Every clue rules out
a run of years.
Phossil puts every week from 1900 to next year on the table and lets the photograph take years off it. The paper, the coat, the car at the curb, the stamp on the back, the birthday of the woman holding the baby. What survives all of them is the answer.
How do you know when my photo was taken?
We do not begin with a guess and then defend it. We begin with ignorance: every week between 1900 and next year, all of them equally possible, none of them favored.
Then the photo starts taking weeks away. A cabinet radio first sold in 1926 removes every week before 1926, because the room cannot contain a thing that does not exist yet. A woman born in 1931 removes every week before 1931. A cloche hat leans hard on the 1920s without forbidding anything either side of them.
The weeks left standing are the estimate. There are usually several years of them, and that is the finding, not a shortfall in it. So a photo carries two things at once: the likeliest year, and the years around it the evidence still allows.
What in the photo tells you the year?
Some clues forbid years outright. They are walls, and nothing softer is allowed to push through them.
- When the person in the photo was alive. A photo cannot have been taken before someone in it was born, or after they died. Where a birth date or a death year has been confirmed in your collection, that window is hard: the published range and the year we file the photo under are both pulled back inside it, and when softer evidence was pointing outside, the photo is flagged rather than quietly moved.
- Technology sitting in the frame. A console, a television set, a car model. A photo cannot predate the thing it contains.
- A photo-lab stamp on the back. Film cannot be exposed after it has been developed, so a stamp is a ceiling.
Others lean without forbidding. They move the weight around rather than removing years.
- A capture date carried by the file itself, when the file is the original. A camera writes the moment down, and a copy passed from phone to phone usually loses it.
- A face read against a birth date you have given us. The sharpest single clue in the pile, and sharpest of all on a young child, because a two-year-old’s face is a narrower statement than a fifty-year-old’s.
- Handwriting on the reverse. Someone writing June 1988 is reporting the day the photo was taken, so it points both directions equally.
- Clothing, hair, and the era the whole photo reads as, which are wide by nature and treated as wide.
- A season or a holiday, which pin the month without saying anything about the year.
- Another photo in the same collection that already carries a date, which lends its date to its neighbors.
This is the direction serious work on the problem points, too. A 2025 paper describes "a probabilistic framework that formally combines visual evidence" to infer the year a photo was taken (Paplham and Franc, arXiv, 2025).
Why is a lab stamp not the date the photo was taken?
Most of what is printed on the back of an old print is not handwriting at all. Across 166 photo backs in one archive, 147 carry a photo-lab stamp and 11 carry handwriting. Of the 120 dates read off those backs, 101 are a month and a year in the machine-printed style the labs used.
A stamp records the day the film was developed. So a roll shot at Christmas and dropped off in January is stamped the following year, and a system that reads the stamp as the capture date puts that photo a year late. Phossil treats a stamp as a ceiling with a short tail running backwards, which lets the estimate sit in December while the stamp says January. A bare year on a stamp gets a flat year, since a December developing is exactly as consistent with it as a February one.
One thing here is not finished. The window a stamp produces is right, and the month written in large type above that window can still be the month the film came back from the lab. So we are not putting a month on stamped backs across an archive until the label is as careful as the range underneath it.
How do the clues combine into one date?
Each clue is a shape drawn over the calendar. A birth date is a cliff. A release year is a step. A cloche hat is a hill. The shapes multiply together, and what is left is a belief spread unevenly across the weeks that survived.
From that belief we publish two numbers and one word. The window is the span that covers 95 out of 100 of the belief. The year we file the photo under is the middle of the tallest plateau, not the average, because when a photo has two candidate eras the average lands in the empty valley between them and puts the photo somewhere neither piece of evidence suggested. The word is the band: Anchored, Estimated, or Rough guess.
When the window comes out wider than 15 years, we publish no year at all. We couldn’t pin this one down, and we won’t pretend otherwise. And a date you enter yourself always wins. Nothing we work out later goes over the top of it. Watch a range close, clue by clue.
How well does it do on photos it has never seen?
Our test set is 62 photos whose dates a family confirmed, with 13 named people, all of them carrying birth dates.
The strict measure splits that set by person, so no one’s face appears on both sides: the engine is fitted on some people and then scored on photos of people it never saw. Measured that way, on the 56 photos where a face could be matched to a birth date, the estimate landed within two years of the truth on 45 of them. The average miss was 1.55 years.
Across all 62, including the ones with no matchable face, the shipped pipeline lands within two years on 47 and within five on 53, with an average miss of 2.08 years. That second figure was measured on the same photos the calibration was fitted to, so it flatters the engine, and the held-out number above it is the one that answers what to expect from your own shoebox.
We published the wrong one once. An earlier calibration got 52 of those same 56 within two years, at an average miss of 1.09 years. It had been fitted on those 56 photos and then scored on them: it was marking its own homework. The number stays on the record and is not used.
Two more corrections came out of scoring ourselves. Of 62 published date ranges, none was narrow enough to fit inside a single calendar year, and 51 of them were being displayed as a single year anyway. None are now. And our widest band, the one that already read Rough guess, was still carrying a small internal claim to being right. Of the 7 photos that landed in it, not one was. It now claims nothing at all.
Where does this stop working?
The test set is one family. Four photos from the 1980s, 55 from the 1990s, three from the 2000s, nothing before 1980, nothing after 2002, and few older faces. Every number above describes photos like those. A 1940s print is outside what we have measured, and we will say so again when we have measured it.
Splitting the test by person keeps a face off both sides of the line. It does not clear the board: a photo with several people in it can still share a person with the half the engine learned from. So even the strict number is a little kinder to us than a complete stranger would be.
The wait between shooting a roll and collecting the prints was fitted on 41 photo backs whose fronts carried a season or a holiday. That is enough to pin the usual wait to within a few weeks. It is not enough to pin the unusual one: 41 backs cannot tell us whether the rare slow case is two months or two years. We chose numbers that keep the rare case rare, and we are writing that choice down here rather than letting it read as a measurement.
And some photos give the engine nothing. A plain sweater, a blank wall, a close crop with no room around it: there is nothing there to rule years out, and the range stays wide. That is the correct answer to that photo. A scan also carries the day you scanned it rather than the day it was taken, so that date is given a weight that cancels itself out, which is why a print photographed today does not come back dated today.
Do I keep what I paid for?
You buy a collection once. It is not a subscription, and nothing you have paid for is taken back from you later.
You can always download your photos, in full resolution, whatever the state of your account. See pricing.
Back to what Phossil does. Once a photo knows when it happened, it can be drawn through time.