Authorship Was Never the Typing

I use a language model when I write these briefs, and I do not hide it. More than one person has asked me some version of whether that troubles me. Sometimes the question arrives as concern and sometimes as an accusation, but the same expectation sits underneath both: that writing is the one place I should have kept the machine out.

I do not share that expectation, and working out why took me longer than I expected, because it turned on a question I could not answer quickly. What does it mean to be the author of something?

What We Can Actually Check

Start with what nobody has ever been able to do: look at a piece of text and see who did the thinking behind it.

The part of writing that carries the most weight never becomes words and cannot be seen at all. It is deciding what is worth saying, what is true, and which source to trust and which to set aside, and then living with that. All of it happens before there is anything to read, and by the time a text exists the evidence of how it came about has gone.

So we settled the question a different way, and we settled it long before anyone had to think about software. We ask whose name is on the work, and we hold that person responsible for what it says.

Ghostwriting is where this is easiest to see, because there the arrangement is admitted out loud. Speeches are drafted by speechwriters. Company statements are written by whoever in the press office had the afternoon free. Sometimes entire books are written by someone other than the author. None of it is secret, and none of it makes the signature void. A name on a piece of writing promises that the person will answer for what it says. It has never promised that they sat there and typed it themselves.

The standard has held. The Globe and Mail’s standards editor wrote in September that a new kind of reader email is becoming increasingly common: was this article written with AI? The same column records that the question comes in only about external contributors, never about staff bylines.

What It Costs to Answer for Something

Answering for a piece of work will not settle authorship on its own. An editor-in-chief answers for a whole newspaper without writing a line of it, and nobody calls them the author of Tuesday’s front page. So something more specific has to be doing the work.

Medical publishing has written that narrower test down, and it did so long before anyone could ask a machine for a paragraph. The International Committee of Medical Journal Editors requires an author to have contributed substantially to the conception or design of the work, or to the acquisition, analysis or interpretation of its data; to have drafted the work or reviewed it critically for important intellectual content; to have approved the final version; and to agree to be accountable for all of it.

The second condition is the one worth stopping on. Drafting is one of two ways to meet it, and the other is reading the draft closely enough to answer for what it says. Producing the words is not required, and on its own it is not enough either. Writing assistance, technical editing and language editing sit on the committee’s list of contributions that do not qualify a person for authorship, and whoever did them is thanked in the acknowledgements instead.

When chatbots arrived, the committee did not write a second test. AI writing assistance is reported in the acknowledgements like any other writing assistance, and a chatbot cannot be listed as an author, because it cannot be responsible for the accuracy, integrity and originality of the work. Both end up in the same place, the acknowledgements, and the criteria did not have to move to put them there.

What makes a brief mine is that it was mine to get right or wrong. I had the idea. I decided what was true, from sources I chose to trust, and I am the one embarrassed if I chose badly. I published it on a day when I could have decided otherwise. If a claim in this brief is false, there is no adviser who handed me the wrong figure and no reviewer that let it through. There is me, and a correction with my name at the top of it.

I spend my working life trying to find the mistake before it ships, so I have a fairly concrete sense of what carrying a consequence looks like. When it goes wrong, the question comes to you.

Where the Outlets Draw the Line

The institutions writing rules about this are protecting something real, and I would rather take on the strongest version of their case than an easy one.

Altinget published guidelines in June for the debate pieces it receives, which permit using a model for idea development and research, and for correcting spelling and grammar, while expecting the reasoning, the arguments and the formulations to remain the writer’s own. The debate editors write that the pieces they receive with heavy AI use read alike in tone, structure and language, with arguments that are often incoherent, impersonal and vague. Underneath that, Altinget states that language and thought are inseparably connected, and that in putting into words what you think, believe and stand for, language is an absolutely necessary tool. The outlet is not confused about what it is doing. It has decided that handing over the sentences means handing over part of the thinking.

The Globe and Mail goes further. Its policy prohibits the use of generative AI for writing or editing outright, and its standards editor spells out the range that covers: everything from grammar and style revisions to structural rewrites and condensing a draft to meet a word count. If it involves copying text into or out of a chatbot, it is not permitted. The same rule binds staff and freelance contributors alike.

The Globe holds the standard I have just been describing at the same time. Its policy lists accountability as a governing principle: the paper is responsible for what it publishes, and any AI process requires human oversight so that the newsroom can stand behind the result. Its accuracy principle treats any model output as a tip from an unknown source, to be confirmed independently before it contributes to anything published.

