ChatGPT
Perplexity
Every week your team saves more documents into SharePoint and your internal wiki. Some of those are AI-written documents, produced in Copilot and saved without a second thought. Once a file is saved, nothing inside it records which sentences a person wrote and which ones the model produced.
That would be a small filing problem if people were the only ones reading. They are not. Copilot answers questions by searching those same libraries, so whatever gets saved there becomes source material for the next answer it gives.
AI document provenance is how you find out, and it is smaller than the name suggests. One metadata column recording where each document came from. It does not exist by default, and adding it costs almost nothing.
Retrieval systems have been shown to favor machine-written text
The mechanism is documented, and it shows up at the search step, well before anything gets retrained. That matters, because retraining is not something your organization does.
Work presented at KDD found that neural retrievers rank machine-written documents above semantically equivalent human-written ones, a behavior they named source bias. On benchmark corpora, the gap ran above 30%.
A separate study at ACL traced what happens when that preference compounds. Machine-written text is retrieved, gets reused, and re-enters the corpus, and across iterations human material is progressively crowded out.
The authors call it a spiral of silence. Longer-term retrieval effectiveness declined in their simulations, though answer accuracy on the questions they tested held up, and that distinction is worth keeping straight. So the risk does not require anyone to retrain anything. It arrives through ordinary use of a system you already run.
How one drafted answer compounds
Someone asks Copilot to summarize your refund policy and saves the output into the team wiki. The next person asks a similar question, and retrieval now favors that document over the original policy page.
Their answer inherits its framing. Saved again, and the third retrieval is competing against two machine-written documents. No individual step is wrong. The corpus tilts anyway.
Stay updated with Simform’s weekly insights.
Nobody has measured whether this happens inside real companies, including yours
Here is where the honest version of this argument diverges from the alarming one. No organization has published an incident showing its AI answers degraded because machine-written material accumulated in its knowledge base.
No study measures what fraction of documents in enterprise SharePoint libraries or retrieval indexes came from a model. The evidence is academic and simulated.
The strongest counter runs deeper than that. Matthias Gerstgrasser and colleagues showed that degradation arises when synthetic material replaces real material each cycle, and that accumulating synthetic alongside real keeps the error bounded.
A SharePoint library is an accumulating estate. Human documents are not deleted when machine-written ones arrive. Related work on curated synthetic data found that lightly curated machine-written material improves models. Machine authorship is not corrupting in itself.
So the case for acting is not that your answers are getting worse. It is that you have no way to find out.
The two questions you cannot currently answer
What share of the documents your retrieval system returned this month were machine-written, and has that share moved since last quarter? Neither is answerable in your tenant today.
Every mitigation the retrieval literature validates, including down-weighting machine-written chunks and scoping retrieval by document metadata, requires knowing which documents those are.
AI document provenance ships for images in Microsoft 365, but not for prose
Your Microsoft estate is well instrumented, which is what makes the gap easy to miss. SharePoint version history records a dated copy of each save with the account behind it. Purview logs that a Copilot interaction happened and which files it touched, while eDiscovery can pull the prompt and response back out of the user’s mailbox. Sensitivity labels and trainable classifiers sort content by category, spotting source code or a resume.
Every one of those controls answers a question about custody. None answers an authorship question, and the chain snaps entirely at copy-paste-save, since knowing a user invoked Copilot at ten past ten tells your retrieval pipeline nothing about the file they saved at noon.
The asymmetry inside a single product makes the point sharply. Every image from Azure OpenAI carries tamper-evident Content Credentials built on the C2PA standard, with no setup required, and a policy setting adds watermarks to AI-generated video and audio across Microsoft 365.
Microsoft’s provenance documentation enumerates images, video, and audio. Prose appears nowhere on that list. Nobody has solved text provenance, and that is an industry-wide problem, not a gap in the platform.
Text watermarking exists, and Google’s SynthID-Text ships inside Gemini, but nothing equivalent runs in Microsoft 365 and no scheme has become the cross-vendor standard that C2PA is for images. Removal also remains straightforward for anyone who wants to.
What to do while the standard catches up
Build the marker yourself on surfaces you already own. Azure AI Search indexers extract document metadata into index fields, and fields marked filterable can be used to filter, boost, or exclude chunks at query time.
That is the mechanical home for an origin value, and it requires no new procurement.
One origin field converts an unanswerable question into a measurable one
Add a custom attribute with four states: human-authored, AI-assisted, AI-generated, and unknown. Default everything to unknown, populate it at ingestion, and mirror it as a SharePoint column so it survives outside the index.
Apply it first where the stakes concentrate, meaning customer-facing answer sources, compliance artifacts, and anything feeding a production retrieval system. The honest limitation is that nothing populates the field automatically when someone pastes Copilot output into a document and saves it, so a large share of your corpus will read unknown for a long time. That share is itself the first number worth watching.
Then measure before you exclude. Reflexively keeping machine-written documents out of the corpus is the wrong first move, since curation beats abstinence and human material needs to stay in the mix. Spend a quarter watching the share, and let the number decide the policy.
Gartner reaches the same prescription from a different direction, arguing that organizations can no longer implicitly trust data or assume it was human generated and calling for the ability to identify and tag machine-produced data as it moves through systems.
Its stated rationale is model collapse, which the counter-evidence above disputes. The instrumentation it recommends survives the dispute intact.
Who owns the field at your size
Governance guidance assumes headcount you do not have. IBM found that 76% of surveyed organizations now have a Chief AI Officer, up from 26% a year earlier, and that sample is large enterprises.
At 500 to 2000 employees, the workable translation names whoever already runs the data estate, and makes whoever deployed the AI accountable for the artifact it produced.
Gay Lea Foods, a dairy cooperative with roughly 1,200 member farms, cut monthly reporting from 24 days after close to under a day by unifying fragmented data onto a single Microsoft Fabric semantic layer: one governed dataset, standard endpoints, every consumer querying the same definitions. Structured data got that treatment years ago. The documents feeding your AI have not.
There is a commercial version of this too. The first vendor in your category who can answer a security questionnaire asking whether a human reviewed this, and answer it with a query, will put that capability next to SOC 2.
A column and an owner. That is the whole cost, and you can carry it before anyone asks. The alternative is finding out what it costs when someone does.
Simform is a Microsoft Azure Expert MSP and Fabric Featured Partner. If the corpus your AI depends on needs an owner and a governance model, our data platform modernization practice builds that discipline into the estate itself.
