You restyle a product shot, it looks great, you post it. Three weeks later a platform slaps an "AI info" tag on it that you didn't add and can't remove, a commenter accuses you of hiding something, and you realise you have no record of which of your 400 assets were generated, edited, or shot on a phone. Nobody was dishonest. There just wasn't a system.
The takeaway up front: labelling isn't one action, it's four layers — embedded provenance metadata, an invisible watermark, the platform's own disclosure control, and a human-readable note — and they fail in different ways, so a workable practice uses more than one. The metadata gets stripped, the watermark survives edits but is invisible to your audience, the platform toggle only works on that platform, and the caption is the only layer a person actually reads.
Why this stopped being optional
Three separate pressures converged, and none of them are going away.
Platform rules. Every major social platform now has a disclosure requirement for realistic synthetic media, and enforcement is a mix of automatic detection and self-declaration. The platforms are reading embedded provenance signals where they exist and applying labels themselves; where they don't exist, they expect you to declare.
Regulation. In the EU, the AI Act carries transparency obligations for generative systems: providers must mark synthetic output in machine-readable form, and those deploying deepfake-style content must disclose it. On the current timetable those provisions apply from August 2026, which is why tools have been racing to build marking in rather than bolt it on. Elsewhere the picture is patchier — the US has no single federal labelling statute, but rules against deceptive advertising still apply, and state laws increasingly cover political deepfakes and digital replicas.
Audience trust. This is the one that actually affects your work. Self-disclosure reads as craft confidence; a system-applied tag on an undeclared image reads as being caught.
Layer 1: Content Credentials (C2PA)
The industry standard here is C2PA — the Coalition for Content Provenance and Authenticity, a cross-industry body whose backers include Adobe, Microsoft, the BBC, Intel, Arm, and Truepic. The user-facing name for its output is Content Credentials.
What it does: attaches a cryptographically signed manifest to the file describing how it was made — which tool, whether AI generation was involved, what edits followed. Anyone can inspect it with a Content Credentials viewer.
Where you'll meet it: several major generators attach C2PA data automatically. OpenAI's image outputs carry C2PA metadata, Adobe's tools write Content Credentials through the editing chain, and camera makers have shipped capture-side support.
Its weakness is brutal. Metadata is a passenger, not a tattoo. Many platforms strip it on upload for size and privacy reasons; screenshotting destroys it entirely; re-encoding through a random online converter usually does too. A file can be genuinely AI-generated, correctly labelled at birth, and arrive at a viewer completely bare. So: attach it, don't rely on it alone.
Layer 2: Invisible watermarks
The complement to metadata is a watermark encoded into the pixels themselves rather than the file wrapper. Google's SynthID is the best-known example, applied to images and video from its own generative models, with a detection portal for checking.
The trade-off mirrors metadata's. A pixel-level watermark survives screenshots, crops, compression, and format changes far better than a metadata block — but it's invisible to your audience, readable only by the matching detector, and tied to the vendor that applied it. It's a forensic tool, not a disclosure to a reader.
You mostly don't apply these yourself; you inherit them from whichever generator you used. Practical consequence: assume outputs from major generators may carry a detectable invisible mark whether or not the metadata survives. Publishing AI work as if it were photography is a losing bet on a timescale of months. The direction of travel is combining all three signals — signed metadata, watermark, and content fingerprint — so that when one is stripped the others still resolve.
Layer 3: The platform's own control
Every major platform has a switch, and using it is the cheapest insurance available.
- Meta platforms apply an "AI info" label based on industry-standard signals in the file, and ask creators to disclose realistic AI-generated content themselves when those signals are absent.
- TikTok reads Content Credentials and auto-labels content that carries them, alongside a creator-facing toggle for AI-generated content.
- YouTube requires creators to declare realistic altered or synthetic content in the upload flow, and surfaces that as a label on the video or in its description.
Two notes. The toggle only labels the copy on that platform — it does nothing for the same asset on your site or in a client deck. And declaring voluntarily beats being detected, because platforms treat undeclared realistic synthetic media as an integrity issue rather than a formatting mistake.
Layer 4: The human-readable line
The only layer your audience actually reads. It costs one sentence and it does the most work.
