Label It \"AI-Generated\" and You Will Trust Your Own Checked Note Less

A bare "AI-generated" label is read as "no human supervision" — lowering trust even in content you personally checked. Research shows naming what the AI actually did shrinks the trust penalty sharply. Here is a four-line front-matter block to replace the binary ai: true flag.¹

When researchers at PNAS Nexus asked 4,976 participants what "AI-generated" means, the most-agreed interpretation — scoring 5.11 on a 7-point scale — was that the content was written entirely by AI, with no human contribution at all.¹ If you add ai: true to a note you wrote yourself, then had an AI draft a paragraph you later fact-checked line by line, that label tells anyone who reads it — including your future self — something that is not quite true. And experimental evidence now shows the inaccuracy costs you: the reader trusts the note less, not because the content is untrustworthy, but because the label claims it was left unsupervised.

This post is about what happens when a reader — or a future you — encounters that label, and what a better one looks like. For the full disclosure block to include in a note (Used AI to: / Did NOT use AI for: / Checked by me, against the source:), see "A Note You Cannot Explain Is Incomplete."² The question here is narrower: what does the label do to the reader before they even reach that block?

What "AI-generated" means to a reader

The phrase "AI-generated" does not describe a role. It describes a production process — in the broadest possible terms. Research on news-headline labeling published in PNAS Nexus in 2024 shows what that broad term is heard as: entirely written by AI with no human supervision.

"This AI aversion is due to expectations that headlines labeled as AI-generated have been entirely written by AI with no human supervision."¹

Altay and Gilardi studied 1,976 US participants in Study 1.¹ An AI label reduced perceived accuracy of the headline by 2.66 percentage points. A "false" label reduced it by 9.33 percentage points — roughly three times as large.¹ The AI penalty is real but not catastrophic for content that is genuinely reliable. What makes it worth attending to is that the penalty is mislabeled trust: the reduction tracks the assumption of "no supervision," not the actual quality of the content.

A second study of 7,579 Americans testing an "AI-Generated" label on images found a −0.14 standard deviation decrease in belief in the labeled claims.³ That paper, by Wittenberg and colleagues, declares a competing interest (the lead author became employed by Meta after the research; two co-authors received Google funding); its result stands as independent corroboration that bare AI labels register as a small but consistent credibility discount — not because the content is defective, but because the label implies it was produced without oversight.

When naming the role shrinks the penalty

If the trust reduction comes from the assumed claim — "no human supervision" — then labeling more precisely should reduce it. Altay and Gilardi tested this in Study 2 by giving different groups different definitions of what "AI-generated" means before showing them labeled headlines.¹

The pattern was clear:

• Weak definition (AI as a tool): trust penalty b = −0.14
• Medium definition (AI as the main generator): b = −0.11
• Strong definition (AI as fully autonomous author): b = −0.02 (P = 0.60 — not statistically significant)

The labels triggered the standard skepticism response only when participants had no definition at all, or when the definition confirmed their autonomy assumption.¹

"The labels only induced skepticism when participants were given no AI definitions (just like in Study 1) or when they were given a Strong AI definition."¹

The practical implication for a personal vault: ai: true reads as Strong. It says nothing about what the AI did, nothing about how much of the note is affected, and nothing about whether a human checked it. A reader — including a future you — fills in the blank with the most common interpretation: fully autonomous, no supervision.

This is worth flagging explicitly: Altay and Gilardi studied strangers reading public news headlines. The note-keeper re-reading their own vault six months later is an inference from that mechanism, not a proven effect on personal note trust. The assumption is reasonable, but it is an assumption.

The kernel tried a binary tag — and removed it

The Linux kernel development community ran this experiment at scale. In kernel 7.0, a standard Assisted-by: trailer was added for patches generated with AI tools. By the kernel 7.3 development cycle, 1,884 of 18,533 non-merge changesets carried the tag — roughly 10 percent.⁴ Jonathan Corbet at LWN noted the actual number of AI-assisted patches was "surely much higher."⁴

"Maintainers still want to know if there is a human involved in the development process."⁴

In October 2026, maintainer Sasha Levin removed the Assisted-by: tag from the kernel's contributor toolkit. The commit message explained the problem in one sentence:

"Its absence says nothing about how a patch was made, and for developers who use these tools for almost everything, its presence says little more."⁵

What remained after the removal was not silence. The kernel's official generated-content.rst document still specifies three questions every AI-assisted patch must answer in prose: what tools were used, which portions of the content were affected, and how was it tested.⁶ And crucially, the human-only Signed-off-by trailer — the developer's certification under the Developer Certificate of Origin — was left entirely intact.

"AI agents MUST NOT add Signed-off-by tags. Only humans can legally certify the Developer Certificate of Origin (DCO)."⁶

The binary tag collapsed into noise. The process description and the human attestation survived. This is field evidence from a production system that iterated to the same conclusion the Altay study suggests: a one-bit label becomes uninformative at scale, and the useful signal is who did what — specifically, whether a human made the final judgment.

The binary generated-by: human|ai field introduced in "Mark What the AI Wrote"⁷ correctly identifies when a note has AI involvement — and that identification matters. The kernel evidence refines it: the binary alone degrades at scale, the way a flag that half of all patches carry carries less information than when it was rare. The question is what to put alongside it.

