Real works, classified

The case corpus

A scale is tested on works, not on examples it wrote for itself. These are real, published, sourced — with the walk to the number shown, so you can disagree with it.

Every claim below is checkable against the linked sources. The classification is ours, read from public reporting of how each work was made; it is never a judgment of the work's quality, and where the makers declare a level themselves, their declaration wins. A case that refuses to classify cleanly is a finding about the scale — bring it.

"Now and Then" — The Beatles

2023

The final Beatles song was completed using MAL, a neural network built for Peter Jackson's "Get Back", to separate John Lennon's voice and piano from a 1977 home demo. Paul McCartney: "nothing has been artificially or synthetically created. It's all real and we all play on it."

The walkQuestion 1: was any generative AI used? No — source separation is trained machine learning that invents nothing, like autofocus or denoise, and the non-generative edge case keeps it at Level 0. Read separation as processing of their own material instead and it is Level 1; either reading is defensible, and both sit strikingly below the "AI Beatles song" headlines of the time.

The binary label called this "made with AI" and made it a controversy. The scale calls it what it was: humans playing on a cleaned-up recording.

Wikipedia — Now and Then (Beatles song)

3 text 0image 4

"Zarya of the Dawn" — Kris Kashtanova

2022–23

A comic book with a human-written story and arrangement, illustrated entirely with Midjourney. The US Copyright Office ruled in February 2023 that the text and the arrangement of images are protectable human authorship, and the individual Midjourney images are not.

The walkFor the whole work: model-made material survives, most of the pixels are the model's, but the story, characters and arrangement — the substance — are the author's: Level 3, with Surfaces declared (text 0, image 4). The images alone land at 4, not 3, because prompting Midjourney does not materially determine the output — the same reasoning the Copyright Office used to refuse them.

The one number plus a Surfaces breakdown says in a line what took the Copyright Office a nine-page letter — and the two analyses agree.

Wikipedia — Zarya of the Dawn US Copyright Office decision (PDF)

The AI Usage Manifesto — this site

2026

The diagnosis, the argument, and every design decision in the scale are the author's. The research and the prose were produced with a large language model, then read, corrected, and signed line by line.

The walkGenerative AI used; new material survives; most of the final words are the model's; the substance — what the document says — is the author's: Level 3, declared on every page and in the metadata.

The declaration, on this site

The 21 translations of this site

2026

Every non-English page of this standard is machine-translated from the English source, and says so above the fold, in the reader's language.

The walkThe translation edge case: a faithful translation inherits the level of the source — 3 — and adds a translation note. Translating a work does not change whose substance it carries.

Twenty-one silent machine translations of a transparency standard would have discredited it on day one. The rule applied to itself is the proof it works.

The translation rule

"A robot wrote this entire article" — The Guardian

2020

The Guardian assigned GPT-3 an op-ed arguing that humans have nothing to fear from AI, feeding it a prompt and an introduction. The model produced eight essays; editors selected the best parts, merged and edited them into one, and explained the process in an editor's note.

The walkThe substance — the assigned argument — and the form both came from the model; the brief carried little work-specific substance of the editors' own. Humans chose, cut, rearranged, reviewed and published with a full note: Level 4. The editing was real, but editors' words do not dominate the result, so it is not Level 2.

A Level 4 done honestly in 2020, before there was a vocabulary for it: the editor's note is everything this scale asks for, in prose form.

The Guardian, 8 September 2020 Wikipedia — GPT-3 (the op-ed and its assembly)

CNET Money explainers

2022–23

From November 2022 CNET published dozens of AI-drafted finance explainers, bylined "CNET Money Staff", each edited by a human before publication — disclosed only behind a byline click, until reporters noticed. A January 2023 audit found serious factual errors and plagiarised passages; corrections and a public accounting followed.

The walkModel substance, model form, human review before publication: Level 4 as the process was described. What failed was the standard Level 4 sets — review must be meaningful and appropriate to the stakes, and errors that survive "editing" put exactly that in question. Sloppy review does not move a work to Level 5; it makes the Level 4 declaration indefensible, which is the point of making it.

The scale does not grade quality. It attaches a name to the review — and CNET's episode shows why readers want that name attached.

Wikipedia — CNET (AI-generated articles)

Associated Press automated earnings stories

2014–

Since 2014 the AP has published quarterly corporate-earnings stories generated by Automated Insights' Wordsmith from structured data, expanding coverage roughly tenfold over what reporters wrote by hand. Each story carries a standing note that it was generated by automation.

The walkModel-produced substance and form, published by a pipeline rather than a reader: Level 5. The template and the data feed were built with care by people — the live-generation edge case says that does not add a human review to each story.

Level 5, honest, useful, and a decade old. The level is not an accusation; the AP's standing disclosure line is exactly the plain-text declaration this scale standardises.

Wikipedia — Automated journalism

Heliograf — The Washington Post

2016

The Post's in-house bot published around 850 pieces in its first year: roughly 300 Rio Olympics reports and alerts, and coverage of about 500 races on election night 2016, faster than any desk could.

The walkAutomated generation and publication at a volume no newsroom could read first: Level 5, per story. Editors built and supervised the system; nobody reviewed each of 500 election updates before it went out.

Digiday, September 2017

Air Canada's support chatbot

2024

The airline's website chatbot invented a bereavement-fare policy — buy now, claim the discount within 90 days — and a customer booked on it. Air Canada argued it should not be liable for what its chatbot said. The BC Civil Resolution Tribunal disagreed: the company is responsible for everything on its site, chatbot output included, and owed the difference.

The walkLive generation shown to a user with no review is Level 5, however carefully the system was configured — that is the live-generation edge case verbatim.

A declaration describes the process; it never outsources the accountability. Level 5 means "no human read this first", not "no human answers for it" — a tribunal has now said the same.

The Register, February 2024

—

Sports Illustrated product reviews

2023 · undeclared

Futurism found product reviews published under authors who did not exist — AI-generated headshots bought from a synthetic-portrait site, invented biographies, no disclosure of AI anywhere. When contacted, the publisher deleted the authors, blamed a contractor, and ended the contract; people involved in the content's creation told Futurism it was AI-generated throughout.

The walkThere is no level to walk to, and that is the finding. Had the same pipeline published the same reviews under an honest Level 4 or 5 declaration, there would have been no story — the deception was not the tool, it was the fake bylines standing where a declaration should have been.

The one dishonest level is the undeclared one. This is what it looks like, and what it costs.

Futurism, November 2023