AI Label Checker

Social platforms decide to attach an AI label in up to three ways. The first is file metadata: structured data written into the file by the tool that created it. The important standards are C2PA Content Credentials — a signed provenance manifest — and the IPTC DigitalSourceType field, whose value trainedAlgorithmicMedia literally means “created with AI”. Meta announced in February 2024 that it reads exactly these indicators; TikTok began reading Content Credentials in May 2024; YouTube auto-labels C2PA-carrying uploads; Pinterest reads the IPTC field. This is the layer this checker inspects, field by field.

The second layer is invisible watermarking: signals such as Google's SynthID are woven into the pixels or audio itself and survive cropping, filters and recompression. They are not stored in metadata, so no metadata reader — including this one — can see them; even Google routes detection through its own detector. The third layer is the platform's own AI classifiers, which judge the content directly; Pinterest states openly that its classifiers detect generative AI “even if the content doesn't have obvious markers”. A complete picture needs all three layers — this tool gives you the first one, honestly labelled as such.

A creator generates an image with Google Gemini and wants to post it to Instagram and TikTok. Dropping the file here produces three findings: a C2PA Content Credentials manifest is present, the IPTC DigitalSourceType reads trainedAlgorithmicMedia, and the XMP CreatorTool names the generating tool. The verdicts follow directly: Meta — very likely labelled, because Meta reads C2PA and IPTC markers and applies its “AI info” label; TikTok — very likely labelled “AI-generated”, because it has read Content Credentials since 2024; YouTube — label locked in, because YouTube says labels arising from C2PA content cannot be adjusted in Studio; Pinterest — very likely labelled “AI modified”, because the IPTC field it reads is present.

Now the opposite case: a straight iPhone photo. The checker finds camera make and model, a date — and GPS coordinates. No AI markers at all, so no metadata-driven label is likely on any platform; but the GPS finding matters for a different reason, because posting it would publish the photographer's home location. That is exactly what the Photo Privacy Cleaner exists to remove. Two files, two completely different risks — both visible in seconds, before anything is posted.

Search results for this topic are full of tools promising to strip markers so posts pass unlabelled. Three facts from the platforms themselves are worth knowing before trusting any of them. First, TikTok states that an applied label “cannot be edited or removed after posting”, and its Community Guidelines prohibit “automation tools, scripts, or other tricks designed to bypass our systems”, with removal or bans as consequences. Second, YouTube documents that labels based on C2PA metadata are among the cases that “can not be adjusted” — deleting and re-uploading the same pixels changes nothing once its classifiers have seen the content. Third, metadata is only one of three detection layers: a cleaned file can still be labelled by SynthID-style watermark detection or by classifiers, which is why this checker never equates “metadata-clean” with “safe”.

There are also legitimate reasons to inspect a file that have nothing to do with hiding anything: a photographer whose real, film-shot photo was labelled “Made with AI” needs to know which marker triggered it before appealing; a social-media manager needs to confirm a client's asset carries the disclosure data their compliance policy requires; a parent wants to know whether a photo carries GPS before it goes on a school account. Reading your own file's metadata is always legitimate. Acting on what you find — disclosing honestly — is the part this tool is built for, and our AI Disclosure Caption Generator writes the caption and shows the in-app step for each platform.

This checker reads metadata, not pixels: it cannot detect SynthID or any other invisible watermark, and it cannot simulate Meta's, TikTok's, YouTube's or Pinterest's internal classifiers, so verdicts are likelihoods based on the markers present, not guarantees. Compressed PNG text chunks (zTXt) are decompressed where the browser supports it; an unsupported browser may report fewer PNG fields. Very large videos are scanned at the box level rather than decoded, and exotic or corrupted files may yield partial results. Platforms also change their pipelines — TikTok's desktop Studio uploads, for example, behaved differently from in-app uploads in one published test — so re-check a file if a platform updates its labelling system.