Edited is a different question from generated
Many disputed images are real photographs that were changed, not scenes invented from scratch. The changes range from ordinary exposure and color adjustments, through AI-assisted tools such as noise reduction, upscaling and object removal, to generative fill that paints new content into part of the frame. A detector trained to spot fully generated images may score a lightly edited photo as real, so it helps to ask what kind of change you are trying to find before choosing a check.
Metadata that records AI edits
The IPTC Digital Source Type vocabulary has a value for this case: compositeWithTrainedAlgorithmicMedia describes media augmented by a generative model, such as inpainting or outpainting, while trainedAlgorithmicMedia describes a fully generated image. Some editors write these values. Google says photos edited with Magic Editor, Magic Eraser and Zoom Enhance in Google Photos include IPTC metadata indicating generative AI edits, and that it uses IPTC metadata to mark blended images made with non-generative features such as Best Take.
Source: Google Photos: transparency for AI edits (opens in a new tab)
Content Credentials can list the edit
A C2PA manifest can record the actions applied to a file and the earlier images it was built from. Adobe announced that Photoshop automatically attaches Content Credentials when Generative Fill or Generative Expand is used and the result is exported. Choose the file in the checker above to see whether a manifest is present, whether it validated, and whether the file declares AI generation or AI composition. The check runs in your browser and does not upload the image. To read the full list of recorded actions, open the original in a C2PA validator.
Source: Adobe Photoshop feature summary (September 2023) (opens in a new tab)
Compare with the original
The strongest evidence of an edit is the unedited capture next to the edited version. Ask the owner whether one exists, and compare the region in question: shadows, reflections, repeated textures, warped straight lines and lettering are worth a close look, though none of them proves anything on its own. Forensic views such as error level analysis can highlight areas that were saved differently, but they are easy to misread and react to ordinary recompression, so treat them as prompts for questions rather than proof.
AI edit or Photoshop?
For most decisions, the tool matters less than the change. A sky replaced by hand can mislead as much as one replaced by a generative model, and an AI denoise filter changes nothing about what happened. Ask what was changed, why, and whether the change was disclosed. If a publication, school or marketplace has a written policy about AI editing, apply that policy to the documented change rather than to a detector score. Image Evidence does not certify authenticity or identify people.
Can you tell if a photo has been edited with AI?
Sometimes. If the file still carries IPTC or C2PA records written by the editing tool, they can show that a generative edit was applied. If those records were stripped, you will usually need the original capture or the editor's own account. General AI image detectors are less reliable on partial edits than on fully generated images.
Is this image AI or Photoshopped?
The file can sometimes tell you. Content Credentials may record the specific actions, and an AI composition value in IPTC metadata points to a generative edit. Without those, the visual result of a careful manual edit and a generative fill can look the same, and neither a detector nor visual inspection can reliably tell them apart.
Does an AI-enhanced photo count as AI-generated?
There is no single standard. Phones apply machine-learning processing to almost every photo, and tools that sharpen or upscale do not usually invent new content. Generative edits that add or remove objects are a different matter. Labels and policies draw this line differently, so describe what was done to the photo rather than relying on a yes-or-no label.
Source: IPTC Digital Source Type vocabulary (opens in a new tab)