BRENNEN

Industries / Media, arts and recreation / Media and entertainment / Publishing

Editing and fact-checking

Editing turns a manuscript into a finished text: shaping its structure, correcting its language to a house style, and, at many magazines, checking every fact before it is printed.

6

Milestones achieved, from 1906 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1906

    A press writes its rules

    The University of Chicago Press turned the style sheet its typesetters and proofreaders used into a 200-page book, Manual of Style, sold for 50 cents. Now in its 18th edition as The Chicago Manual of Style, it is used by writers, editors, proofreaders and indexers; the press says sales of all editions passed one million copies with the 14th edition in 1993.

    chicagomanualofstyle.org

    Credit

    University of Chicago Press

  2. 1923

    Checkers join the staff

    When Briton Hadden and Henry Luce started Time magazine in 1923, they added a research department to make sure every printed word could be verified before publication, the job now called fact-checking. The New Yorker began checking thoroughly in 1927, after an error-filled profile of the poet Edna St. Vincent Millay led her mother to threaten a libel suit.

    cjr.org

    Credit

    Briton Hadden and Henry Luce, Time; The New Yorker

  3. 2006

    A memoir meets police records

    The Smoking Gun, a website, spent six weeks checking James Frey's memoir A Million Little Pieces against police reports, court records and interviews, and reported in January 2006 that he had wholly made up or greatly embellished his criminal past and his time in jail. The book had sold more than 3.5 million copies and, by Nielsen BookScan's count, more copies in the US in 2005 than any title but the latest Harry Potter.

    LessonThe memoir's facts were tested only after it was a bestseller, and by outsiders using public records.

    thesmokinggun.com

    Credit

    The Smoking Gun

  4. 2018

    Checkers miss invented reporting

    Der Spiegel, a German news magazine known for one of the world's largest fact-checking departments, found that its award-winning reporter Claas Relotius had invented quotes, scenes and in some cases whole stories. It came to light because a co-author, Juan Moreno, re-reported a shared story at his own expense.

    LessonCheckers confirmed details that could be looked up, such as how far a town lies from Tallahassee, but the magazine's own account says the fraud lay in on-the-ground reporting, which rests on trust in the reporter.

    spiegel.detime.com

    Credit

    Juan Moreno; Der Spiegel

  5. 2023

    Dahl's words kept in print

    Puffin, the children's imprint of Penguin Random House in the UK, issued new editions of Roald Dahl's books with what The Daily Telegraph counted as hundreds of changes to wording, made with The Roald Dahl Story Company and the group Inclusive Minds. After an outcry, the publisher announced on 24 February 2023 that 17 titles would also stay in print unchanged as The Roald Dahl Classic Collection, and Dahl's publishers in the US, France and the Netherlands said they would not use the changes.

    LessonRewording a dead author for modern readers met enough resistance that the publisher kept both versions on sale.

    npr.org

    Credit

    Penguin Random House Children's UK

  6. 2025

    An AI writer gets through

    On 7 May 2025 WIRED published a feature pitched by a freelance writer that turned out to be an AI fabrication. When editors grew suspicious, two outside AI-detection tools judged the text likely to be human-written. WIRED retracted the story and said it had not gone through a proper fact-check or a senior editor's review, which first-time contributors should get.

    LessonThe detection tools missed the fake; the standard checks for a new writer, skipped this time, were the safeguard.

    wired.com

    Credit

    WIRED, in an account published on 21 August 2025

Watch for

  1. Screening for AI before print

    Press Gazette, a UK magazine about journalism, launched Made by Humans with the software company Flip-Pay and began testing it in its own publishing system in September 2026. The tool labels text, images, audio and video "AI Generated", "AI Assisted" or "Human Made" before publication, using Pangram's detector for text; Pangram says its latest model gives one false positive in 10,000 cases and can tell lightly edited human writing from generated text.

    Achieved whenan outside test on edited magazine copy, including human writing tidied with AI tools, publishes the tool's error rates.

    inpublishing.co.ukmediacopilot.ai

    Claimed by

    Press Gazette, Flip-Pay and Pangram, September 2026

Open

  1. 01

    Planted errors for copy editors

    A public test for AI copy-editing tools. Take book chapters, plant known errors of each kind a copy editor is paid to catch (spelling, slips from the house style, a character whose name or eye colour changes, a wrong date, a repeated word, figures that do not add up), and score each tool, and two professional copy editors, on how many errors they find and how many correct sentences they change.

