Contents
- What you need before you check an image
- OpenAI: C2PA metadata, plus a second layer
- Google DeepMind: SynthID exists, but you can't check it yet
- Meta and Instagram: on-platform labels, plus a new crackdown
- Adobe: Content Credentials and the "screenshot-proof" claim
- Microsoft Paint and Photos: the watermark nobody announced
- Where every one of these checks breaks down
- What to do when you can't verify an image
- FAQ
OpenAI, Google, Meta and Adobe say they label AI-generated images. Microsoft’s images carry a watermark too, but one the company never announced. None of them do it the same way, and none of the methods are foolproof. We checked all 11 of the official pages and tools a reader would actually need, live, today. 8 loaded normally. 3 returned an error. Here is what each platform actually does, and how to check an image yourself.
What you need before you check an image
You don’t need special software. A browser and the image file itself are enough for most of this.
Before you start: – The image file itself — a screenshot of an image often strips the very data you’re trying to find – A browser, for the Content Credentials Verify tool at contentcredentials.org/verify – Patience for one caveat: none of these tools can find a watermark that was never embedded in the first place
We ran a simple test before writing a word of this piece. From one machine, we hit all 11 official pages and tools tied to AI-image labeling with a direct automated request — the same kind of check a reader’s own script or link-checker would make.
| # | Page / tool | What it’s for | HTTP status |
|---|---|---|---|
| 1 | C2PA Verify tool (contentcredentials.org/verify) | Upload any file, see embedded credentials | 200 |
| 2 | C2PA.org | Spec and list of adopters | 200 |
| 3 | OpenAI Help Center, C2PA in DALL·E 3 | Official metadata explainer | 403 |
| 4 | OpenAI blog, provenance policy post | Official OpenAI policy | 403 |
| 5 | Google DeepMind SynthID page | Explains SynthID | 200 |
| 6 | Meta Newsroom, Feb 2024 labeling post | Official Meta policy | 200 |
| 7 | Adobe Content Authenticity “Inspect” tool | Adobe’s own verification tool | 200 |
| 8 | Adobe Content Authenticity home | Landing page | 200 |
| 9 | Adobe Helpx, Content Credentials overview | Official Adobe explainer | 403 |
| 10 | Instagram Help Center | General help portal | 200 |
| 11 | Instagram/Meta creator blog | Aug 2026 AI-profile-label policy | 200 |
Checked 2026-09-20, 20:45 KST, from a South Korea network. The three 403s — on openai.com, help.openai.com and helpx.adobe.com — look like bot or geo protection, not an outage. A signed-in person using a normal browser likely sees these pages render fine. Still, it means a reader trying to pull up OpenAI’s own explainer with a script, a link-checker, or certain privacy browsers may hit a wall where Google, Meta and Adobe’s core pages don’t.
OpenAI: C2PA metadata, plus a second layer
OpenAI adds C2PA metadata to images made or edited with DALL·E 3 in ChatGPT and through the API, and says it does the same for Sora video. As a second, harder-to-strip layer, OpenAI has built in Google DeepMind’s SynthID watermarking for images generated through ChatGPT, Codex, or the API — an actual technical partnership between the two companies, not just a policy statement.
The metadata does add to file size. API-generated PNGs grow by roughly 3%. API WebP files grow by roughly 5%. ChatGPT-generated WebP files grow by roughly 32%.
To check an OpenAI image, run it through contentcredentials.org/verify — it reads the C2PA manifest or SynthID signal directly. One catch: both help.openai.com and openai.com returned that 403 in our live check today, so if you go looking for OpenAI’s own explanation of how this works, you may need a regular browser session rather than a direct link.
OpenAI’s own material reportedly says the metadata “can be sometimes removed by platforms, editing tools, or file conversions.” That’s secondhand — pulled from search results after openai.com blocked our direct request — so treat it as likely accurate but unconfirmed.
