Back to Reference
Contents
Reference
438 words
Jun 25, 2026

AI services use embedded image metadata to ground their answers

When you upload a photograph, Heratio's metadata extractor reads the EXIF, IPTC, and XMP tags the camera and your editing software wrote inside the file. Beginning with issue #750, those tags are no longer just stored - they are now fed to the AI services as context hints so the model can disambiguate entities it would otherwise have to guess at.

What you will notice

When you run AI tools (NER, handwriting recognition, "AI Suggest Description", document extraction) against an information object that has a digital object attached, the AI sees a short prefix in its prompt:

Hints from image metadata: date=1969-07-20 20:17:40; location=28.0473,-26.2041; creator=Neil Armstrong; subjects=Apollo, Moon, NASA. Use these to disambiguate entities.

Practically this means:

  • Better dates. The NER service stops inventing dates - it sees the real EXIF DateTimeOriginal and grounds extracted date entities accordingly.
  • Better places. When GPS coordinates are embedded, the AI knows where the photo was taken and is far less likely to confuse a place name like "Paris" between France and Texas.
  • Better creator attribution. IPTC By-line and XMP dc:creator are surfaced so the AI doesn't fabricate a photographer.
  • Better subjects. IPTC Keywords already curated by a cataloguer are made visible to the AI so its suggestions stay aligned.

What the AI does not see

  • Suppressed GPS. If the Privacy team has flagged a coordinate as sensitive (issue #751 PII findings), the location hint is removed entirely from the AI's prompt. The AI cannot leak a coordinate it was never given.
  • Empty metadata. If a digital object has no embedded EXIF / IPTC / XMP - for example, a scanned page from an old film negative - no hint prefix is added and the AI sees the same prompt it saw before this change.
  • Other digital objects. Hints are computed per-file. The AI only sees hints from the specific image it is processing.

Audit trail

Every AI call that consumes hints writes an inference_context_used row to the inference receipt chain. Operators can ask, for any AI inference: "exactly what hints did the model see?" The answer is recoverable from /admin/ai-compliance/receipts.

Settings

There is no new operator-facing toggle. The behaviour is automatic: if metadata has been extracted into the property table (which ahg-metadata-extraction does on upload), the hints are visible to the AI. To stop hint injection, clear the metadata properties for the digital object.

GPS coordinates respect the existing meta_extract_gps operator toggle - if GPS extraction is disabled at the metadata layer, the AI never sees coordinates either.