Photo Metadata

Photo metadata

Overview

Metadata embedded in an image file travels with the file. Photo CMSs (Photo Mechanic, Lightroom, Capture One) and wire intake systems read a photo's caption, credit, and rights from its IPTC and XMP blocks, not from a separate document or the email it arrived in. If the caption, credit, alt text, and license are not inside the file, they are gone the moment the photo is downloaded, forwarded, or re-uploaded.

One exiftool pass writes the EXIF, IPTC, and XMP layers together and leaves every other tag (camera settings, shot time) untouched. Modern software reads XMP first, legacy IPTC-IIM second, EXIF only for date and GPS, so write XMP everywhere and add IIM as a compatibility copy on JPEG/TIFF (HEIC, AVIF, and WebP have no IIM slot at all; see reference.md).

Two things changed since this workflow was "caption, credit, copyright." First, how an image was made now belongs in the metadata: the IPTC Digital Source Type field labels a camera photo versus an AI-generated one, and platforms (Meta, Google) and the EU AI Act increasingly read it. Second, cryptographic provenance (C2PA / "Content Credentials") is arriving on wire images and cameras, a signed layer exiftool can read but not write. Both are covered below.

A capable model already knows the field names. The hard part is not the mechanics, it is the judgment below. Lead with that.

When to use

• Prepping press photos for a wire so partner newsrooms can search, credit, and republish them
• Adding required photographer attribution and a reuse license before publishing or sharing
• Labeling how an image was made, a straight photo, an AI-generated illustration, an AI-edited composite
• Batch-tagging a shoot (a folder of images)
• Making images accessible (embedded alt text) and rights-clear (copyright or Creative Commons)
• Reading and sanity-checking the C2PA Content Credentials on an image that arrived from an agency

When not to use: editing pixels (this is metadata only); writing alt text for an HTML <img> (use accessibility-compliance); preserving web pages as evidence (use web-archiving); signing a Content Credential (exiftool can't, use c2patool, below).

The discipline (what agents get wrong)

These are the failures a capable agent makes anyway. They matter more than any tag name.

1. Caption only what is visible. Describe what the frame shows, not what you were told. Do not infer events, intent, identities, relationships, or legal status you cannot see. "Demonstrators gather to protest a court ruling" is a claim about facts not in the frame; "A crowd holds signs outside a courthouse" is the photo.
2. Label people from visible evidence. Name an agency or role only from a visible marking, a labeled vest, a uniform, a badge, a patch. Otherwise write "officers in tactical gear," "a man in a blue shirt." Never assert someone's immigration or legal status (no "detainee," no "undocumented") unless it is unambiguous in the frame.
3. Always write alt text, it is not the caption. Write both: a short screen-reader description in XMP-iptcCore:AltTextAccessibility and the publishable caption in IPTC:Caption-Abstract. IPTC keeps these deliberately distinct, the caption states facts and is shown on the page; the alt text is read aloud by a screen reader, so do not just copy one into the other. Agents routinely write the caption and skip the alt text.
4. Label how the image was made, and never lie about it. If an image is AI-generated or AI-edited, say so in XMP-iptcExt:DigitalSourceType; if it is a straight photo, digitalCapture states that plainly. Do the honest thing and label it; do the diligent thing and, on an inbound file, never strip an existing Digital Source Type or C2PA credential, that erases a disclosure someone made on purpose.
5. Strip GPS when the location could endanger someone. A protester, a source, an abuse survivor, a minor, embedded coordinates can reveal a home or a safe house. Remove GPS from the published derivative (-gps:all=) while keeping the caption and credit; keep a full-GPS archival master only where location is editorial evidence. GPS is the single highest-risk tag in the file.
6. Keep structured fields neutral. Editorial framing or a contested label belongs in Headline, never in City, Caption-Abstract, or the location fields. Partner newsrooms apply their own language; clean structured fields let them.
7. Verify the round-trip from source. Read the metadata back from the written file, not from your buffer. After any upload or transfer, re-read it from the destination, a 200 response proves the bytes were accepted, not that the metadata survived. Most social platforms re-encode on upload and strip IPTC, XMP, GPS, and C2PA (see reference.md), so "I embedded it" is not "it arrived."

Quick reference, the fields that carry the weight

Digital Source Type values (fully AI → trainedAlgorithmicMedia, AI-edited → compositeWithTrainedAlgorithmicMedia, straight photo → digitalCapture), the full IPTC controlled vocabulary, the IPTC-IIM byte limits, the PLUS/Google-licensing and Creative Commons field sets, the C2PA tooling, and the AP caption recipe: see reference.md.