Google DeepMind: SynthID exists, but you can’t check it yet
SynthID embeds imperceptible watermarks across 4 content types: images, video, audio and text. Google says the image and video versions are designed to survive cropping, filters, frame-rate changes and lossy compression. The audio version is described as inaudible and resistant to noise, MP3 compression and speed changes. Text watermarking works differently — it nudges token-selection probability during generation rather than touching pixels or sound.
SynthID is built into the Gemini app, where you can upload content and ask whether it’s AI-generated. It also runs inside Lyria for music and NotebookLM for podcasts.
Here’s the gap. A separate “SynthID Detector” portal exists for more thorough checking, but it is in early testing and gated behind a waitlist limited to journalists and media professionals. There is no open, public tool where an ordinary reader can drop in an arbitrary image and get a SynthID answer today.
Meta and Instagram: on-platform labels, plus a new crackdown
Meta’s February 2024 policy, announced by Nick Clegg, covers three cases. Content made with Meta’s own AI generator gets an “Imagined with AI” label. Content carrying detectable industry-standard metadata — C2PA or IPTC — gets labeled automatically. Users posting AI-generated video or audio are required to self-disclose with a label.
Meta was blunt about the limits in that same post:
“It’s not yet possible to identify all AI-generated content, and there are ways that people can strip out invisible markers.”
At the time, Meta said it was developing its own watermarking technology, called Stable Signature, aimed at making markers harder to remove.
There’s a newer, sharper policy on top of that one. As of August 31, 2026, Instagram requires an “AI-generated profile” label on accounts built around AI personas, replacing an earlier, optional “AI creator” label. Accounts that skip the label face reduced reach. A platform spokesperson framed the reasoning this way: “We’ve heard from our community that they don’t like coming across a profile that seems human, only to find out later that the person featured is AI-generated.” The policy followed reporting by The Guardian and The New York Times on fake AI influencers, some falsely claiming medical credentials while pushing supplements. Accounts that believe they were penalized wrongly can appeal through Instagram’s Account Status dashboard.
Neither Meta nor Instagram appears to offer a standalone reader-facing checking tool. Labels show up directly on the post or profile instead.
Adobe: Content Credentials and the “screenshot-proof” claim
Adobe’s Content Credentials system is its C2PA-based provenance layer. The “Inspect” tool at contentauthenticity.adobe.com/inspect lets anyone upload a file and see what’s embedded.
Adobe describes a 3-part model meant to be durable: cryptographically signed metadata, an invisible watermark using Adobe’s own TrustMark algorithm, and a digital fingerprint backed up in its cloud — built so any one surviving layer can restore the others. The company’s own marketing language calls the result “screenshot-proof.” Worth sitting with that phrase for a second: the fact that Adobe needed to add a watermark and a cloud-backed fingerprint on top of plain metadata suggests the metadata alone doesn’t survive a screenshot.
One sourcing note: helpx.adobe.com 403’d in our live check, and contentauthenticity.adobe.com itself returned an empty, JavaScript-rendered body. We pulled that marketing language from search snippets instead of a clean direct read — treat it as secondhand until someone rechecks it live.
Microsoft Paint and Photos: the watermark nobody announced
This one doesn’t fit the pattern of the other four. Microsoft never announced it as a feature. It was found through independent reverse engineering by a researcher named Xusheng Li, and as of an August 25, 2026 report from The Register, Microsoft had not confirmed the mechanism publicly.
The technical picture, from Li’s write-up: on a 512×512 test image, 193,376 of the image’s 262,144 pixels changed — about 73.8% of the frame — to embed one 16-byte GUID. That GUID is also written into the image’s C2PA manifest as a “soft binding,” under an algorithm Li’s post labels com.microsoft.invismark.1. If the watermarking step fails, Paint treats the whole generation as a failure and returns nothing; Photos, by contrast, logs the error but still hands back the image.
