AI Transparency Law for Creatives: The Label That Doesn’t Survive a Repost
On August 2, 2026, California started requiring AI generated images to carry an invisible label proving where they came from. The most popular AI image generator on the planet still doesn’t have one. That’s the AI transparency law for creatives on day one: real fines, a hard deadline, and a hole in the middle big enough to drive a fabric cart through.
What changed on August 2
California’s AI Transparency Act, Senate Bill 942 as amended by Assembly Bill 853, became operative on August 2. Any AI image, video, or audio system with more than a million monthly California users now has to do three things: offer a free public tool anyone can use to check whether a piece of content came from their system, let people add a visible “AI generated” label if they choose to, and automatically embed machine readable data in every output showing who made it, what system made it, and when.
Fines start at five thousand dollars per violation per day. The bigger change is who can bring the case. City attorneys and county counsel can now enforce this, not just the state attorney general. That’s new for AI regulation in this country, and it means a lot more people with standing to sue.
The timing wasn’t an accident. The European Union’s AI Act hit full enforcement on Article 50, a similar disclosure requirement, the same day, across all 27 member states. Two of the largest regulatory bodies in the world picked the same date.
Who's covered, and who's already behind
OpenAI and Google both built compliant tools ahead of the deadline. Midjourney didn’t. It joined the Content Authenticity Initiative, the standards group whose rulebook this law borrows from, back in 2023. Three years to prepare, and as of enforcement day it still ships no watermark and no provenance data. One of the most widely used image generators in the country, out of compliance from hour one.
That split matters for the platforms themselves, not just for users. For OpenAI and Google, compliance is already a selling point, proof to worried customers and regulators that they built this before anyone forced them to. For Midjourney, every day of noncompliance is now a live legal exposure, especially with city and county attorneys able to bring a case on their own.
This only applies to tools with more than a million monthly California users. Most of the smaller, niche AI tools you use for color work or pattern generation fall outside the law entirely. This was written for the giants, and the giants split into two camps: the ones who built the thing, and the one who joined the club and skipped the homework.
The lineup, tool by tool
Here’s what each major generator offers right now for checking its own output.
| AI Tool | Company | Content Type | Free Public Detection Tool | Note |
|---|---|---|---|---|
| ChatGPT / Sora | OpenAI | Image | openai.com/verify | Launched May 2026, checks for both C2PA and SynthID |
| Gemini | Image, video, audio | Ask Gemini directly in the app | Dedicated detector portal is waitlist-only | |
| Copilot | Microsoft | Image, video, audio | contentcredentials.org/verify | Video and audio watermarks are opt-in, not default |
| Firefly | Adobe | Image, video | contentauthenticity.adobe.com/inspect | Â |
| Eleven Labs | ElevenLabs | Audio | elevenlabs.io/ai-speech-classifier | Free, no sign-in |
| Flux | Black Forest Labs | Image | None confirmed | Â |
| HeyGen | HeyGen | Video (avatar) | None confirmed | No first-party tool found |
| Midjourney | Midjourney, Inc. | Image | None | Still noncompliant as of this writing |
A few things stand out.
- Gemini’s own dedicated SynthID Detector, the one that would let anyone check the watermark directly, is limited to journalists, researchers, and media professionals on a waitlist. Everyone else gets pointed to the Gemini app and a typed question instead. Nobody’s explained the reasoning in plain terms, but the likely one isn’t complicated. Give the public a detector with no limits, and you’ve also handed out a free testing ground for figuring out which edits make the watermark disappear. Keep the guest list short and that kind of probing gets a lot harder. Curious choice, though, for a law that’s supposed to guarantee free public access.
- Copilot’s video and audio watermarks have to be switched on by an admin or user. The underlying metadata gets attached either way, but the visible label most people would notice doesn’t show up unless someone flips it on.
- Flux and HeyGen: I couldn’t find a first-party public detection tool for either, as of this writing. Both embed C2PA-style metadata, but neither runs its own consumer verify page the way OpenAI, Adobe, and ElevenLabs do.
