Two years ago that trailer was the interesting part. It flagged the unusual thing. I think we're close to the point where the interesting line is the other one: what did the human do?

Cambridge Dictionary seems to agree. On 17 November 2026 it named human its Word of the Year. Lexicographer Colin McIntosh's reason: "AI is changing how we talk about being human." (announcement)

The adjective is changing sides

For centuries, human creation needed no disclaimer. It was simply creation.

Then AI arrived, and we labeled the exception: AI-generated, AI-written, AI-edited, AI art, AI slop.

Now the exception is becoming ordinary, and the ordinary is becoming exceptional. So we're beginning to label the human: human-written, human-created, human-edited, human-reviewed.

Linguists call this markedness. The default is unmarked. The exception carries the label. Nobody writes "typed" on a letter, but "handwritten" is worth saying. When the default changes, the label moves:

Phase Assumed (no label) Labeled
Early generative AI made by a person AI-generated
Now, in more places drafted with a model Human Authored, human-reviewed

Cambridge's own examples show it. The Authors Guild launched a "Human Authored" mark for books. Bill Gates writes about "human reserved jobs." Andy Weir wonders if he's among the last "fully human writers." In each phrase, "human" is a qualifier on something nobody used to qualify.

The handwriting comparison has a limit worth stating. A typewriter changes the method; the person still picks every word. A model can change the author. So "human" is being asked to do a harder job than "handwritten" ever did.

"Human" is four claims in one word

When I see "human" next to AI work, it means one of these:

  • Human-authored: a person wrote the substance. Doesn't promise it's good.
  • Human-reviewed: a person checked it. Doesn't promise they wrote any of it.
  • Human-in-the-loop: a person has a step. Doesn't promise they could stop anything.
  • Human-accountable: a named person answers for it. Doesn't promise they did the work.

Every developer has seen the gap. An approved PR can be 100% model-written. A "human-in-the-loop" pipeline can be someone clicking Approve on hundreds of items an hour. The bare adjective lets the reader assume the strongest claim.

Git already solved half of this

We've had a convention for separating roles for a long time. The Linux kernel workflow uses trailers like Signed-off-by: and Reviewed-by: so that writing a change and checking it are recorded separately. AI tools slotted Co-Authored-By: into the same pattern.

What's missing is the human side at the same precision. Something like:

Fix retry backoff in ingestion worker Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Reviewed-by: Anuj Sadani Human-Edits: rewrote error handling, kept generated tests 

Human-Edits: isn't a git convention. I'm proposing the shape, not reporting a standard.

For non-code artifacts (docs, reports, generated summaries) I'd put the same thing in a provenance block in the output spec:

provenance: drafted_by: model model: claude-opus-5-5 authored_by_human: partial # none | partial | full  human_edits: "restructured sections 2-4, rewrote conclusion" reviewed_by: anuj review_scope: "facts and sources checked; tone not reviewed" accountable: anuj # a person, never a team 

Two fields do most of the work:

  • authored_by_human: partial admits the middle ground where most real work lives. The adjective can't.
  • review_scope forces the reviewer to say what they didn't check. That's the part every label leaves blank.

What a real label looks like

The Authors Guild mark is a good model because it has a process behind it. You register the work and accept a license. It covers text "fully authored by one or more human beings and not generated by AI," with a de minimis exception for spelling and grammar tools (guidelines).

Threshold, registration, named exception. That's a spec. "Human-crafted" on a landing page is a sticker.

Human, for now

My bet: in routine work the label fades the way "typed" did. Nobody will tag a generated test fixture. Where provenance carries risk (production code, legal, medical, published writing) it hardens into fields and certifications, because an adjective can't say partial and can't name a reviewer.

We used to label what machines could do. Now we're labeling what humans still do. The useful version of that label won't live in the copy. It'll live in the metadata.

Perhaps the biggest sign of AI becoming normal isn't that we stop calling it artificial. It's that we start calling ourselves human.

Drafted with AI assistance. The argument, the edits and the responsibility for every claim are mine. Long-form version with tables on my site.