Repository navigation
AI provenance metadata in DOCX custom XML parts #1546
Description
Context
AI-generated Word documents are everywhere — reports, proposals, analyses created by LLMs. But the DOCX format carries no metadata about AI generation, trust level, or source provenance.
python-docx supports custom XML parts (customXml), which is the natural place to store this.
The Question
Is there a recommended pattern for storing custom provenance metadata in DOCX files via python-docx? Specifically:
- Can
customXmlparts be read/written through the current API? - Is there interest in a helper for AI provenance (e.g.,
document.core_properties.ai_generated = True)?
Why This Matters
EU AI Act Article 50 (August 2, 2026) requires transparency metadata on AI-generated content. DOCX is a primary format for enterprise content. Having a standard way to mark AI provenance in Word docs is becoming a compliance requirement.
Existing Approach
AKF embeds provenance into DOCX custom XML:
<akf:metadata>{"v":"1.0","claims":[{"c":"Q3 report","t":0.85,"src":"SEC 10-Q"}]}</akf:metadata>But curious if python-docx has its own approach or if this should be handled at the application level.
Activity
HMAKT99 commentedon Mar 28, 2026
Created a working example: python_docx_provenance.py — shows embedding AI provenance in DOCX custom XML using python-docx + AKF.
scanny commentedon May 18, 2026
@HMAKT99 this is a very interesting question and I'm sure is of growing importance.
Off the top of my head I'd be strongly inclined to use one of the properties mechanisms rather than
CustomXmlif it would get the job done.There is the notion of:
docProps/core.xml): "core" as in "Dublin-Core", like standard library and information-science classifiers such as created on, author, etc. This is a fixed schema of properties.docProps/app.xml), perhaps known as extended properties sometimes: These are things like company, manager, editing time, page count etc. This is also a fixed schema and not yet implemented inpython-docx.docProps/custom.xml) (not to be confused withCustomXmlfunctionality associated with structured-data tags (SDT). This schema is wide open so seems like the best fit off the top of my head.It seems like it would be important for folks to use a shared schema for AI generation metadata. It's not going to be much good if readers don't know where to find the metadata and how to interpret it.
It seems like these folks have something going in that regard: https://c2pa.org/
Looks like what they propose can't fit in a key-value pair though, so that would require a
CustomXmlpart it looks like. Also, what they're up to seems to be pretty strongly oriented toward digital media like images and video, so I'm not entirely sure how that would apply to a generated document.Regular authorship attribution could be accomplished with core-properties for author and description. And it seems like custom-properties could add a lot too, just not things like a cryptographic signature perhaps, like you might use to determine whether something had been modified since the metadata was applied.
What's your point of view on this?