Is this old photo AI-restored? How to tell, and why labels matter

The photo of your great-grandmother came back from a restoration app looking wonderful. Her eyes are clear, the crease across her cheek is gone, her dress is a soft blue. Twenty years from now, your grandchild will look at it and believe she wore blue. She may have worn brown. Nobody will know, unless someone wrote it down.

Last checked: 3 October 2026. We make The Memory Palace, a small app made by one person, Bram, in Leuven, Belgium. Its free photo restorer labels every result, and this article explains why.

The short answer

You can often see AI restoration in the photo itself: faces sharper and smoother than everything around them, eyes and teeth that look too regular, and invented detail in hair, lace or lettering. You can sometimes read it in the file: a metadata field called IPTC Digital Source Type, or signed Content Credentials (C2PA). A missing label proves nothing, because metadata is easily lost. That is why the best label is the one you write yourself, next to the original scan.

What AI restoration actually does

A classic restoration removes dust and scratches and evens out faded tones, working with what is in the photo. An AI restorer goes further: where information is missing, it fills it in with what is likely. A blurred face is redrawn as a sharp one. A black and white photo gets colours that the model considers plausible. The result often looks better than the original ever did, and that is exactly the problem: part of what you see was never photographed.

This is not a reason to avoid AI restoration. It is a reason to say that it was used.

Signs you can see in the photo

• The face is sharper than the rest. Many restorers rebuild faces with a separate model. The face looks crisp while hands, clothes and the background stay soft.
• Too regular. Eyes of exactly the same size and shine, a full row of even teeth, ears that look alike. Real faces in old photos are rarely that tidy.
• Skin like porcelain. Wrinkles, freckles and the grain of the film disappear.
• Invented detail. Lace, embroidery, hair and lettering on a shop sign can turn into patterns that look right from a distance and make no sense up close.
• Colours that are too even. Every face the same tone, uniforms in a bright modern colour, a sky with no grain at all.
• Compare with another copy. If a cousin has the same photo, lay them side by side. A missing button or a changed collar shows at once.

Signs you can read in the file

Kind of label	What it is	How to check	Can it get lost?
Visible note next to the photo	A line on the page or in the app, such as "Restored and coloured with AI", used for example by the free restorer on thememorypalace.ai	Read it where the photo is shown	Yes, as soon as the photo is saved or shared without the note
IPTC Digital Source Type	A metadata field. trainedAlgorithmicMedia means made with a trained AI model; compositeWithTrainedAlgorithmicMedia means enhanced or corrected with generative AI	ExifTool, or another metadata viewer that shows XMP	Yes, metadata is often stripped by messaging apps, social media and screenshots
Content Credentials (C2PA)	A signed record of how a file was made and edited, an open standard backed by Adobe, Google, Microsoft, the BBC and others	contentcredentials.org/verify	Yes, in the same way as other metadata
Your own caption	A line in the file name, the album or on the back of the print: what was changed, with which tool, and when	Read it	Only if the caption is separated from the photo

In practice: Google Photos shows "Edited with AI tools" and adds Content Credentials when a generative AI edit is saved. MyHeritage places a small palette or wand symbol in the corner of photos it colourised or enhanced. Adobe attaches Content Credentials when qualifying generative AI tools are used in Photoshop, and adds no visible label. Many other restoration tools do not say whether they label the result; our comparison of free restoration tools lists what each one states.

To read the IPTC field with ExifTool, a free program for Windows, Mac and Linux, you can type: exiftool -XMP-iptcExt:DigitalSourceType photo.jpg. If the answer contains trainedAlgorithmicMedia, the file says it was made or reworked by a trained AI model.

Why a label matters for family photos

Old photos become evidence. For a genealogist, a photo is a source: the uniform tells the regiment, the dress tells the decade, the shop sign tells the street. An invented badge or an estimated colour can send someone looking in the wrong place.

The next generation cannot ask. You know which version is the restored one. Your grandchildren will not, unless it says so.

The law is moving the same way. In the EU, Article 50 of the AI Act applies from 2 August 2026. Providers of AI systems that generate or manipulate images must mark the output in a machine-readable format, so that it can be detected as artificially generated or manipulated. Standard editing that does not substantially change the image is an exception. Your family album is not covered by the rule, but the reasoning applies there too.

How to label your own restored photos

1. Keep the original scan. Store it next to the restored version, with the same name plus "original".
2. Say what changed. One line is enough: "Scratches repaired and coloured with AI, October 2026. Colours are an estimate."
3. Write down what you know. If an aunt remembers the dress was dark green, add that. It is worth more than any model's guess.
4. Keep the label with the photo. In a caption in your album or app, not only in the metadata, which is easily lost.

What our own restorer does

The free restorer of The Memory Palace says before you start that the restoration is done with AI, and shows a note under every result: "Restored and coloured with AI. The colours are an estimate, not a memory." The restored file also carries the IPTC Digital Source Type trainedAlgorithmicMedia in its metadata. If you would rather not add colour, there is an option to keep the photo black and white and only repair it. You can try it with one photo, free and without an account, on the restore page. Want to compare it with other tools first? See our overview of free old photo restoration tools.

Sources

• Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations: eur-lex.europa.eu, readable version at artificialintelligenceact.eu/article/50
• IPTC Digital Source Type vocabulary: cv.iptc.org/newscodes/digitalsourcetype
• C2PA and Content Credentials: c2pa.org
• Content Credentials Verify: contentcredentials.org/verify
• ExifTool: exiftool.org
• Google Photos Help, Content Credentials: support.google.com/photos/answer/16496549
• MyHeritage Education, photo features: education.myheritage.com/article/myheritage-photo-features
• Adobe Help, AI content labeling: helpx.adobe.com/creative-cloud/apps/generative-ai/ai-content-labeling-faq.html