Sometimes — but only in one direction. A growing number of AI image generators embed metadata that identifies the tool used to create an image, most commonly through the C2PA Content Credentials standard. If that metadata is present and intact, it can be a strong signal. If it's absent, that tells you almost nothing.
Tools that follow the C2PA standard write a signed manifest into the file's metadata: which application or model generated it, and sometimes a chain of subsequent edits. Some tools also write simpler software or "generator" tags outside the formal C2PA structure. Either way, this is text-based data sitting in the same part of the file as EXIF and XMP.
Because it's just metadata, it survives only as long as no one touches it. A screenshot, a re-save through almost any image editor, an upload to a platform that strips metadata on ingestion, or a one-click metadata-removal tool will erase it completely — while leaving the image itself looking exactly the same. Someone doesn't need any technical skill to make an AI-generated image lose its "AI-generated" label; opening and re-saving it in most apps is enough.
Separately from metadata, some companies embed an invisible signal directly into the pixel data itself — Google's SynthID is the best-known example. This kind of watermark is designed to survive things that destroy metadata: cropping, screenshotting, recompression, format conversion. It requires the company's own detector to read, since it isn't a human-readable tag; it's part of the image data.
You can check exactly what's embedded in a given file — including any C2PA manifest — with AiEraser's View mode, which reads the raw metadata blocks without altering the file.