How Accurate Is AI Photo Dating? Measured Benchmarks

When a photo shows its date, our AI almost always gets the year right. When it doesn’t, most photos still land close, and tagging your family’s faces closes much of the gap.

Four children on a sofa, with an orange camera date stamp reading 96 2 14 in the corner of the print.

date on the print

The photo shows its date

An orange corner date, a lab stamp on the back, or a handwritten note.

96 to 99%land within a year of the true date

1 monthtypical miss, on average
95 to 99%in the right order

Two parents and a young boy on a picnic blanket in the mountains, with no date printed anywhere on the photo.

no date anywhere

Nothing on the photo says when

Uploaded just as they are, in no particular order. The most common box, and the hardest.

83%land within five years of the true date

31%within a year
77%in the right order

The same picnic photo, with each face outlined and labelled with a name and birth year.

Ray, born 1932 Danny, born 1962 Joan, born 1935

Same photos, family faces tagged

Name a face once and add a birth year. The photo itself needs nothing written on it.

91%land within five years of the true date

53%within a year up from 31%
82%in the right order up from 77%

See every number, row by row

In one sentence: prints that carry a date land 95 to 99% of photo pairs in the right order and 96 to 99% within a year of the true date; bare prints in no particular order land 77% of photo pairs in the right order and 83% within five years, and 82% and 91% once faces are tagged. Measured on 1,300 scanned prints with true dates from 1986 to 2005, verified by a person, five runs each. Example scans are illustrative: a generated sample family with invented dates. The benchmark prints are private family archives and are never shown.

Find the row that looks like your box

✦

Every benchmark at once. Find your row, then read across: three bars per chart, always in the same order.

No family information Family roster Tagged faces No demonstrated difference, drawn level

What your prints carry	In the right orderof photo pairs	Within 1 yearof photos	Within 2 yearsof photos	Within 5 yearsof photos	Within 10 yearsof photos	Typical missyears, lower is better
The print carries a dateRead off the print itself, front or back. Family information changes nothing measurable here.
A date on the front The orange date a camera burned into the corner of the print. date on the front 300 prints · 21 rolls and albums · 1989 to 2004 · 5 runs · a tagged face on all 300	99%99%99%99%=99%99%=	98%98%98%98%=98%98%=	99%99%99%99%=99%99%=	>99%>99%>99%>99%=>99%99%=	100%100%100%100%=100%100%=	0.10.10.10.1=0.10.1=
A lab date on the back The date the photo lab machine-printed on the back when it developed the roll. date on the back 300 prints · 39 rolls and albums · 1998 to 2003 · 5 runs · a tagged face on all 300	95%95%95%96%=95%96%=	96%96%96%98%=96%98%=	99%99%99%99%=99%99%=	>99%>99%>99%100%=>99%100%=	100%100%100%100%=100%100%=	0.10.10.10.1=0.10.1=
A note on the back Someone wrote the occasion, or the month and year, on the back in pen. note on the back 300 prints · 13 rolls and albums · 1986 to 2003 · 5 runs · a tagged face on all 300	>99%>99%>99%>99%=>99%99%=	99%99%99%99%=99%98%=	99%99%99%99%=99%99%=	99%99%99%99%=99%99%=	100%100%100%100%=100%100%=	0.10.10.10.1=0.10.1=
Nothing on the print carries a dateThe year has to come from the picture. This is where tagging faces pays.
Bare prints, album order kept Nothing on the front, nothing on the back, but each album was scanned in page order. 1st 2nd 3rd 4th Uploaded in album order. Chips show where each print truly belongs in time. 400 prints · 46 albums and rolls · 1986 to 2005 · 5 runs · a tagged face on all 400	83%83%83%84%=83%88%=	37%37%37%49%=63%63%+26	63%63%63%71%=83%83%+20	94%94%94%92%=98%98%+4	>99%>99%>99%>99%=>99%>99%=	1.91.91.91.7=1.01.0−0.9
The most common archive Bare prints, no order Nothing on the front, nothing on the back, and they go in as they come out of the box. 3rd 1st 4th 2nd Same four prints, uploaded in no particular order. Chips show where each print truly belongs in time. 400 prints · 46 albums and rolls · 1986 to 2005 · 5 runs · a tagged face on all 400	77%77%77%78%=82%82%+5	31%31%31%37%=53%53%+22	54%54%54%56%=74%74%+20	83%83%83%84%=91%91%+8	98%98%98%98%=98%98%=	2.72.72.72.4=1.71.7−1.0

Every figure is the mean of five runs over the same frozen prints, scored against dates a person verified. Point at any bar (or tab to it) for the full reading. Measurement ranges add the 95% interval and the spread across the five runs. Example scans are illustrative: a generated sample family with invented dates.

Reading the bars

Three bars per benchmark, always left to right: no family information, a typed family roster, faces tagged. The number above each bar is its score.

In the right order

Take any two photos with different true dates. It counts when the delivered timeline puts the earlier one first.

Typical miss

Mean absolute error: how far our date lands from the true date on average, counting early and late the same.

Points

Percentage points, printed under the bar that earned them. Going from 31% within a year to 53% is a gain of 22 points, marked +22.