The Financial Times draws the line in the same place, and had to invoke it after publication. In August it added a note to the top of a column: AI had been used to condense a longer draft before submission, and its editorial code specifically prohibits the use of AI in the writing process.

Not every paper draws the line there. When the author of a Wall Street Journal op-ed confirmed using AI to write it, the paper’s editorial page editor defended publishing it, saying the question is whether a contribution reflects the author’s original argument and whether the author has the standing and credibility to make it.

A survey Journalisten ran in June asked eleven Danish media organisations how they intend to use artificial intelligence, and six of them answered. TV 2 is still experimenting and taking its time, and says it will use these tools in a balanced way. Set beside the Globe’s policy, the other five answers show what all of these rules share. Nobody is arguing that a named human should stop being answerable. The Globe says so explicitly while banning the drafting. DR Nyheder says the same while allowing it: a journalist there who uses DR’s in-house AI tool to draft a passage must, in DR’s own phrasing, stand behind every single word as if the journalist had written the words themselves. TV 2 Østjylland publishes no AI-generated articles or finished journalistic products, while using a model for drafts, idea development, proofreading and research, with a person reading and approving everything it has touched. JydskeVestkysten holds that it must always be people who assess, edit and take responsibility for what is published, and Nordjyske puts it as plainly as anyone, saying that a model can be a tool in producing journalism but never releases the responsible journalist or editor from editorial responsibility.

Everyone here agrees that a named human has to answer for the work. Where they divide is over whether a model may write the sentences at all. DR Nyheder allows it, and so does TV 2 Østjylland, provided a person stands behind what comes out. The Journal’s editor argued much the same way in defending an op-ed its author had written with AI. Altinget, Kristeligt Dagblad, the Globe and the Financial Times do not allow it. Their reasons differ where they give them: Altinget holds that the formulations have to be the writer’s own, Kristeligt Dagblad treats journalism made by people as a value in itself, and the Globe is protecting the voice and the meaning of the piece. I part ways with all of them, though not over whether a model can be trusted with a sentence. My disagreement is about the byline, and whether the writer who reworks that sentence is still the author of it.

Side by side, the positions look like this. Altinget’s governs the debate pieces it publishes rather than its own newsroom, the Journal’s is an editor’s defence rather than a published rule, and the medical editors’ is not about journalism at all.

Where the Line Leads

Take the Globe’s rule and apply it evenly.

Every item on its list is ordinary newsroom work when a person does it. A copy editor corrects grammar and style. An editor restructures a piece that buries its point. A sub-editor cuts a draft to fit the space. None of that has ever cost a writer the byline. The same changes made by a model are prohibited, whatever the writer does with them afterwards.

If I wrote every word of a brief myself and then asked a model to fix the grammar and cut it to length, I would have broken the Globe’s rule. If an editor did the same, nobody would say I had broken anything. My own process goes further than that, since I ask for a first draft, but the rule does not turn on how far I go. It is broken at the first grammar fix.

The medical editors ask a different question, and they ask it of everyone: what the person contributed to the substance of the work, whether they reviewed the draft critically, whether they approved it, and whether they will answer for it. A hired writer who supplied the prose and nothing else is acknowledged rather than credited, and a model is handled under the same heading. The Globe adds a condition the medical editors never needed — that a human produce the words — and applies it to one kind of help only.

The Globe does give its reasons, and they are not thin ones. The editor responsible for AI in its newsroom takes a prompt as ordinary as improve the grammar: it might swap one word for a synonym, which is fine, or it might rewrite the paragraph, which can strip out the writer’s voice and change what the sentence says. I have watched a model do the second when asked for the first, so the worry is not invented. The guidelines also keep non-public information out of these tools, outside cases the paper has approved and a senior editor has signed off. Neither reason reaches the byline. A copy editor can flatten a voice, and an editor can change a meaning while cutting a paragraph to length, and the piece stays the writer’s.

What a Second Mind Adds

The objection I find hardest to answer is that a colleague and a model are not the same kind of help.

An editor is a person, and so is a speechwriter. When a politician signs a speech that somebody else drafted, a mind did the work somewhere in the chain: someone weighed the claims, decided what was true, and would have noticed if it had not been. Hand the drafting to a model and nobody does that. A model has no view about whether a sentence is true. It has a view about which words tend to follow which.

At the drafting stage, that is right. The medical editors put the same point as a rule: a chatbot cannot be an author, because it cannot be responsible for the work. I have no quarrel with that.

