Mark What the AI Wrote: Note Provenance in the Age of Model Collapse

The blur is real, and it now has a name. Merriam-Webster made "slop" its 2025 Word of the Year: machine-made filler at scale.¹ Open your own vault and the quieter question follows: which of these notes did you actually write? Mark what the AI wrote, in a field you control.

The landmark warning comes from training, not note-taking. In Nature in 2024, Ilia Shumailov and colleagues named "model collapse": "a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting the training set of the next generation."² They measured a hard result: indiscriminate use of model-generated content in training causes irreversible defects, and the tails of the original content distribution disappear.² But hold the scale straight. A personal vault does not statistically collapse. The paper is a macro-corpus, model-training phenomenon; borrowing it for your notes is an analogy for provenance loss, the slow inability to tell what you thought from what a model drafted. The analogy earns its keep because the human fix rhymes with the machine one: keep the human-written source of truth legible and separable. You cannot separate what you never labelled.

What most people believe: an AI draft is just another note

Most people treat an AI-drafted paragraph as one more note in the pile. You pasted it, you lightly edited it, and you assume you will remember it was the model's. The belief is reasonable and, at ten notes, true. At a thousand notes, across months, memory fails, and the drafts and your own thinking merge into one undifferentiated vault.

The stakes are exactly the ones Steph Ango, Obsidian's CEO, argues for plain files. "In the fullness of time, the files you create are more important than the tools you use to create them. Apps are ephemeral, but your files have a chance to last."³ A file that outlasts its app is only worth keeping if you can still tell what is in it. A note whose authorship you have forgotten is a file that has quietly lost part of its meaning while keeping all of its bytes.

The pivot: lose the provenance, then lose the source of truth

The conventional view fails on one point: provenance is not recoverable after the fact. Once an AI draft sits unlabelled beside your own words, no later effort reconstructs which was which. The genericness creeps in quietly: un-sourced claims, borrowed phrasing, a house style that is the model's and not yours. What you cannot distinguish, you cannot trust.

The collapse paper is worth reading precisely because the effect is structural, not a quirk of one system. Shumailov and colleagues found it "can occur in LLMs as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs)" — a property of recursive generation itself, not of any single model family.² The vault analogy tracks the structure, not the statistics: a system that keeps feeding on its own undifferentiated output drifts toward the average. Your defence against the drift is the same one the researchers name for models: keep the real, human-authored data marked and in the mix.

The argument: mark what the AI wrote, in a field you own

Mark it at the source. Add one plain-text field to a note's frontmatter header, generated-by: human or generated-by: ai, and the provenance travels with the file. It is a label you write, read, grep, and export. No hidden vendor flag, no separate database. The note carries its own answer to who wrote this.

The frontmatter header is the right home because it is already where portable metadata lives, and it is human-readable. If you have not used one, the mechanism, the YAML frontmatter metadata header, is a companion piece on this blog; this essay is about one specific field to put in it. Ango's second line is the reason the field belongs in the file and not in an app: "if you want to create digital artifacts that last, they must be files you can control, in formats that are easy to retrieve and read."³

There is a standards-shaped answer to provenance already, and it is instructive that it does not fit here. C2PA, the cross-industry Coalition for Content Provenance and Authenticity, certifies "the source and history (or provenance) of media content": images, video, audio.⁴ It was built for pixels and waveforms, not for the plain-text note you keep. That gap is the argument for a text-native convention. A field in your own Markdown is the version of provenance a vault can actually use.

The honest boundary: an analogy, a debate, and a label that is not a lock

Three honest limits keep this from overclaiming. The vault does not literally collapse; that is the paper's result at model scale, borrowed as analogy. The science itself is contested. And a generated-by: field is a voluntary note to yourself, not proof — closer to a comment than a signature. Say all three plainly.

Take the science first, because the honest version is a live debate, not a settled scare. One camp shows the danger is escapable. Gerstgrasser, Schaeffer, and colleagues find that "accumulating the successive generations of synthetic data alongside the original real data avoids model collapse" — mix fresh human data in rather than replacing it, and the loop breaks.⁵ Another camp is less reassuring: Ali Borji, reviewing the same result, argues "the outcomes reported are a statistical phenomenon and may be unavoidable."⁶ Both are worth holding. The mitigation that both discussions point at, keeping genuine human authorship in the corpus and knowing which is which, is exactly what a provenance field lets you do in your own notes.

On the paper's standing: in March 2025 the authors published an Author Correction that amended a single equation typo, not the finding.⁷ A correction is not a retraction. Cite the corrected record, and neither overstate the result nor wave it away.