No demonstrated difference

The change was smaller than the measurement error across five runs, so we do not claim it. Those bars are hatched, marked with an equals sign, and drawn level with the no-information bar.

Every photo gets a date

Coverage was 100% in every row, arm and run. A skipped photo would have counted as a miss.

Should you add family information?

✦

Three kinds of box, three answers.

Optional

Your prints carry dates

Rows 1 to 3. A typed roster and full face tagging both landed inside the measurement error.

In the right order95 to 99%
Within a year96 to 99%
Within five years99% or better
Typical missabout 1 month

Type a roster if you like. It will not hurt, and it is ready if you add undated prints later.

One minute

Bare prints, typed roster

Names and birth years, typed into a form. No faces picked out, no biometrics.

In the right order77% → 78%
Within a year31% → 37%
Within five years83% → 84%
Typical miss2.7 → 2.4 years

Every one of those moves sits inside the error bars, so we report no demonstrated difference. It is the seed list for tagging.

The big lever

Bare prints, faces tagged

Name a person once, add their birth year, and every photo they appear in gains an anchor: who is there and how old they look.

In the right order77% → 82%
Within a year31% → 53%
Within five years83% → 91%
Typical miss2.7 → 1.7 years

All four gains held in every one of the five runs. Face grouping asks for your consent first, and new tags trigger an included re-check of the dates you already have.

What a roster is

The form appears when you confirm an upload. Skip it and it is offered again with your results, where an included re-check applies it. List the people who are actually in the prints.

What a tag does

Two parents and a young boy on a picnic blanket, with each face outlined and labelled with a name and birth year.

Ray, born 1932 Danny, born 1962 Joan, born 1935

The same bare print as row 5. A boy who reads as about seven, born in 1962, puts it in 1969, and the adults agree. Both apps ask for your consent before any face is grouped.

One free lever: keep the album order

The same 400 bare prints, run in page order and shuffled. Keeping the order protects against the big misses, in every one of the five runs. Exact years barely move (37% versus 31% within a year, inside the error).

94% within five years, album order kept 1.9 years typical miss

83% within five years, order lost 2.7 years typical miss

Remember when one photo was taken? Pin it. After dating finishes, the app asks whether you know the year of a few photos spread across your timeline; a rough year is plenty, and the estimates around each answer are re-tuned at no charge. Measured in August 2026 on an earlier 200-print album-order set, not the matrix above: six pinned years took the exact-year rate from 42% and 59% to 94% and 82% in two runs and cut wrong-decade placements to about 2%. Those figures belong to that older set.

How we measured it

✦

The real systemThe same Timeline Scan AI a paying job gets, with the date readers switched on. No lab setup.
Five runs, ranges publishedOur AI is not perfectly repeatable, so every figure is the mean of five independent runs, and the spread is one tap away under Show measurement ranges.
True dates checked by a personThey come from camera timestamps, lab stamps, notes and family records, verified by hand before the set was frozen. Our AI never sees them.
Nothing gives the date awayFilenames, folders and file data are scrubbed. Each of the 400 bare prints was checked by a person for anything on it that could reveal its date.
Re-run before anything shipsThe whole matrix is re-run before any change to the dating goes live. A change that does not hold up does not ship.

Orientation: where the 99.8% figure comes from

✦

99.8% of scans turned right side up 998 of the 1,000 test scans

Turning a scan right side up is a separate job from dating it, measured separately. One detector does it everywhere: the web service, the iPhone app, and our free FrontBack Scanner. It reads each print in all four quarter-turns before deciding, so a hard photo gets four opinions instead of one.

Measured on 1,000 scanned prints from the same family archives: 500 left exactly as they came off the flatbed and 500 given random quarter-turns. When the detector cannot tell (a dark frame with no vertical reference), it keeps the scan as you laid it rather than guess, with a one-click rotate in the preview either way. Built on the MIT-licensed deep-image-orientation-detection model by Duarte Barbosa. Last re-measured June 2026.

Honest limitations

✦

One family archive so far. All 1,300 prints come from one family, split across 13 to 46 albums and rolls per row. Enough to measure this system honestly; not proof the numbers transfer to every family’s prints.
One reviewer, and private ground truth. One person verified the dates, and the prints are private family photos, so nobody outside can re-run the benchmark. We say so rather than imply independence we do not have.
The prints span 1986 to 2005. Treat pre-1980 accuracy as unmeasured rather than promised.
Run-to-run variance is real on bare prints. On the album-order row, the within-a-year rate swung between 32% and 41% across the five runs. That is why every figure is a mean of five and the ranges are published.
Exact years are harder than order. A bare print’s exact year is hard even for a person. If you need every photo within a year, expect to nudge some estimates; the review timeline exists for that.

For comparison, MyHeritage reports about 60% of PhotoDater’s estimates within 5 years of the actual date. Ours is 99% or better when the print carries a date, 83% on bare prints in no order, and 91% once faces are tagged. Every photo-dating number depends on how much evidence the photos carry; our side-by-side comparison covers when each tool fits.

How Accurate Is Timeline Scan’s AI Photo Dating?