Restoring the Photo in the Box. What AI Repairs, and What It Invents.

Two people sit on a front step, squinting at whoever held the camera. The print has faded to the colour of weak tea. A crease runs down the middle, missing her face by a few millimetres, and the bottom corner has gone altogether.

They're your grandparents, probably. Someone has kept the photo in a box for forty years, untouched, because until recently nothing could be done.

Now an image tool (ChatGPT or Gemini, because Claude doesn't make images) will hand back a repaired, coloured version in a few minutes. People cry when they see the result, and this is one of the best things AI does. But some of what you'll see was never there.

When the AI removes a crease, it's guessing what was under the crease. When it sharpens a blurred face, it's drawing a face that fits the blur. When it colours a black-and-white dress, it's choosing a colour the photo never recorded.

Most of the guesses are good. But sometimes a guess gives your great-grandmother slightly different eyes, or a cardigan that was never that green. Then the result is partly a painting of the original.

That's no reason to leave the photo in the box. Keep the original, label the new one, and ask the person who remembers.

Photograph it flat, then put the print away

Lay the print flat by a window, with the flash off, and hold your phone straight above it so the edges come out square. Take three and keep the sharpest. If you have a scanner, a scan beats a photo of a photo.

Save the picture somewhere safe, with a plain name like nan-and-grandad-step-original.jpg, because this is the real one.

Everything from here is a version of it, and in twenty years someone will want to know which was which. The print goes back in the box, and the AI never needs it.

Ask for repair, and nothing else

A photo you upload is a photo you've shared with a company. For most old family pictures that's fine, as long as you make the choice on purpose. Image tools also limit how many pictures you can make, so a whole shoebox takes a few evenings.

Do the repair first, on its own, so you can see where repair stops and invention starts.

Restore this old photo. Remove the creases, scratches, stains and fading, and gently sharpen it. Don't change anyone's face, expression, age or features. Don't add or remove anything, and keep it black and white. Keep the whole frame, including the edges.

Set the result beside the original and look. Are the faces the same people, and is the background still the background? If something has drifted, say so, and the AI will fix it. For example, "you've changed her mouth, so put it back as it was."

If half of Grandad's face is missing, what comes back is a believable face. It's a guess at Grandad, lovely and not him, so write that on the back.

Give it the colours you know

If you want colour, ask for it separately, and tell the AI every real detail you have. "Her coat was red, his car was black, this was taken in Jamaica in about 1962."

Each detail is one less guess, but even so, colour is always invented. The AI is picking a 1962 from all the pictures it learned from, not from the day itself. So treat the result as what the scene probably looked like, and save this version under its own name, ...-restored-colour.jpg.

Take it to whoever remembers

This is where a nice picture becomes a family thing. Take the coloured version to whoever's oldest and ask, is that right?

They'll tell you the dress was blue, and that the step was outside the house on Beckford Road. They'll tell you the man on the left wasn't Uncle Len, it was his brother, and here's a story about him.

The AI guessed, but they know. Go back and fix what they told you, and write the story down next to the photo.

Print it, and give it back

Print the photo properly, on photo paper, with the words underneath, and give it to the person who kept the original for forty years.

Nan and Grandad, Beckford Road, 1962, the week before they moved.