# Human-made disclosure

A human-made disclosure is a producer-attached marker asserting that an artifact was made without AI. "100% written by James, not AI" in a blog footer, "all art hand-drawn, never AI" in a game's store page, "hand-coded" on a project README. It is the mirror image of the [[slop-marker-convention]]: same structure, opposite claim.

## Why it can't be evidence

Nothing verifies it. The producer can lie at zero cost, and unlike a slop marker — where the incentive runs toward silence, so declaring costs you something — the incentive here runs toward declaring whether or not it's true. A dishonest human-made label is strictly profitable. That makes the marker's information content close to zero when you consider it as a claim about the artifact.

[[no-ai-statements|Zhan]] concedes this immediately and argues the label was never doing that job. What it carries is the fact of its own display: the producer chose to spend time on something they could have shortcut, and wanted that known. Read as a stance rather than a claim, unverifiability stops mattering — you're not being told a fact about the text, you're being told an allegiance by the author. The nearest existing analogy is "Handmade in [country]" on kitchenware, which nobody audits either and which people still pay a premium for, knowing the machine version is often more precisely made.

## What it's a response to

It exists because the alternatives failed:

- **Detection.** Surface features carry no reliable signal, and trying to use them creates the degradation loop in [[ai-detector-arms-race]].
- **Voice and quality.** The "if it reads as AI, it was weak anyway" defense breaks on writing that legitimately has no voice (technical documentation, dissertations) and on the fact that models can now produce voice on request.
- **Reader burden.** Assigning the work of telling to the audience requires them to track tells that change every model release.

Once those are out, the only party who knows how the thing was made is the person who made it, and the only question left is what shape their declaration takes.

## Compared with the neighbors

Three conventions in this wiki solve the same problem with different costs and different guarantees:

| Convention | Who declares | Cost to declare | What it protects against |
|---|---|---|---|
| [[human-made-disclosure]] | author, claiming human | none | readers defaulting to "assume AI" |
| [[slop-marker-convention]] | author, claiming machine | reputational | readers spending attention unaware |
| [[credibility-as-slop-test]] | author, staking their name | reputational | bad work either way |

Human-made disclosure is the cheapest and weakest of the three, and its weakness is structural rather than fixable. Its advantage is that it needs no prior relationship — a reader who has never heard of you can still see the footer, where [[credibility-as-slop-test]] only functions if your name already means something to them.

## Limits

- **Free to fake**, and free-to-fake markers erode as they spread. If the label becomes standard, it stops distinguishing anyone.
- **Binary in a non-binary space.** "AI-assisted" covers everything from spellcheck to a model drafting the argument, and a yes/no footer can't express where a piece sits. Zhan's own version handles this by being about a category of use ("something as fundamental as communicating with fellow humans") rather than a percentage.
- **It can harden into a purity test.** The declare-human marker sits close to the refusal positions in [[anti-llm-discourse]] and can read as one even when the author, like Zhan, isn't arguing that nobody should use the tools.

## Cross-references

- [[no-ai-statements]] — the argument for it
- [[ai-detector-arms-race]] — the failure it routes around
- [[slop-marker-convention]], [[readme-not]] — the inverted convention
- [[credibility-as-slop-test]] — the reputation-staking alternative
- [[simonw-zig-anti-ai]] — the institutional-policy version of the same instinct