included in the content accessed through their platform.
• States that violations of these watermarking provisions would constitute unfair or deceptive practices under the FTC Act.

Analysis: The development of content provenance standards through dedicated NIST task forces would be valuable in establishing best practices for watermarking. While the proposal requires that application providers enable users to include the model name and version in the metadata of AI-generated content, application providers and model developers should be required to automatically embed this information in the metadata in order to facilitate tracing back to the relevant developer and provider. This enables greater accountability and incentivizes design choices that incorporate protections against deepfakes by default. However, the bill does not address how users, developers, and platforms might be held accountable for the harmful deepfakes they create, publish, and distribute outside of simply requiring watermarking of these deepfakes.

DEFIANCE Act

The DEFIANCE Act of 2024 is sponsored by Sens. Durbin, Hawley, Klobuchar, and Graham.

• Imposes civil liability for disclosing, soliciting, and intending to disclose a “digital forgery”.
• Defines “digital forgery” as “any intimate visual depiction of an identifiable individual” created through means including artificial intelligence, regardless of whether it includes a label.

Analysis: The bill explicitly states that sexually explicit deepfakes which include disclaimers should still be subject to civil liability, and it also prohibits the production and possession of sexual deepfakes in certain cases. However, it is not clear that it limits a provider’s ability to supply individuals with deepfake-generating services, nor does it incentivize developers of AI systems to ensure that their models cannot produce this content. Furthermore, we note that the definition for a “digital forgery” may be too broad in that it could also capture manipulated content created by means other than artificial intelligence, such as traditional image editing software. AI-powered production and alteration of visual content arguably necessitates a distinct approach from traditional tools, as AI technology trivializes the process of producing realistic deepfakes at scale.