Remove the AI Label Before You Post

A label is a decision made from your file

When a platform marks an image as AI, it has usually not looked at the picture and judged it. It has read a statement inside the file — written by the app that generated or edited it — and acted on that statement during upload.

Two kinds of statement do almost all of the work. A C2PA manifest, also called Content Credentials, is a signed record of which tools made and changed the image. An IPTC digital source type is a plain field in the file's XMP whose values include one meaning "created by a trained algorithm". Either can be enough.

Some platforms add a second layer that reads pixels instead, and some rely on you to declare AI use yourself. Which layer a platform uses decides whether cleaning the file before upload helps at all.

What each platform reads

Each platform below has its own page with the details; this is the short version.

• Instagram, Facebook and Threads — Meta reads both C2PA and IPTC AI metadata and shows an "AI info" label. With neither in the file, there is nothing to read.
• TikTok — reads Content Credentials on uploaded images and videos, and separately requires creators to label realistic AI content of people and scenes.
• LinkedIn — shows a CR badge on any image carrying Content Credentials, and the badge names the tool and who signed the record.
• Pinterest — shows an "AI modified" label using image metadata and its own classifiers, which work even on files with no metadata at all.
• X — has labelled some AI images "Made with AI" automatically since early 2026, but has not published what triggers it. A published test found a ChatGPT image labelled while Gemini and Grok images were not.
• YouTube — relies on creators disclosing realistic altered or synthetic content. Nothing in an uploaded file is read for that rule.

Where the trigger sits inside an image

The C2PA manifest lives in a different place in each format: an APP11 segment in a JPEG, a caBX chunk in a PNG, a C2PA chunk in a WebP, and a uuid box in a HEIC. The IPTC source type sits in the XMP packet, which every one of those formats can carry. A generator name in the EXIF Software tag is a third, weaker signal.

The tool above finds all three, shows you which are present, removes them and then re-reads the cleaned file to confirm none remain. Nothing is uploaded; the file is read and rewritten in your browser.

What cleaning cannot change

Metadata removal deals with metadata. Three things sit outside it, and no tool that leaves your pixels untouched can affect them.

• Invisible watermarks embedded in the pixels, such as Google's SynthID or the watermark TikTok adds to AI content. They are designed to survive editing and re-saving.
• Classifiers that judge the picture itself, like Pinterest's. A cleaned file looks exactly as the original did, so it scores exactly as the original did.
• Disclosure rules. Meta requires you to label photorealistic AI video and realistic AI audio, TikTok requires labels on realistic AI content, and X has announced penalties for creators who post undisclosed AI images of armed conflict. Those obligations attach to the content, not the file.

When removing it is the right call — and when it is not yours to remove

The label is meant to tell viewers the truth about an image. When it does not — a photograph you took, labelled machine-made because of one retouching stroke — wanting it gone is reasonable.

Check whose metadata it is first. Several generators' terms prohibit removing provenance from their output: Adobe's (Firefly, and Photoshop and Lightroom's generative features), Microsoft's Image Creator, Canva's AI features, xAI's Grok and Black Forest Labs' FLUX API. For those, the honest fix is at the source — retouch without generative tools, or disclose — and this site's page for each generator says what it allows.

When the image was generated, the label is accurate. Removing the metadata does not make it a photograph, and presenting it as one may break the platform's rules and mislead people relying on the label.