2026-09-06

Why a platform labels your real photo “AI” — and what it is actually reading

You post a photograph you took yourself, on a phone, with your own hands, and a small tag shows up under it: AI info. Or the exact opposite happens — an obviously synthetic image scrolls past with no tag at all. Both are common, and both happen for the same reason. With rare exceptions the platform is not examining your picture. It is reading a few hundred bytes of metadata that some earlier piece of software either wrote into the file or didn't.

That distinction used to be trivia. In 2026 it isn't: the labels stopped being a voluntary experiment and became a legal obligation on both sides of the world, and the number of cameras, phones and editors that write these signals is climbing fast. Here is what is actually inside your file, why a real photograph sometimes trips the tag, and what an ordinary crop or resize does to the whole arrangement.

The label is metadata, not pixel forensics

There are three things a platform can find in an uploaded image, and only one of them involves looking at the image at all.

1. A C2PA manifest, branded as Content Credentials. This is a cryptographically signed record travelling inside the file. The C2PA's own FAQ describes what a manifest can carry: "how the content was created, what tools or processes were used, when and where it was made, and how it has changed over time." Note the second half — a credential is a history, not a verdict. It is also, per the same FAQ, "typically embedded in the asset" but able to be "separated" from it, a detail that turns out to matter enormously.

2. An IPTC field in the XMP packet. Long before Content Credentials, the news industry's metadata body defined a Digital Source Type property, and it added a controlled value for images that came out of a model: trainedAlgorithmicMedia. The IPTC's guidance on the vocabulary is where the recommendation lives, and the IPTC notes that this property is used by Meta, Google, Apple and Pinterest, and that C2PA adopted it too. It is one line of text in a file — that is the entire "detection" for most images.

3. An invisible watermark, or your own answer to a checkbox. Some generators also encode a signal into the pixels themselves so it survives a re-save, and the large platforms ask uploaders to self-declare. Self-declaration is the only one of the three that involves a human.

Three signals go in at the top of a file's life, one process at the bottom wipes all three, and the label at the end is only ever as good as what survived.

Why real photographs started getting tagged

The clearest account of this comes from the company that caused the most of it. Meta's Our Approach to Labeling AI-Generated Content and Manipulated Media announced a "Made with AI" label applied "when we detect industry standard AI image indicators or when people disclose that they're uploading AI-generated content." Photographers immediately found it on pictures they had taken with a camera.

Meta's own update to that post, dated July 1, 2024, explains why, and it is the sentence worth remembering: "labels based on these indicators weren't always aligned with people's expectations," because "some content that included minor modifications using AI, such as retouching tools, included industry standard indicators that were then labeled 'Made with AI'." A generative fill on a distracting sign, an AI-assisted noise reduction, an object eraser — the editor wrote the indicator, and the platform, reading only the indicator, could not tell a two-second cleanup from a prompt-to-image. The label was changed to the vaguer "AI info," clickable for detail. A second update on September 12, 2024 refined it further: where the signals suggest content was only modified or edited by AI tools, the label moves into the post's menu rather than sitting on the post, while content generated by AI keeps the visible tag.

So the answer to "why does my photo say AI?" is usually not that anyone thinks you faked it. It is that a piece of software in your edit chain honestly declared itself, and the declaration is coarser than the truth.

Why there is suddenly much more of this

Two laws came due within six months of each other.

In the EU, Article 50 of the AI Act (Regulation (EU) 2024/1689) applies from 2 August 2026. The European Commission's own FAQ on the transparency obligations states the core duty plainly: "providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, must ensure that AI-generated or manipulated content are marked in a machine-readable format and detectable as artificially generated or manipulated." The same FAQ records a narrow grace period — systems already on the market before 2 August 2026 have until 2 December 2026 for that marking duty — and confirms that content generated before the application date does not have to be labelled retroactively. The Commission has since published guidelines on the Article 50 obligations alongside a Code of Practice on Transparency of AI-generated Content; Meta announced on 28 July 2026 that it is signing it.

One clause in Article 50(2) is worth reading if you edit photos for a living, because it draws the line this whole article sits on: the marking obligation "shall not apply to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data" provided by the user. Cropping, straightening, exposure, denoise, resizing — the ordinary furniture of editing — is not what the marking duty is aimed at. Whether a particular tool falls inside or outside that line is exactly the question the platforms have been getting wrong in both directions.

Korea's clock ran slightly earlier. The Framework Act on the Development of Artificial Intelligence and Establishment of a Trustworthy AI Ecosystem (Act No. 20676, promulgated 21 January 2025) took effect on 22 January 2026, and its Article 31 requires operators to mark generative-AI outputs as AI-generated, with a stricter requirement — a notice or mark users can clearly recognise — for outputs that are hard to tell from reality. The Ministry of Science and ICT's announcement of the enforcement decree (12 November 2025) set out the detail and flagged a grace period of at least a year before penalties.