Good disclosure is specific about what the AI did:
- "Concept image generated with AI." — clear, honest, sets expectations.
- "Product photo restyled with AI; the product itself is unretouched." — precise, and defuses the obvious worry.
- "Illustration: AI-generated. Copy: written by me, edited from an AI draft." — separates the two, which readers care about more than creators expect.
Weak disclosure hides in a footer at six-point grey, or uses vagueness like "made with modern tools." If a reasonable person would feel misled after learning how the image was made, the label failed regardless of what the metadata says.
Put it in the caption or immediately adjacent to the asset for social, and in the image alt text for web — alt text travels with the image in more contexts than a caption does.
The one rule about photographs of real things
There is a hard line and it's worth stating separately: do not present a restyled or generated image as documentary evidence of something that happened. Product shots, listings, before-and-after images, news, testimonials, property photos — in these contexts an AI-restyled image isn't a stylistic choice, it's a claim about reality.
This isn't primarily an ethics lecture; it's where the actual legal exposure sits. An AI-enhanced product image that shows a product looking materially different from what ships is a deceptive advertising problem long before it's an AI-labelling problem. Restyling backgrounds, lighting, and composition is normal creative work; changing the item itself is not.
The same principle governs style fusion work: transform the look, preserve the truth of the subject.
A workflow you can actually keep
Labelling fails when it's a final-step chore. Build it into the pipeline — the same argument as in the end-to-end AI creative workflow.
- Name files at generation time. A prefix like
aigen_oraiedit_on every generated or heavily AI-edited asset. Trivial, and it makes your file browser a provenance record. - Keep a one-line log per published asset: date, tool, prompt or recipe, what the AI actually did, and any source image you supplied. A spreadsheet is fine.
- Check what your generator attaches. View your own outputs in a Content Credentials viewer once, so you know which layers you get for free.
- Export without stripping. Some export settings and third-party converters drop metadata silently — the creative utility-belt roundup is exactly the kind of ad-hoc tooling that discards provenance data. Verify once per tool in your chain.
- Set the platform toggle at upload, every time, for realistic imagery.
- Write the caption line, specific about what the AI did.
- Store the original. If a claim ever needs answering, the unedited source plus the log is the answer.
Total added cost after setup: under a minute per asset.
Where FusionZap fits
FusionZap is pre-launch, so this is what's being built rather than a live feature list: the fusion generator is planned to label its outputs as AI-generated as part of the pipeline, so the provenance layer is produced by the same run that produces the image. Combined with the fusion recipes, that means the record of which two inputs merged into this output exists by default rather than living in someone's memory.
The layers you add yourself — the platform toggle and the caption — remain yours. No tool can write an honest sentence about your own work for you.
FAQ
Do I have to label AI-assisted work if the AI only helped a bit? The practical test is whether a reasonable viewer would feel misled on learning how it was made. AI cleaning up your grammar generally doesn't need a label; an AI-generated image presented as a photograph does. When it's genuinely borderline, disclosing costs you nothing.
Can I remove Content Credentials from my own images? Technically yes — metadata is removable, and it happens accidentally all the time through uploads and conversions. But deliberately stripping provenance from realistic synthetic media is the behaviour platform integrity systems and the EU transparency rules are aimed at, and invisible watermarking may persist regardless.
Does this apply to AI-written text too? Text has no equivalent embedded-provenance standard in wide deployment, and detection tools for writing are unreliable enough that some institutions have stopped using them. Disclosure for copy is therefore mostly a matter of platform policy, client agreements, and your own editorial standards.
What if my client asks me not to label? Separate the reasons. Not wanting a label on a stylised concept illustration is a taste preference. Not wanting one on a realistic product or property image is a request to mislead a buyer, which is a risk you'd be carrying alongside them. Say so plainly and put the recommendation in writing.
Build it in, don't bolt it on
Pick your prefix, start the log, check what your generator attaches, and write one honest caption sentence per asset. That's the whole practice, and it takes an afternoon to set up.
If you want the provenance layer produced by the same step that produces the work, see what FusionZap is building — labelled outputs are part of the planned fusion pipeline rather than a box you have to remember to tick.
