How Accurate Is Timeline Scan’s AI Photo Dating?

By Spencer Baker, founder · Last updated: October 31, 2026

Timeline Scan, the AI photo dating service, estimates dates for scanned family photos. Here is how it performed on prints with dates checked by hand.

The short answer

  • A toddler holding a pot in the hallway, with an orange camera date stamp reading 9 2 ’93 in the corner of the print. Date on the photo

    99 to 100%

    right within a year

  • A boy holding a fishing rod by a lake, with no date printed anywhere on the photo. No date anywhere

    96%

    within five years

  • The same lake photo, with the boy's face outlined and labelled Spencer, born 1992. Spencer, born 1992 No date, faces named

    Over 99%

    within five years

Photos like yours

✦

Photos with a date on them

A toddler holding a pot in the hallway, with an orange camera date stamp reading 9 2 ’93 in the corner of the print.

A date on the front

Over 99% right within a year

Naming faces: not needed

See the details

The orange date some cameras burned into the corner.

300 prints from 21 rolls and albums, taken 1989 to 2004.

Test 1 results
Photos onlyFaces named
Within a yearover 99%over 99%
Within 5 years100%100%
Right orderover 99%over 99%
Exact day95%95%
Typical misswhen not exact2 months3 months

Measured October 15, 2026: three runs with photos only and one with faces named, with no date-stamp pre-check.

In every run, with or without faces named, 299 of 300 prints were within a year; the one miss was 366 days off.

The back of a photo print, machine-printed by the lab with a frame code and the date 07.17.01. Back of the print

A lab date on the back

99% right within a year

Naming faces: not needed

See the details

The date the photo lab printed on the back.

Usually a few days after the picture was taken. 300 prints from 39 rolls and albums, taken 1998 to 2003.

Test 2 results
Photos onlyFaces named
Within a year99%99%
Within 5 years100%100%
Right order98%98%
Exact day87%87%
Typical misswhen not exact3 months▲ 2 months
The back of a photo print with a handwritten note reading Spencer July 3, 1994. Back of the print

A note on the back

Over 99% right within a year

Naming faces: not needed

See the details

A month and year, or an occasion, written by hand.

300 prints from 13 rolls and albums, taken 1986 to 2003.

Test 3 results
Photos onlyFaces named
Within a yearover 99%99%
Within 5 years100%over 99%
Right orderover 99%over 99%
Exact day78%77%
Typical misswhen not exact36 days2 months

Differences this small are within measurement error.

Photos with no date

1st 2nd 3rd 4th

Still in album order

100% within five years

Naming faces: may help

See the details

No date anywhere, with each album scanned in page order.

400 prints from 46 albums and rolls, taken 1986 to 2005.

Test 4 results
Photos onlyFaces named
Within a year77%89%
Within 5 years100%100%
Right order94%94%
Exact day7%▲ 19%
Typical misswhen not exact7 months5 months

Naming faces scored higher here, but albums varied too much to call it proven.

3rd 1st 4th 2nd Most common

Loose, in no order

96% within five years

▲ Over 99% with faces named

See the details

Loose prints scanned as they came out of the box.

The same 400 prints as Test 4, shuffled.

Test 5 results
Photos onlyFaces named
Within a year65%▲ 77%
Within 5 years96%▲ over 99%
Right order89%▲ 92%
Exact day7%14%
Typical misswhen not exact14 months▲ 9 months

▲ means a clear improvement, more than chance could explain.

What do these measures mean?
Within a year
Our date was no more than a year from the real one. Within five years works the same way.
Right order
Take any two photos from different dates: did the older one come first?
Exact day
Our date matched the real day exactly.
Typical miss
For photos we didn't date to the exact day: how far off our date was, on average, early or late.
Faces named
We named the people once, with the year each was born, as you can in the app.

Should you name faces?

✦
  • Dates on your photosNot needed
  • No datesYes, it helps most
  • Still in albumsScan in page order
A boy holding a fishing rod by a lake, with the face outlined and labelled with a name and birth year. Spencer, born 1992
Undated photos within a year

65 of 100

Photos only

77 of 100

Faces named ▲ 12 more

Why does naming faces help?