The signal is fragile, in both directions

More devices are writing credentials than ever. Google's announcement of C2PA Content Credentials on Pixel (10 September 2025) says the Pixel 10 lineup is "the first to have Content Credentials built in across every photo created by Pixel Camera," at C2PA Assurance Level 2, and that Google Photos attaches Content Credentials to JPEGs that already have them when they are edited — with AI tools or with ordinary ones. TikTok, in a post dated 9 May 2024, said it would read Content Credentials to auto-label AI content made elsewhere, and start attaching credentials to TikTok content that "will remain on content when downloaded."

And yet the whole edifice rests on metadata surviving the trip. It frequently doesn't. The C2PA FAQ concedes the point directly — manifests are "typically embedded in the asset" but "can be separated" — which is why the coalition also specifies durable credentials backed by soft bindings such as invisible watermarking or fingerprinting, so a stripped credential can be looked up again rather than lost. Those soft bindings are not universally deployed, so in practice, today: a photo with no label may simply be a photo that went through something that didn't preserve metadata, and the absence tells you nothing at all.

What an ordinary edit does to all of this

Every tool that changes an image's pixels or its format writes a new file. That includes ours. NearIMG's tools — compress, resize, crop, rotate, convert and the rest — run in your browser and re-encode the image there. The file you download is a freshly encoded JPG, PNG or WebP: no EXIF block, no XMP packet, no C2PA manifest. Nothing about the original's metadata survives, because none of it is copied across.

Which is a feature and a liability at the same time, and it is worth knowing which one you are getting:

Good if the metadata was the problem. A photo about to be posted publicly loses its GPS coordinates and capture times on the way through. That is the same mechanism, seen from the useful side.
Bad if the credential was the point. If you are a photographer whose provenance chain matters — a newsroom submission, a stock library, a contest that asks for an unmodified original, an insurance or legal file — do not route that file through any re-encoder, ours included. Keep the original untouched, deliver copies, and ask the recipient what they need before you flatten anything.

One more thing we should say plainly, since this post is about honest declarations. NearIMG has an optional AI upscale mode: it runs a real super-resolution model locally in your browser to enlarge small photos. It writes no marker of any kind into the output, because our encoder writes no metadata at all — so if you publish an AI-upscaled image somewhere that expects a disclosure, the disclosure has to come from you.

And the obvious inversion of this article: no, stripping metadata is not a way to get an AI label removed. Platform rules ask you to disclose synthetic content regardless of what the file says, self-declaration is a separate signal from metadata, watermark-based detection is designed to survive a re-save, and the marking duty in both the EU and Korean regimes sits on whoever generated the content. Removing a truthful signal to dodge a label is a policy violation with a plausible paper trail, not a clever trick.

How to check a photo yourself

On your phone, without uploading anything. Google Photos on Android and iOS shows this: open the photo, tap More, then About, and scroll to "How this was made." Google's help page lists the wordings you may see, including "Media captured with a camera," "Edited with AI tools," "Edited with non-AI tools" for ordinary crops and rotations, and "May have been edited by AI tools" when a third-party app didn't say. It is on the mobile apps, not the web version.
On a computer. The Content Authenticity Initiative publishes an inspector at verify.contentauthenticity.org that reads a file's Content Credentials. We have not verified whether it inspects the file locally in the browser or sends it to a server, so treat it like any other upload and don't feed it anything sensitive.
For the raw XMP field. The IPTC publishes a metadata viewer, and any EXIF or XMP tool will show you whether DigitalSourceType is present and what it says.

If your own photograph got tagged

Work backwards through the edit chain rather than arguing with the label. Which app touched the file last, and does it have a generative feature you used — object removal, generative fill, sky replacement, AI denoise or sharpening, an "enhance" button that is quietly a model? That is almost certainly what wrote the indicator. If the edit genuinely was cosmetic and you would rather not carry the tag, the fix is to redo that step with a non-generative tool, not to launder the file afterwards. And if the tag is right — if there is real generated content in the frame — leave it, or better, say so in the caption where a human will actually read it.

The short version

The AI label on a photo is metadata, not analysis: a signed C2PA manifest, an IPTC DigitalSourceType value, sometimes an invisible watermark, sometimes just the box you ticked. Meta's own July 2024 update explains the false positives — retouching tools wrote industry-standard indicators, and "Made with AI" read them as authorship. The rules got real in 2026: Article 50 of the EU AI Act applies from 2 August 2026 with machine-readable marking required of generators, Korea's Framework Act took effect 22 January 2026, and both carve out ordinary assistive editing. Meanwhile any re-encode — a crop, a convert, a compress, in NearIMG or anywhere else — produces a new file with none of the signals in it, which is why a photo with no label tells you precisely nothing, and why the originals are the ones worth keeping untouched.