It does not show that the brief stops being mine. A second mind is a real advantage: two people checking a speech will catch more than one person alone, and I do not have that. Nobody checked the draft I was handed, and I know it, which is precisely why the checking is the work. There is no adviser to point at when a figure turns out wrong. The missing mind means the checking falls entirely to me, and so does the blame when it fails.

The other half of the objection is Altinget’s, and it is the more complex one: language and thought do not come apart, so a draft is never only prose. That is true. A draft comes back with choices already made inside it. Some of them are about wording, and some of them are about what the argument actually is, and those of the second kind are judgements whether I asked for them or not. They are also easy to miss, because a sentence that already sounds finished invites you to agree with it. That is why a draft is where the work starts. If prose carried no thought, reworking it would be decoration. It carries thought, so reworking it is the part that decides whether the piece says what I mean.

How One of These Briefs Gets Made

Since the order of the work is most of my answer, here is what it looked like for The Retention Economy.

For years my Instagram feed pushed me clips I never found welcome, and marking them not interested changed almost nothing. I could not let that sit, so I did what my education pushes me towards, which is to ask why and keep asking until the answer is more than a complaint. There is no artefact for this part. Open tabs, half-read papers, and arguments I have with myself over weeks.

Eventually I had a hypothesis: the platform was doing this for a reason. I asked Claude to help me find research on something I could not yet name precisely, it pointed me towards studies, and I read the relevant sections of them. They did not confirm the hypothesis so much as sharpen it, because what the research described was an incentive structure that finds the lever on its own. That decision is one of the most fully mine in the brief, and no reader will ever see it happen. At this point, nobody has written a sentence of the brief.

Then I saw what these feeds actually do, which is to catch a passing thought and hand it straight back, the opposite of what meditation trains. That became the spine of the piece. Working the conclusion out, I remembered a half-heard study showing that screen time follows a social gradient, and Claude helped me find it again. Neither the turn to meditation nor the detour to social gradients came from a model.

Only then did I ask for a first draft, which is the step the objection is about. What came back was raw material in roughly the wrong order, and the hours after it went into composing, rewrite after rewrite. This argument before that one. This concession cut. This example moved to where it could carry its weight. Near the end I ask a model to read the draft back against the sources and argue with it, and I decide which of its objections hold. I also check all sources manually, to ensure that the model did not hallucinate. Could I have started from a blank page instead? Yes, and I have done so in the past. The hardest stretch of a brief, for me, is always the one between a finished argument and a first draft, and handing that stretch over buys the hours back for the research and the judgement and the shaping.

What the Objection Gets Right

None of this defends handing over the whole process. If I let a model decide what I think, settle what is true, and pass me an argument I never examined, the objection is right and a reader should walk away. That failure is not hypothetical.

In May, Berlingske sent a journalist home pending an investigation after Information found fabricated quotes and experts in an article about a pesticide ban. Two universities confirmed they had no staff by the names quoted. What had happened, by the paper’s own account, was that a model had been asked to summarise experts’ statements from an earlier Berlingske article, and the earlier article had, by mistake, never been loaded into the tool. Asked to summarise something it had not been given, it produced people who did not exist. The editor-in-chief said an internal tool had been used in breach of guidelines the paper considers very clear. Politiken had a comparable case a year earlier, with a fact box that attributed books to an author who had not written them.

I want to be careful about what that case proves, because it is tempting to claim too much in either direction. A named human was answerable there, and that did not stop it. The failure was in the checking, on the way in and on the way out, and the standard I have been defending is the one that was not met. If anything the case is the argument for making that checking the job. The consequence landed on the journalist, because the model could not answer for any of it, and that is why it cannot be an author.

There is a second thing the outlets ask for, and I think they are right to. DR Nyheder requires that AI use be declared to readers. Kristeligt Dagblad declares it where it is clearly relevant to them. Nordjyske puts a line at the foot of its housing-market summaries saying the article was written by Ritzau’s text robot from land registry and building register data. The Globe’s standards editor writes that where a newsroom discovers only after publication that AI was used, readers must be told. That is a fair requirement, and it is a different question from the one this brief is about. Whether a reader should know a model was involved is about honesty. Whether the work is mine is about whose judgement it carries. The medical editors hold both at once: the assistance is disclosed, and the author is still the author. Conceding the first costs me nothing. The note at the foot of every brief now says what the tool did, and this brief is the longer version of that note.

So I will keep telling readers that a model was involved. What I will not accept is that this makes anyone but me the author of the briefs on my own website.