AI Disclosure is the Law
The EU AI Act, the California AI Transparency Act, and New York’s new synthetic performer advertising law, all require the disclosure of AI-generated content.
(Left to right) Presenters: Simon Erskine Locke (standing), Andy Parsons (Adobe), Samantha Rothaus and Richard Eisert (Davis+Gilbert).
The new disclosure obligations under these laws were the focus of a breakfast roundtable discussion organized by Davis+Gilbert and Tauth Labs, supported by H/Advisors and CommPRO in September.
The takeaways: If you publish AI-generated content for audiences in Europe, you may need to incorporate human- and machine-readable disclosures. If you are an AI company with more than one million monthly users, you may need to embed machine-readable disclosures in AI-generated images, video and audio, and offer users both a visible label, as well as tools to verify the source of the content.
If you are not an AI company and are only communicating to U.S. audiences, at this time, you do not have a broad legal obligation to disclose that content is AI generated. There are important exceptions, including ads featuring AI-generated “synthetic performers” in New York and political ads covered by a number of state election deepfake laws.
Make no mistake, the trend is toward disclosure for AI-generated content. Communicators, marketers, and technology companies need to understand their obligations, and anticipate where the puck is headed. This Davis+Gilbert slide deck is a valuable primer on the specific requirements for companies under these laws.
At the breakfast there was a discussion of whether the requirement to disclose AI-generated content makes sense in a world in which AI touches nearly everything. This is to some extent a moot point. The law is the law. Lawmakers have been rushing to pass legislation to address the serious and consequential issues of deepfakes, fraud, and misinformation, which have scaled with the mass adoption of AI tools. Disclosure is a focus.
The consequences of the failure to comply with the EU AI Act are significant. Companies that breach the Article 50 transparency obligations can be fined up to €15 million or 3% of total worldwide annual revenue. Similarly, U.S. AI disclosure laws have teeth – California’s carries civil penalties of $5,000 per violation, per day.
Big tech at this point has settled on several core technologies to address AI disclosure requirements. Two are built into both Google’s Gemini and Anthropic’s Claude: SynthID, which provides a way to identify content generated by AI, and the C2PA content authentication standard, which Tauth Labs is bringing to the communications and other industries. SynthID is a watermarking tool and C2PA provides a human and machine-readable way to cryptographically prove provenance and generate visible labels for AI-generated content (and a lot more). Read Content Authentication: 10 Use Cases.
While the breakfast focused on legal obligations for AI disclosure, we also need to ask what clients or consumers would expect of our organizations to share around the use of AI. This should be factored into compliance, policy, and ethics discussions.
Watermarking and content authentication technologies provide ways to comply with the law and policy decisions to disclose AI use. C2PA authentication can, for example, be used to incorporate machine-readable disclosures without an “AI GENERATED” label, where that is not an obligation. Disclosures can be read using what we call a Trust Shield built into content to provide a means to verify authenticity and review disclosures. Verification tools can also be used for this purpose.
There is an important final point to consider that was touched on in our panel discussion. Will legal, ethical, or policy-driven requirements to disclose that content was generated all or in part by AI change how we use the technology?
If disclosure, and the ability to verify that content is from you or your company, pushes AI content creators to take additional steps to fact check and be more accountable for the content that is produced, in my book, that is a good thing. And, if it creates a higher barrier for fraud and misinformation, we should welcome it.
Trust in the real world is based on disclosure. It is lost when we don’t share relevant information. This applies as much to a secondhand car, as it does to an op-ed developed for The Wall Street Journal. Using content authentication to provide cryptographic proof of its source and how it was generated is arguably the right thing to do. It is reasonable to debate whether this needs to be done explicitly through a visible label or by building this into content in ways that may be less intrusive.
It may be tempting to see AI disclosure requirements as legal over-reach and perhaps even non-sensical, but done right, disclosure can serve as a trust anchor in the AI world.