    The brief
    Credit · unclaimed
  2. 02

    Quotes traced to their sources

    A public test of tools that check a nonfiction book's quotations and notes against the sources they cite. Take 20 public-domain or openly licensed nonfiction books whose notes point to sources that are themselves free to read, plant 50 errors in each (a quotation altered or given to the wrong person, a figure changed, a claim the cited source does not make), and score each tool, and a professional fact-checker on a sample, on how many planted errors they flag and how many sound passages they flag by mistake.

    The brief
    Credit · unclaimed

Open 01 · the brief

Planted errors for copy editors

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test for AI copy-editing tools. Take book chapters, plant known errors of each kind a copy editor is paid to catch (spelling, slips from the house style, a character whose name or eye colour changes, a wrong date, a repeated word, figures that do not add up), and score each tool, and two professional copy editors, on how many errors they find and how many correct sentences they change.

Why it is still open

Copy-editing tools reach publishers on their makers' claims. The nearest public benchmark, EditPropBench (Garvin Kruthof, May 2026), tests whether AI editors carry a factual change through the rest of a manuscript; it uses synthetic machine-learning papers, not books, and does not score style or consistency. FlawedFictions (Kabir Ahuja, Melanie Sclar and Yulia Tsvetkov, 2025) plants plot holes, such as a fact that a later passage contradicts, into human-written stories, and found that leading models did poorly, and worse as stories grew longer; it does not test spelling, house style or the needless changes an editor must avoid. This research found no public test of AI tools against copy editors on book text.

arxiv.orgarxiv.org

Built with what exists

Chapters from public-domain books, from openly licensed recent fiction and some written for the test, so that no model has seen every text; a short house style sheet written for the test; a script that plants errors and keeps the answer key; AI models through their public interfaces; Python for scoring.

  1. First step

    Plant 50 errors in one public-domain chapter and run two AI tools over it.

  2. Done looks like

    20 chapters with 50 planted errors each; for at least three AI tools and two copy editors, the share of errors found by type and the number of needless changes per 1,000 words; chapters and answer key released openly.

Open 02 · the brief

Quotes traced to their sources

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of tools that check a nonfiction book's quotations and notes against the sources they cite. Take 20 public-domain or openly licensed nonfiction books whose notes point to sources that are themselves free to read, plant 50 errors in each (a quotation altered or given to the wrong person, a figure changed, a claim the cited source does not make), and score each tool, and a professional fact-checker on a sample, on how many planted errors they flag and how many sound passages they flag by mistake.

Why it is still open

Book publishers rarely check facts. Emma Copley Eisenberg reported for Esquire in 2020 that in almost all book contracts the author, not the publisher, warrants that the text is accurate, that most nonfiction books are not fact-checked, and that a thorough check by a freelance checker costs about $8,000 to $10,000. In May 2026 The New York Times found more than half a dozen misattributed or invented quotations in Steven Rosenbaum's The Future of Truth (Matt Holt/BenBella), whose author said he had used ChatGPT and Claude in research and writing. Tools exist: citefact, an open-source program first released in August 2026, checks that each direct quotation appears word for word in the cited source and asks an AI model whether the source supports each claim, but it reads only author-year citations, not the numbered notes most trade books use, and publishes no accuracy figures; Der Spiegel's fact-checking department has built a prototype that pulls claims from articles and checks them with web searches, also without published accuracy figures. This research found no test of such tools on book-length nonfiction.

yahoo.comkirkusreviews.compypi.orgjournalists.org

Built with what exists

citefact, with the books' notes converted to author-year form; an AI model that reads long documents and searches the web, such as Claude Fable 5.2 (see the Note "Three working devices"), or an open-weight model; cited sources from Project Gutenberg, the Internet Archive and open-access journals; each book run several times, since one pass can hide unstable answers (see the Note "All twenty times"); Python to plant errors and score.

  1. First step

    Convert one public-domain nonfiction book whose sources are free to read into citefact's input format, plant 20 errors, and count how many it flags.

  2. Done looks like

    20 books, 1,000 planted errors and the answer key released; for each tool, the share of planted errors found by type and the false flags per 100 pages; on two books, the same measures for a professional fact-checker, with hours and cost set beside each tool's cost per book.