What to write instead of ai: true

The kernel's surviving fields — prose process description plus human-only attestation — suggest a parallel for a personal vault. Replace a bare ai: true with four lines that name who drafted the note, who checked it, when the check happened, and what portion the AI actually touched.

drafted-by: <tool-name>
checked-by: me
checked-on: 2026-11-06
ai-portion: "abstract and structure; all numbers verified against primary sources"

Each line does something specific:

• drafted-by names the AI role (generator of a draft) — this is the Weak or Medium definition from the Altay study, the kind that reduces the trust penalty rather than invoking the Strong-autonomy assumption.
• checked-by: me is the human attestation — the analog to Signed-off-by. Its presence asserts that a person reviewed the content. Its absence says nothing about process, just as the kernel's removed tag did.
• checked-on timestamps the attestation so you know whether the checking predates a source update.
• ai-portion describes scope — the kernel's "which portions of the content were affected" question applied to a note.

The ai-portion field doubles as a note to future self. "Abstract and structure; all numbers verified against primary sources" tells a reader — including you at six months' distance — exactly where to trust the note's shape and where to re-verify independently.

A note can be correctly marked generated-by: ai per #226 and still read as unsupervised if the label stops there. The four-line block makes the supervision visible. For the full reasoning about why explaining what you used AI for completes the note, "A Note You Cannot Explain Is Incomplete" works through the operational form in detail.²

The caveats

The studies tested strangers, not note-keepers. Altay and Gilardi studied US and UK online participants rating news headlines; Wittenberg and colleagues studied Americans rating social media images. Neither experiment tested whether the same mechanism operates when you read your own past notes. The inference is plausible but not directly confirmed.

Draxler and colleagues found, in two studies (n₁ = 30, n₂ = 96), that higher levels of a person's influence on a text increased their sense of ownership over it.⁸ A note generated entirely by AI, left unedited, may genuinely feel less like yours — not as a cognitive distortion but as an accurate perception. The goal of the four-line block is not to claim ownership you did not earn; it is to reflect it accurately when you did.

AI use also carries costs when disclosed in public contexts. Reif, Larrick and Soll found, across four preregistered experiments (N = 4,439), that people who disclose AI use "anticipate and receive negative evaluations regarding their competence and motivation."⁹ That finding is about professional disclosure to colleagues — your personal vault is a different context. But if a note travels: a research note shared with an editor, a draft reviewed by a collaborator — the label matters there too, and the four-line format gives the reader more to work with than a bare flag.

A label is not an audit. checked-by: me is a voluntary self-authored assertion, not a cryptographic certificate. It is closer to a commit message than to a signature. Its value is recall and honesty, not verification.

Frequently Asked Questions

What does a bare "AI-generated" label tell a reader? It tells them the content was most likely written entirely by AI with no human involvement. Research shows this is the dominant interpretation of the phrase — scoring 5.11 out of 7 on an agreement scale — even when the actual note was drafted by AI and then carefully checked by a person.¹

Why does the trust penalty happen even with accurate AI-assisted notes? The penalty tracks the implied claim, not the actual quality. When a reader assumes "no supervision," they discount the note regardless of whether it was checked. Labeling more precisely — naming what the AI did and who verified it — reduces the assumption and therefore reduces the penalty.¹

Does naming what AI did completely remove the trust penalty? In the Altay and Gilardi experiments, a Strong AI definition (fully autonomous authorship) made the penalty non-significant (P = 0.60). That corresponds to a Weak or Medium definition equivalent to "tool, checked by a person." It suggests the penalty can be largely erased by specificity, though the studies tested headlines, not personal notes.

What happened to the Linux kernel's Assisted-by: tag? It was removed in October 2026 (commit 1e035297ffda, Sasha Levin) because at roughly 10 percent adoption across 18,533 changesets, its presence said little more than its absence.⁵ What remained: prose disclosure of which tools were used and what they did, plus the human-only Signed-off-by attestation.⁶

Should I add this four-line block to every AI-assisted note? To any note where the AI contribution was substantial — drafted a paragraph, suggested structure, wrote a summary. For spelling corrections, grammar passes, or minor word substitutions, the kernel's generated-content.rst lists exactly that class of work as out of scope for disclosure.⁶

What if I cannot remember what the AI drafted? Write what you do know. ai-portion: "some structure, possibly some prose; not sure of the boundary" is more honest than ai: true and more accurate than silence. The goal of the field is recall, not precision.

How does this relate to the binary generated-by: human|ai field? The binary field introduced in "Mark What the AI Wrote" answers whether the note has AI involvement.⁷ The four-line block refines it by naming the role. Both are useful; neither replaces the other. A note with generated-by: ai and no other provenance metadata has identified itself as AI-involved — the four-line block fills in what that involvement actually was.

What is checked-on for? It timestamps the attestation. Sources update. If you verified every statistic in a note on a given date, checked-on tells you whether that verification predates a later revision to the source. It is especially useful for notes about research findings or technical documentation that evolves.

A bare label is a placeholder, not a provenance record. The word "generated" in "AI-generated" describes a production process without naming a role; it implies a result without describing a check. Four lines replace the placeholder with a record — who drafted, who verified, when, and what portion.

MNMNOTE stores your notes locally on your device, in open Markdown you own. mnmnote.com

References