For three years, the only question that really mattered about AI content was whether anyone could tell. That question closed on 2 August 2026, and AI content disclosure stopped being a matter of taste and became a matter of law.
On that date, Article 50 of the EU AI Act came into force and California’s SB 942 took effect alongside it. Nine days later, Anthropic confirmed that Claude models released on or after 2 August embed an imperceptible, machine-readable signal into the text they produce. Three days after that, Google made its visible watermark optional across its generative models, while leaving the invisible SynthID mark and C2PA metadata firmly in place.
Read those three moves together and the new rules take an obvious shape. The label a human can see is becoming a choice. The one a machine can read is becoming infrastructure.
The old calculation was about detection. You used AI, you cleaned up the output, and you asked yourself whether it would pass. Plenty of marketers made peace with that arrangement.
That arrangement no longer exists. Anthropic’s watermark is woven into the model’s output rather than stamped on afterwards, so it travels with copied-and-pasted text and can survive light editing. Google’s SynthID does the same for visual and audio media. The mark is invisible to your reader, but it is legible to any platform, publisher or client running a detector.
So the question is no longer whether anyone will know. It is whether you were the one who told them. Being discovered is a fundamentally different event from disclosing, and the gap between those two outcomes is where your credibility now lives.
Two caveats are worth holding onto. The watermark signals that a model was involved, not that it wrote the whole thing, so running your own draft through Claude for a proofread can trigger it. And heavy rewriting or translation can degrade the signal. Neither caveat is a strategy. Building a content operation around watermark evasion is a bet against every major lab, three regulators and the direction of the entire industry.
Here is the part nobody enjoys. The Interactive Advertising Bureau’s updated AI Transparency and Disclosure Framework, released on 18 August 2026, cites an NYU Stern finding that disclosing generative AI involvement cut an advertisement’s click-through rate by 31.5 percent.
That is not a rounding error. That is a third of your traffic, and it explains why so much of the industry has quietly hoped this conversation would go away.
The IAB’s answer is not to hide the cost but to spend it deliberately. Caroline Gigerovich, the organisation’s VP of AI, put the principle plainly: “not every use of AI needs a label, labeling everything teaches consumers to ignore labels.” The framework argues for targeted disclosure, on the reasoning that a 31.5 percent penalty is worth paying where AI use could genuinely mislead someone, and wasteful where it could not.
There is a second-order effect here that matters more than the headline number. Once labels become common, unlabeled content starts to read as more trustworthy simply because it lacks a label. Researchers call this the implied-truth effect. Disclose everything and you devalue your own signal. Disclose nothing and you eventually get caught by a detector. The value sits in disclosing precisely.
The framework surfaces a perception gap that should stop most marketing teams cold. 82 percent of advertisers believe consumers feel positive about AI in advertising. Only 45 percent of Gen Z and Millennial consumers actually do. That is a 37 point gap between what the industry assumes and what its youngest customers report.
The consumer data underneath is more interesting than a simple backlash story. 76 percent of US adults say distinguishing AI-made from human-made content is extremely or very important to them. Yet 73 percent of Gen Z and Millennial consumers say that knowing a brand used AI would either increase their likelihood of purchase or make no difference at all. Among Gen Z specifically, 34 percent see AI-using brands as creative while 30 percent see them as inauthentic.
Set that against the 31.5 percent click-through drop and you get the real finding. This is a stated versus revealed preference problem. People say they do not mind. Their thumbs say otherwise, at least in a feed where a label is the only new information. Which means the label is not the variable you should be optimising. The relationship the label lands inside is.
The IAB framework is unusually specific, and the specifics will relieve most content marketers.
Disclose: AI-generated images and video produced from prompts, some categories of synthetic voice, synthetic avatars, digital twins of deceased people, digital twins of living people placed in fabricated scenarios, and chatbots that a person could mistake for a human.
Do not disclose: routine post-production, clearly fantastical imagery, authorised synthetic voices of real people, generic synthetic voices, background music and audio enhancement, and text or copy.
Note that last one. Written copy sits outside the disclosure recommendation even as it becomes the most reliably watermarked format in existence. The mark and the label are now separate systems answering to different masters, and you need to plan for both.
The jurisdictional layer sits on top. In the US you can use a standardised sparkle icon or a plain text label. In the EU, Article 50 requires disclosure without mandating a specific icon. California’s SB 942 and New York’s synthetic performer law, live since June 2026, add their own requirements, and South Korea’s revised AI Basic Act arrived earlier this year. If you publish across borders, you are already subject to the strictest of these.
Under the IAB’s August 2026 framework, no. Text and copy sit outside the recommended disclosure categories. That said, copy produced by recent Claude models carries an invisible watermark regardless, so plan for the mark even where no label is required.
Partially, and unreliably. Anthropic acknowledges that heavy rewriting or translation can degrade a statistical text watermark, and that a determined person can erase it. Google’s SynthID is similarly resilient rather than indestructible. Neither is a defensible basis for a content workflow.
Article 50 of the EU AI Act has applied since 2 August 2026 and requires providers and deployers of generative systems to make synthetic output detectable and to inform people when they are interacting with AI. If you run campaigns into EU markets, you are inside its scope regardless of where your business is registered.
The NYU Stern research cited in the IAB framework found a 31.5 percent drop in click-through rate when generative AI involvement was disclosed. This is precisely why the framework recommends labeling selectively, reserving the label for uses that could otherwise mislead.
Almost everyone is going to treat this as a compliance exercise: the minimum label, applied as late as possible, in the smallest available type.
That leaves the more interesting position open. Platforms are already building filters for automated content, and audiences are learning to read provenance the way they learned to read sponsored tags. In that environment, a brand with a published policy about how it uses AI, and more importantly where it deliberately does not, is making a claim its competitors cannot cheaply copy.
The watermark was always coming. Stop asking whether your content will be identified as AI-assisted, and start deciding what you want it to say about you when it is.
Not sure which side of the line your content sits on? Book a free consultation with Social Lady and we will audit your workflow and build you an AI disclosure policy that protects your reach and your credibility.