Name someone once, add the year they were born. The web app and the iPhone app both ask your consent before grouping faces.

Spencer was born in 1992 and looks about five here, which puts this photo near 1997. When nothing on a print says when, knowing who is in it is the strongest clue left.

On undated prints in no order it took “within a year” from 65% to 77%, “within five years” from 96% to over 99%, and cut the typical miss from about 14 months to about 9 months.

If a date is printed or written on the photo, naming faces adds little. With a lab date on the back, it helped a little.

Keeping album order costs nothing. On the same 400 undated prints it took “within five years” from 96% to 100% and the typical miss from about 14 months to about 7 months.

Why you can trust these numbers

✦
  • 1,300 real family prints
  • Dates checked by hand
  • No hints left in the files
  • Every photo counted
  • Re-checked after changes
How we kept the test fair

Five tests on 1,300 family prints with known dates. Run and checked by Spencer Baker, founder of Timeline Scan. The set was frozen in August 2026.

  • The same dating model you get. The app now runs without a date-stamp pre-check; reading text on the back is unchanged.
  • What was the pre-check? A separate step that only adds a crop, a close-up of the date, when it judges a print to carry a visible stamp.
  • Test 1: measured October 15, 2026 with the dating model alone, no date-stamp pre-check: three runs with photos only and one with faces named. The first photos-only run reuses completed October 14 predictions.
  • Tests 2 to 5: one published run per arm using the October 14, 2026 setup with the pre-check configured.
  • Test 3's faces run was repeated on October 15, 2026 with the same setup, after we corrected two face tags.
  • Dates checked first. A person verified dates from camera stamps, lab stamps, notes or family records before testing. These answers were not supplied to the AI.
  • No hidden hints. File names, folders and file data were wiped. A person checked all 400 undated prints for anything that could give away the date.
  • Every photo counts. Every photo received a date. A skipped photo would have counted as a miss.
  • One Test 3 print hit the AI provider’s content filter. It was dated from its back note and nearby photos, as in the live service.
What these numbers don’t tell you
  • One family archive. The 1,300 prints come from 46 albums and rolls. Other families may get different results.
  • One era. Prints were taken 1986 to 2005. Photos before 1986 have not been measured.
  • Private photos. One person checked the dates. Nobody outside can re-run the test.
  • Repeated photos. Test 1 uses the same 300 prints: three runs with photos only and one with faces named. Its comparison pairs albums, not runs.
  • Test 1 setup. Measured October 15, 2026 with the dating model alone, no date-stamp pre-check.
  • Single runs. Tests 2 to 5 use one published run per arm with the October 14, 2026 setup and pre-check configured. These and the Test 1 faces range measure album differences, not repeatability.
  • Exact years are harder than order. Expect to change a few undated photos by hand.
  • Example photos are real prints. They are childhood photos of Spencer Baker, who built Timeline Scan, from the same family archive. The rest of the test prints stay private.
Every number from the benchmark
  • Date on the print: 98 to 100% of photo pairs in the right order and 99 to 100% within a year of the true date.
  • No date, in no order: 89% of photo pairs in the right order and 96% within five years, and 92% and >99% once faces are named.
Swipe sideways to see all the numbers.
Timeline Scan dating accuracy by test, October 2026. Small print under each figure is its 95% range.
TestWayIn the right orderphoto pairsWithin 1 yearWithin 2 yearsWithin 5 yearsWithin 10 yearsExact dayTypical missaverage when not exact, in days
Test 1A date printed on the front300 prints · 21 albums and rollsPhotos only3 runs>99%99 to 100>99%99 to 100100%100100%100100%10095%91 to 9964 days5 to 169
With faces named1 run>99%99 to 100>99%99 to 100100%100100%100100%10095%90 to 9883 days25 to 239
Test 2A lab date on the back300 prints · 39 albums and rolls1 full runPhotos only98%95 to 9999%98 to 100>99%99 to 100100%100100%10087%81 to 9286 days40 to 133
With faces named98%96 to 9999%98 to 100>99%99 to 100100%100100%10087%82 to 92▲ 64 days25 to 113
Test 3A handwritten note on the back300 prints · 13 albums and rolls1 full runPhotos only>99%98 to 100>99%99 to 100>99%99 to 100100%100100%10078%69 to 8436 days15 to 79
With faces named>99%98 to 10099%97 to 10099%97 to 100>99%99 to 100100%10077%68 to 8367 days15 to 197
Test 4No date, album order kept400 prints · 46 albums and rolls1 full runPhotos only94%88 to 9677%60 to 8998%95 to 99100%100100%1007%2 to 14214 days139 to 293
With faces named94%86 to 9689%81 to 9499%97 to 100100%100100%100▲ 19%10 to 27158 days108 to 224
Test 5No date, no particular order400 prints · 46 albums and rolls1 full runPhotos only89%82 to 9165%59 to 7187%81 to 9296%93 to 99>99%99 to 1007%3 to 11439 days349 to 536
With faces named▲ 92%86 to 94▲ 77%69 to 83▲ 94%92 to 97▲ >99%99 to 100>99%99 to 10014%8 to 19▲ 265 days206 to 333
  • Typical miss is the average for photos not dated to the exact day.
  • ▲ marks a clear improvement with faces named. The paired 95% range of the difference excludes zero. Other measured differences are within measurement error.
  • The 95% ranges use 1,000 re-samples of albums and rolls, plus each arm’s own runs. Test 1 compares each option’s run average on the same album samples; runs are not paired.
  • Results and range endpoints round to whole percentage points. An imperfect result that would round to 100% appears as >99%.
Benchmark results reuse terms