What this does for creatives who feel like AI stole from them
Not much, and I think it’s important to be clear about that. This law is about labeling what comes out of these systems. It has nothing to do with what went into them, so it doesn’t touch the separate fight over whether these companies trained on artists’ work without consent or payment. That argument is playing out in copyright lawsuits, not in transparency statutes. What this law gives you is a slightly better way to tell, after the fact, whether a specific image came from one of these systems. That’s one small piece of evidence in the “did they copy my style” conversation, not an answer to it.
What this does for creatives who use AI
This is the part that matters more for most of you, because most of you aren’t fighting AI companies. You’re using their tools the way I use mine: as a reference, the same way I’d use a photograph for one of my landscape quilts. Nobody accuses me of stealing from a sunset.
Here’s what might help. You now have an actual legal backdrop for the disclosure habit some of you already practice. If you tell people “I used AI to generate a color reference, then I quilted the piece myself,” a law just showed up agreeing that disclosure is the right call. Once labeling becomes normal on the platforms people use every day, saying you used an AI reference stops sounding like a confession and starts sounding like a process note, the same as saying your background fabric came off a Cricut.
Here’s what doesn’t help. This law labels generated images. It does not label finished quilts. If you build a mockup with AI and then hand quilt the actual piece, nothing about this law touches your finished work, and nothing about it proves which parts were AI assisted and which weren’t. That’s still yours to say, in your own words, same as it’s always been.
It also hands people a new way to get the wrong idea. Someone might expect a clean label or a clear detection result and treat its absence as proof a piece is handmade. It isn’t proof of anything. If you’re building your own marketing on the presence or absence of a watermark, you’re standing on ground that isn’t there. And if the AI tool you use never hits a million monthly California users, none of this applies to it at all. Regulation isn’t vetting your tools for you. That’s still your job.
How platform stripping breaks the AI transparency law for creatives
That invisible, machine readable label gets embedded the moment an image is created. It does not survive the trip through social media.
Upload that image to Instagram. Repost it to X. Send it through WhatsApp. The metadata gets stripped somewhere in that chain, every time, and current plans don’t fix this until 2027. The law told every covered platform to label its work, and most of them did, right up until Instagram got its hands on it. Technically followed. Practically pointless. Very Amelia Bedelia of the whole system.
So a label existing at the moment of creation tells you that one specific export, on that one specific day, was labeled. It tells you nothing about the copy you find three reposts later on somebody else’s feed. A label’s presence, once an image has traveled through social platforms, tells you almost nothing. Its absence tells you even less. It might mean the tool is noncompliant, like Midjourney. It might mean the platform already stripped it. It might mean a person made it with a camera. From the outside, you can’t tell which.
If you want to see the limits of this for yourself, OpenAI’s detection tool is free and doesn’t require an account, at openai.com/verify. Upload an image and pay attention to what it doesn’t tell you as much as what it does. It only checks OpenAI’s own outputs, and even when it confirms something is AI made, it never tells you who ran the prompt. It’s Catfish for pictures: it can confirm the thing is fake, but it will not tell you who’s been messaging you.
Where this connects to what I already teach
This law only covers the exit: whether the finished image carries a label. My Conscious Creative’s Guide: AI Tools That Won’t Steal Your Soul is about the entrance: whether the tool was built on training data anyone consented to, and whether there’s a real ethics policy behind it instead of a marketing page. Vet the tool going in. Don’t count on the label coming out.
Sources: California AI Transparency Act, Senate Bill 942 as amended by Assembly Bill 853, California Legislative Information. Additional coverage from Tech Times, August 2, 2026. Tool details from OpenAI’s verification page, Google’s SynthID Detector announcement, Adobe’s Content Authenticity inspection tool, ElevenLabs’ AI Speech Classifier, and Microsoft’s Content Credentials documentation.