The Register’s reporting adds a linkability wrinkle: Paint resends the prior promptGenerationId with each new request, so a run of images from one session can, in principle, be tied together on Microsoft’s server. The outlet frames the discovery against the EU Code of Practice on Transparency of AI-generated Content, which requires signatories to mark AI content in machine-readable form — and notes Microsoft’s approach appears to go further than that rule requires. Microsoft did not immediately respond to The Register’s request for comment.
This is not an opt-in label a user can toggle. It’s baked into the local generation pipeline.
Where every one of these checks breaks down
Every platform’s own material admits, in one way or another, that its labeling method is not reliable end to end.
| Platform | What its own material admits |
|---|---|
| OpenAI | Metadata reportedly removable by platforms, editing tools, or file conversions (quote sourced secondhand) |
| Meta | Labels depend on detecting known markers; markers can be stripped |
| Adobe | Added a watermark and cloud backup specifically because plain metadata alone isn’t screenshot-proof |
| SynthID checking tool isn’t publicly available at all yet |
I’d put the detection-dependency problem at the top of that list. Meta’s own wording only labels content “when we can detect industry standard indicators” — a failure to detect, not just a deliberate strip, also results in no label, and that quieter failure mode probably affects more images than deliberate removal does.
None of this means the tools are useless. It means the absence of a label proves nothing — an unlabeled image could simply have had its markers stripped, or never had detectable ones to begin with. Presence of a label is stronger evidence, but only for the platforms a reader can actually check against right now: OpenAI and Adobe through contentcredentials.org, and C2PA adopters generally. Google’s SynthID and Meta’s on-platform labels don’t offer that same reader-facing lookup today.
My honest read: the gap between “we watermark our images” as a headline and “here’s a public tool that proves it” as a follow-through is wider at some of these companies than others. Adobe and OpenAI both point you to a working checker. Google and Meta, for different reasons, currently don’t.
What to do when you can’t verify an image
Start with contentcredentials.org/verify. It reads C2PA manifests from OpenAI, Adobe and other C2PA adopters, and it’s the closest thing to a universal tool among the five platforms covered here.
If that comes back empty, that alone doesn’t settle anything — the image might be AI-generated with its metadata stripped, or made with a tool that never added metadata in the first place. For Google-origin content, there’s currently no open public checker at all, so a manual judgment call is often the only option. For Meta and Instagram content, look at the label on the post or profile itself rather than searching for a separate verification step — none seems to exist yet.
If you’re trying to judge a whole account rather than a single image, on Instagram specifically, the “AI-generated profile” label (or its absence, post-August 2026) is now the signal the platform itself is pushing account owners toward.
FAQ
Does every AI-generated image get labeled automatically? No. Labeling depends on the platform, the tool used, and in Meta’s case, on whether the platform can actually detect a marker — its own material says detection is a precondition, and markers can reportedly be removed or fail to survive editing and file conversion.
Can I check a Gemini-made image for a SynthID watermark myself? Not through an open public tool as of this check. Google’s SynthID Detector exists but is waitlist-gated for journalists and media professionals; ordinary readers can ask the Gemini app directly, but there’s no standalone verification page to run an arbitrary file through.
Is the Microsoft Paint watermark something I can turn off? No word on that. What’s clear is that it’s built into the local generation pipeline rather than offered as an optional label — not something you toggle from a settings menu.
What’s still open: whether Microsoft ever comments on the Paint and Photos findings, whether Google opens the SynthID Detector beyond its current waitlist, and whether those three 403s we hit today are just a bot wall or something a bit more deliberate. Worth another look in a few months.
Image credits
- Content credentials logo by Coalition for Content Provenance and Authenticity (C2PA), licensed under PUBLIC DOMAIN (via Openverse)
- Photo by Andrew Neel on Pexels
- Homepage of GOOGLE DEEPMIND Website magnified on logo with magnifying glass (53147449198) by Jernej Furman from Slovenia, licensed under BY (via Openverse)