On the field itself: it proves nothing to anyone but you. A generated-by: marker is a self-authored label, honoured by your own discipline: anyone can write human over an AI draft, and the file will not object. Its value is not authentication; it is recall. You, and the next AI session you open over the vault, can tell the human source of truth from the machine draft. That is a discipline, not a guarantee, and treating it as more would be its own small dishonesty.

The practice: what to do tomorrow

Start with one field and one folder. Add generated-by: to the frontmatter of any note you draft with a model. Stage raw machine output in an _ai/ folder before you promote and edit it into your own words. Review that folder weekly. The discipline is boring, and boring is exactly what keeps a vault honest over years.

Four concrete moves, in order:

• Add the field to every AI-touched note. One line in the header. human when you wrote it; ai when the model did; mixed when you edited a draft into your own.
• Stage before you merge. New machine output lands in _ai/ and stays there until you have read and rewritten it. Nothing enters the main vault unmarked.
• Keep the human source of truth in the mix. The accumulation finding is your rule of thumb: never let a note become a summary of a summary of a draft. Fresh human writing breaks the loop.⁵
• Grep it when it matters. Because the marker is plain text, grep generated-by:\ ai lists every machine-drafted note in seconds — no plugin, no export, no lock-in.

The field looks like this in a note's header:

---
title: Q3 planning notes
created: 2026-07-12
generated-by: ai        # human | ai | mixed
source-session: 2026-07-12-planning-chat
---

The AI drafted the outline below. Everything under "My take" is mine.

And the staging folder keeps the unreviewed drafts visibly apart until you have made them yours:

vault/
├── projects/
│   └── q3-planning.md          # generated-by: mixed  (reviewed, edited)
├── _ai/                        # raw model output, not yet promoted
│   ├── q3-planning.draft.md    # generated-by: ai
│   └── competitor-notes.md     # generated-by: ai
└── daily/
    └── 2026-07-12.md           # generated-by: human

The input side has a mirror: a policy for what you decide not to feed the AI in the first place. Output-side marking and input-side policy are the two halves of keeping a vault yours.

Frequently asked questions

How do I keep track of what the AI wrote in my notes? Add a generated-by: field to the frontmatter of any note a model drafted — human, ai, or mixed — and stage raw output in an _ai/ folder until you rewrite it. Because the marker is plain text in the file, you can grep it, sort by it, and export it, with no plugin and no vendor lock-in.

What is model collapse, and does it affect my own data? Model collapse is the decay a generative model suffers when trained recursively on its own output; Shumailov and colleagues named and measured it in Nature in 2024.² It is a training-scale, macro-corpus phenomenon. Your personal vault does not literally undergo it; the term is an analogy for provenance loss, the creeping inability to tell your thinking from a model's draft.

How do I mark AI-generated content in Markdown? Use a frontmatter field rather than a body tag, so the label is machine-readable and travels with the file. generated-by: ai is a convention you own — not a hidden flag written by a tool you cannot read. The frontmatter header is where portable note metadata already belongs.

Does mixing real data prevent model collapse? At training scale, one line of research says yes. Gerstgrasser and colleagues find that accumulating synthetic data alongside the original real data — rather than replacing it — avoids collapse.⁵ The personal-vault translation is a rule of thumb: keep fresh human writing in the mix and never let a note become a copy of a copy.

Is model collapse real, or overblown? It is real and contested — hold both. The 2024 Nature result stands; a 2025 Author Correction fixed only an equation typo, not the finding.⁷ Since then the debate is live: one analysis argues accumulation avoids collapse,⁵ another that the effect may be unavoidable.⁶ Honest writing cites the debate, not just the scary sentence.

Isn't a generated-by: field just a comment that proves nothing? Correct, and that is the honest framing. It is a voluntary, self-authored label for your own recall, not cryptographic authenticity. Media provenance standards like C2PA exist for that, and they scope themselves to "media content" — images and video, not plain text.⁴ The field's value is that you, and the next AI session, can tell source of truth from draft.

Why not rely on C2PA or Content Credentials for my notes? Because C2PA certifies the provenance of media content — images, video, audio — by its own specification, not plain-text Markdown.⁴ A note is text you keep and edit for years. A frontmatter field is the text-native equivalent: readable, portable, and yours to maintain, rather than a signature format designed for a different kind of artifact.

Provenance is not something you reconstruct later; it is something you record while you still know the answer. Mark the note when you make it, and the vault stays legible to the one reader who matters most — the future you who has forgotten which words were ever yours.

Because your notes are plain text with an open, human-readable header, the marker is a field you add and control rather than a flag you cannot read, and that is the shape mnmnote.com is built around.

References