One pass that writes all three layers

CAPTION="A crowd holds signs outside the Mercer County Courthouse, Friday, June 19, 2026, in Trenton, N.J. (Dana Rivera/Example News Collective)"
ALT="A crowd of people holding handmade signs stands on the steps of a stone courthouse."
# how the image was made, a full IPTC CV URI (see reference.md for all values)
DST="http://cv.iptc.org/newscodes/digitalsourcetype/digitalCapture"

exiftool -codedcharacterset=utf8 -overwrite_original -P \
  -EXIF:Artist="Dana Rivera" -XMP-dc:Creator="Dana Rivera" -IPTC:By-line="Dana Rivera" \
  -IPTC:Credit="Example News Collective" -XMP-photoshop:Credit="Dana Rivera / Example News Collective" \
  -IPTC:Caption-Abstract="$CAPTION" -XMP-dc:Description="$CAPTION" -EXIF:ImageDescription="$CAPTION" \
  -XMP-iptcCore:AltTextAccessibility="$ALT" \
  -XMP-iptcExt:DigitalSourceType="$DST" \
  -IPTC:Keywords="protest" -IPTC:Keywords+="Trenton" \
    -XMP-dc:Subject="protest" -XMP-dc:Subject+="Trenton" \
  -EXIF:Copyright="(c) 2026 Example News Collective. Licensed CC BY 4.0." \
    -IPTC:CopyrightNotice="(c) 2026 Example News Collective. CC BY 4.0." \
    -XMP-dc:Rights="(c) 2026 Example News Collective. Licensed CC BY 4.0." \
  -XMP-xmpRights:Marked=True \
    -XMP-xmpRights:WebStatement="https://creativecommons.org/licenses/by/4.0/" \
    -XMP-xmpRights:UsageTerms="Licensed CC BY 4.0. Credit: Dana Rivera / Example News Collective." \
    -XMP-cc:License="https://creativecommons.org/licenses/by/4.0/" \
    -XMP-cc:AttributionName="Dana Rivera / Example News Collective" \
  -IPTC:City="Trenton" -IPTC:Province-State="New Jersey" \
    -IPTC:Country-PrimaryLocationName="United States" -IPTC:Country-PrimaryLocationCode="USA" \
  "-IPTC:DateCreated<EXIF:DateTimeOriginal" "-IPTC:TimeCreated<EXIF:DateTimeOriginal" \
    "-XMP-photoshop:DateCreated<EXIF:DateTimeOriginal" \
  photo.jpg

-P preserves the file's modification time; drop it if you want the write to touch the timestamp. Extended accessibility descriptions for complex images (charts, infographics) go in XMP-iptcCore:ExtDescrAccessibility, a separate field from the caption, added only when the alt text plus surrounding text can't convey the image.

Then verify from the file (the step agents skip):

exiftool -G1 -s -IPTC:By-line -IPTC:Caption-Abstract -XMP-iptcCore:AltTextAccessibility \
  -XMP-iptcExt:DigitalSourceType -XMP-cc:License -IPTC:Keywords photo.jpg

Label how an image was made (AI and synthetic)

XMP-iptcExt:DigitalSourceType records origin from the IPTC controlled vocabulary. The value is a full URI, exiftool does not validate it, so a bare word or a typo is silently accepted and useless. The three every newsroom needs:

BASE="http://cv.iptc.org/newscodes/digitalsourcetype"
# a straight camera photo, worth stating even for real news images
exiftool -XMP-iptcExt:DigitalSourceType="$BASE/digitalCapture" photo.jpg
# fully AI-generated (a trained model produced the whole image)
exiftool -XMP-iptcExt:DigitalSourceType="$BASE/trainedAlgorithmicMedia" ai.jpg
# a real photo edited with generative AI (inpaint / outpaint / generative fill)
exiftool -XMP-iptcExt:DigitalSourceType="$BASE/compositeWithTrainedAlgorithmicMedia" edited.jpg

Meta and Google read this field to auto-label AI content, and the EU AI Act's machine-readable-disclosure duty (Article 50, enforcement from August 2026) is pushing it from nice-to-have toward required. IPTC 2025.1 adds companion fields, AISystemUsed, AISystemVersionUsed, AIPromptInformation, AIPromptWriterName (exiftool ≥ 13.40). Full vocabulary and the retired terms to avoid: reference.md.