Dating Undated Photographs and Documents
Written for archivists, cataloguers, genealogists and researchers holding material that nobody bothered to date.
Most collections have a box like this: photographs with nothing on the back, letters without a year, forms with the date field left blank. Somebody writes "1930s?" on the folder, and that guess travels into the catalogue, family trees and citations.
A good date estimate is not a guess with more confidence. It is a range built from evidence, with the evidence written down. This post covers which clues narrow a date, how to combine them, where AI helps and misleads, and how to record the result.
Two kinds of evidence
Internal evidence is what the item says or shows: names, events, prices, signs in the street, the spelling of the text.
External or physical evidence is what the item is: paper, print process, card mount, letterhead, stamps, seals, the format of a form. Context counts too: the folder, the donor, dated items in the same collection.
Physical evidence often gives a broad window of decades. Internal evidence can be sharp but is easy to misread. Good estimates use both.
The logic: after this, before that
Archival dating rests on two limits:
- Terminus post quem (TPQ): the earliest possible date. A photograph showing a building cannot be earlier than the building. A letter mentioning an event cannot be earlier than the event.
- Terminus ante quem (TAQ): the latest possible date. A postmark, a dated accession stamp, or the death of the person who wrote the caption on the back.
State every clue as one of these limits, or as a "most likely" window. "The clothes look 1920s" is an impression. "The tram line in the background opened in a known year" is a limit.
One trap runs through all of this: manufacture is not use. Stationery stays in drawers for years, forms are filled in long after printing, and cars stay on the road for decades. A manufacturing date is a TPQ, almost never a TAQ.
Documents: what narrows the date
Dated references in the text. Events, prices, a newspaper the writer has just read. Titles and ranks help too: a person addressed as "Doctor" or by a military rank had reached that stage by the time of writing. Cross-check each reference; writers recall old events as well as new ones.
Letterheads, stamps and seals. Institutions change names, addresses and emblems, and those changes are often documented. A place name can do the same work: the Russian capital was named Petrograd from 1914 to 1924, so stationery printed "Petrograd" was printed in that window, though old stock could be used later. A Hebrew letterhead or seal of the State of Israel cannot be earlier than May 1948. Postage stamps have documented issue dates, and a postmark, when legible, is usually the strongest single limit on a letter or postcard.
Script and orthography. Russian spelling was reformed by decrees of 1917 and 1918, which removed letters such as ѣ, і and ѳ. A text consistently in the new spelling points to the reform years or later. Old spelling is not a terminus ante quem: émigré publishers kept it for decades, and many private writers kept the spelling they had learned at school. The post on pre-revolutionary Russian OCR covers the letters themselves. In Yiddish, Soviet publishing from the 1920s spelled words of Hebrew origin phonetically (עמעס rather than אמת). That points to a Soviet context after the revolution, not to a specific year.
Typewriter or handwriting. A typed document says little on its own. Its value is comparison: the same machine, with the same damaged letters, can link an undated sheet to dated letters.
Printed forms. Many official forms carry a small printer's line at the foot: the press, an order number, sometimes a year of printing. That year is a TPQ for the form, not the date it was filled in.
Photographs: what narrows the date
Format and process. Card mounts, print sizes, borders and processes changed over the decades. Reference resources such as the Image Permanence Institute's Graphics Atlas help identify processes, and a conservator can confirm. Treat format as a window of years, often decades, not a point.
The studio imprint. A studio name and address on the mount can often be checked against city directories or advertisements for the years that studio operated at that address.
The verso. Pencil notes, dedications, studio stamps, an inventory number from an earlier collection, a photo-paper manufacturer's mark. Read everything, including faint pencil, as the photo cataloguing guide describes. A dated dedication is a strong TAQ, though the print may be years older.
Signs, vehicles and buildings. Shop signs and street names can be checked in directories. Vehicles give a TPQ from the year the model appeared. Buildings that were built or demolished at known dates give limits in both directions, which makes them among the most useful clues in street scenes.
Clothing and hairstyles, with caution. Fashion is real but weak evidence. It spread unevenly, older people kept older styles, and studios lent costumes. Use it to support a range, never to fix a year.