Version 1.0, October 31, 2026. These terms apply only to the aggregate benchmark results published on this page. © 2026 Timeline Scan LLC. All rights reserved. For reuse permission, contact support@timelinescan.com. Uses permitted by applicable law are unaffected.

This notice grants no rights to photographs (including example images), face data, or individual photo records. The underlying family archive remains private.

99.8%

of scans upright

How we tested it

998 of 1,000 test scans. Measured separately from dating.

From the same family archive: 500 scans as they came off the scanner and 500 turned at random.

The checker looks at each photo all four ways up. When it can’t tell, it leaves the scan as you laid it.

The same checker runs in the web service, the iPhone app and our free FrontBack Scanner. Built on the MIT-licensed deep-image-orientation-detection model by Duarte Barbosa. Last re-measured June 2026.

Questions

✦
Do I need to name the people in my photos?

Only for photos with no date on them. That is where it helps most.

If a date is printed or written on the photo, naming faces adds little. With a lab date on the back, it helped a little.

On loose photos with no dates, it helped more than anything else: within five years rose from 96% to over 99%, and within a year from 65% to 77%.

Will it work on photos older than 1986?

Yes, it dates older photos the same way. Our test prints were taken 1986 to 2005, so we have no accuracy number for older ones yet. Try some of your own scans free before you pay.

Can AI tell the exact day a photo was taken?

If a date is printed or written on the photo, most come back with the exact day: 78 to 95 of every 100 in our tests.

When one misses, it is off by about 1 to 3 months on average.

With no date, it estimates the year. On loose prints in no order, a miss was off by about 14 months on average, or about 9 months with faces named.

What if it gets a date wrong?

Open the photo and press Change date to type the right one.

How does Timeline Scan compare with MyHeritage PhotoDater?

MyHeritage says about 60% of PhotoDater's guesses land within 5 years of the actual date.

By the same measure, ours was 100% with a date on the print, 96% on loose prints with no date, and over 99% once faces were named.

Results depend on what your photos show. Our side-by-side comparison explains when each tool fits best.

Try your photos

Ways to try

The free 50-photo trial lets you judge the dating on your own scans before you pay.

Curious about one print first? Try the free what year was this photo taken checker, no account needed.

Try It Free