Human-written content labels, explained

A content label is a short, standardised claim about how something was made, attached to the thing itself. “Human authored.” “Made without generative AI.” “AI-assisted.”

They are proliferating, and most of them mean nothing, because most of them skip the part that makes a label a label: published criteria and someone other than the maker standing behind them.

Three models, in ascending order of usefulness

Self-declared badges. The maker adds a graphic. No criteria, no verification, no consequence for being wrong. These are the most common and the least informative: the label carries exactly as much weight as the person displaying it, which means it adds nothing to what their own reputation already said.

Criteria-based certification. A third party publishes rules, the maker attests against them, the body can revoke the mark. The Authors Guild’s Human Authored programme is the clearest working example: fully human-written text, with a de minimis allowance for tools like spellcheck, and explicit permission to use AI for research, brainstorming and outlining. That specificity is what makes it a claim rather than a mood.

Cryptographic provenance. C2PA Content Credentials attach signed metadata describing how a file was created and edited. Strongest in principle, and constrained in practice by two things: it records what the signing software asserted, and platforms routinely strip metadata on upload.

The design decision everything hangs on

Does the label describe the process or the output?

Process labels (“no AI was used in making this”) are intuitive and nearly impossible to comply with honestly. AI is in the spellchecker, the camera’s noise reduction, the search engine that found the reference, the transcription of the interview. A rule that technically excludes almost everyone gets ignored, and a label everyone ignores is worse than none.

Output labels (“the text was written by a person”) are narrower and enforceable. The Authors Guild chose this, and it is why their scheme works: the question is who composed the sentences, not what software was open.

Anyone designing a label for images, music or video has to make the same choice in a medium where the boundary is much blurrier. What is the photographic equivalent of spellcheck? Computational photography means every phone image has been substantially machine-constructed before the photographer sees it.

That question does not have a settled answer yet. Whoever answers it well will set the vocabulary for their medium.

What a good label looks like

• Narrow. One specific claim, not a general virtue.
• Checkable. A registry entry, not just an image file.
• Revocable. With a stated process for complaints.
• Legible. A reader should understand it without a glossary.
• Honest about its limits. A label that claims more than it can support gets discredited by its first scandal, and takes the category with it.

Why this matters more than it looks

Labels are not really about individual works. They are about vocabulary.

Whoever defines the terms (what counts as human-made, where the line falls, what wording appears on the badge) ends up setting the frame everyone else argues inside, including regulators. That has happened before with organic, fair trade and cruelty-free, in each case long before the law caught up.

The definitional ground for human-made work is currently unoccupied in every medium except books.