Combining weak clues into a range
Take a hypothetical street photograph. The print format suggests a window of several decades. A shop sign carries a family name that appears in a city directory for a span of about fifteen years. A building in the background is known to have been completed in a certain year, which gives a TPQ. The verso carries an accession stamp from a donation, which gives a TAQ.
None of these dates the photograph alone. Together they narrow it to the overlap, perhaps a decade. The latest TPQ and the earliest TAQ define the range; everything else refines the likely part of it. When two clues conflict, record the conflict and leave the range wide rather than discard the inconvenient clue.
Where AI helps, and where it misleads
AI is useful at the evidence-gathering stage:
- Reading verso notes, stamps and seals, including handwriting and non-Latin scripts.
- Spotting datable details: signage, vehicles, uniforms, posters and landmarks a busy cataloguer may miss.
- Cross-checking named events, people and institutions against reference sources, when it shows the source.
It misleads in predictable ways:
- Confident single-year guesses. Asked "when was this taken?", a model will often answer with one year and no reasons. It looks precise and is rarely defensible.
- Style-based dating presented as fact. "1920s fashion" is an impression, and a model will state it in the same tone as a legible postmark.
The rule: the model proposes clues and a range, with the reason for each limit; a person checks the limits. Confidence labels need checking too: the confidence-calibration case study shows a "certain" category that was right only 56% of the time before calibration.
Recording the date honestly
A date estimate needs four parts in the record:
- A display date for readers, the way a cataloguer would write it: "c. 1930s", "[1934?]", "before 1938".
- A machine-readable form in EDTF, the Extended Date/Time Format developed at the Library of Congress and later incorporated into ISO 8601-2. Examples:
193X(some year in the 1930s),1934~(approximately 1934),1934?(uncertain),1930/1939(an interval). For "one year in this span, we do not know which", EDTF also offers[1930..1939], and[..1938]for "1938 or earlier". The post on AI metadata and archival standards shows how these map to MODS and EAD. - The evidence: a short note listing each limit and its source. "TPQ: building completed (city records). TAQ: accession stamp on verso."
- A confidence mark, such as confirmed, probable or tentative.
The note matters most: a future cataloguer with a new clue can tighten the range only if the old reasoning is visible. Genealogists know the same problem from name variants: a conclusion without its evidence cannot be corrected, only repeated.
A short checklist
- Separate internal evidence from physical evidence.
- State each clue as a TPQ, a TAQ or a likely window.
- Remember that manufacture is not use: paper, forms and mounts give earliest dates.
- Read the verso completely, including faint pencil and stamps.
- Treat old orthography and clothing as weak evidence.
- Build the range from the latest TPQ and earliest TAQ; record conflicts.
- Ask AI for clues with reasons, not for a year.
- Record a display date, an EDTF value, the evidence note and a confidence mark.
- Leave the date empty when there is no evidence.
Our catalog drafting service applies the same discipline: it drafts dating for documents and photographs, including notes on the back, marks each identification as confirmed, probable or tentative, leaves the field empty when the source is silent, and a person on your staff approves every record. If you have a box of undated material and want to see how far the evidence goes, get in touch.
Frequently Asked Questions
How can I tell how old a photograph is?
Collect several independent clues rather than relying on one: the print format and mount, the studio imprint and any notes or stamps on the back, and datable things in the image such as signs, vehicles and buildings. Each clue gives an earliest or latest possible date, and together they define a range.
Can AI date an undated photograph?
AI can gather evidence quickly: it reads verso inscriptions and stamps, spots signage and objects, and checks named events. It should not be trusted when it offers a confident single year based on style alone. Treat its output as a proposed range with reasons that a person verifies.
What do terminus post quem and terminus ante quem mean?
Terminus post quem is the earliest date something can be, for example the date a building in the background was completed. Terminus ante quem is the latest date, for example a dated postmark. A date range sits between the strongest pair of such limits.
How should an uncertain date be recorded in a catalogue?
Keep a display date for readers, such as 'c. 1930s', and a machine-readable form in EDTF, such as 193X or 1930/1939. Add a short note on the evidence and a confidence mark, so the next person can check the reasoning.