Culture · Essay

AI can sort your family album. It cannot caption it truthfully.

Face grouping, search and transcription have genuinely changed what a photo collection can do. But the two hard jobs, knowing who matters and what actually happened, cannot be inferred, and 'AI restoration' that invents detail is a falsified record.

Written to last.

By Confinity Editorial · 2026-04-23 · 6-minute read

Quiet tools, kept out of the way.

Machine learning has quietly changed what an ordinary photo collection can do, and the useful changes deserve their own list, because they sit next to some heavily marketed ones that a family archive should refuse.

The four things that genuinely work

Face grouping. Modern face detection can cluster thirty years of photographs by who appears in them, which turns "somewhere in here are pictures of Dad" into a folder. It does not know the name; it knows the sameness. One person supplying the name once turns a cluster into a record. Search by content. Finding "the beach ones" or "the pictures with the dog" without anybody ever having tagged anything removes the single biggest barrier to actually using a large collection. Date and place recovery. Scanned prints carry no metadata, but models can now estimate era from clothing, cars and film stock, and match backgrounds to places, giving a sortable skeleton to a shoebox of undated prints. Estimates, to be checked, but a scaffold where there was nothing. Transcription. The least glamorous and most valuable: an hour of a grandmother talking becomes searchable text attached to her voice. The recording stays the artefact; the transcript makes it findable. This is the machine help Confinity actually uses: tidying as you capture, suggesting connections between the people you name, training on nothing.

The two jobs that cannot be inferred

A model can sort every photograph you own and still not know the only two things a descendant will ask. Who matters. Eighty photographs of a school friend who moved away, three of the great-aunt who raised your father: frequency is not importance, and no algorithm can rank a family's people from the outside. The three photographs are the treasure, and only somebody inside the family knows it. What happened. The caption a grandchild needs is not "four people, outdoors, 1970s", which is where machine description tops out. It is "the day they reopened the shop after the fire, and the only time anyone saw Grandad cry". That sentence cannot be inferred from pixels. It has to be told, by someone who was there, and there is a deadline on it nobody can see. Sorting is now nearly free. Meaning still costs what it always cost: a person's attention, applied in time.

The restoration line, and why we will not cross it

Then there is "AI restoration", and here a family archive has to draw a line that photo apps will not draw for you. There are two different operations sold under that name. One recovers: descreening, scratch removal, contrast recovery, careful sharpening that works with the detail actually present in the scan. That is restoration in the conservator's sense, and it is fine. The other generates. Ask a model to "enhance" a blurry face and it does not reveal your great-grandmother's face; it invents a plausible face and paints it over hers. The result is sharper, more shareable, and partly fictional, and once it circulates in the family it silently replaces the genuine image. An invented face on a real ancestor is not an improvement to the record. It is a fabrication inside it, unlabelled, permanent, and increasingly impossible for a descendant to detect. This is why Confinity holds the position it does: no generated imagery in the archive, ever. The blur is the truth of what survived. A record's whole value is that it can be believed, and believability does not survive invisible invention.

The album, divided correctly

So the digital family album lands in a clean division of labour. Machines: grouping, finding, dating, transcribing, the clerical work they are built for. Humans: naming, choosing, explaining, the meaning only they possess. And a hard boundary at generation, because the one thing an archive must never do is make things up. An album sorted by a machine and captioned by the family beats either alone, and beats by miles the collection most families actually have: perfectly stored, wholly unexplained, waiting for questions that get harder to answer every year.