Writers learn what not to write — Editorial Intelligence

I have spent a lot of time thinking about what AI should remember. I am starting to think I have been missing half of it.

I wrote something this week that I did not like.

That sounds like the least interesting possible observation from somebody who has spent most of his career writing. Bad drafts happen. You cut them, rewrite them and move on.

This one was useful because I had not written it.

I had given a detailed Sage Advice brief to Microsoft Copilot and asked it to produce a first version. The draft was competent. It followed the structure. It covered the obvious advice. There was nothing spectacularly wrong with it.

I just did not want to publish it.

I wrote about the wider experiment in The AI draft is not the article, because the useful result was discovering that a generated draft could work as a scaffold rather than as finished copy.

But one sentence from that piece has stayed with me:

It gave me something concrete to dislike.

The more I think about that, the less flippant it seems.

A surprising amount of editorial judgement is formed in opposition.

You learn what you believe partly by encountering an argument you think is wrong. You develop a voice partly by hearing language you never want to use. You learn what a good headline is partly by writing bad ones. You learn where a claim breaks because somebody challenges it. You learn to recognise empty prose after producing enough of it yourself.

Eventually you accumulate a long private list of things you do not want to do again. That may be part of what we mean when we talk about editorial taste.

And it raises an awkward question about AI. What happens when we make it much easier to avoid the mistakes that used to build the judgement?

The pain was not the point

I want to be careful here because there is an easy, nostalgic version of this argument that I do not believe.

Writers do not need to suffer.

A bad manager tearing apart somebody’s work without explanation does not build judgement. Being made to rewrite something six times because stakeholders cannot agree does not create craft. Spending years producing low-value copy simply because that is how previous generations learned is not an apprenticeship worth preserving.

But I did learn by getting things wrong. A lot.

Part of the reason I now have a reasonably strong sense of what works and what does not is that I have written things that did not work, chosen the wrong angle, buried the point, overcomplicated the structure, trusted the wrong detail and occasionally defended sentences that deserved to die.

I also learned through feedback, and that is messier than the neat apprenticeship story suggests. Some feedback was excellent. Some was subjective. Some was contradictory. Some was political. Some of it was unfair.

Learning to write professionally meant learning how to interpret that too.

An editor or stakeholder being unhappy does not automatically mean they are right. Equally, deciding that feedback is unfair does not mean there is nothing to learn from it. Sometimes the useful lesson is about the writing. Sometimes it is about the audience. Sometimes it is about the organisation, the power around the work, or the fact that two people can want fundamentally different things from the same piece.

Experience gives you a better chance of telling the difference. The useful part was never the pain.

It was the sequence underneath it:

mistake → consequence → feedback → interpretation → memory

You wrote an opening that buried the point. Somebody stopped reading.

You trusted the strongest quote in an interview and later realised it was not actually the most important one. You used a statistic because it looked impressive and then discovered it could not support the claim you wanted to make.

An editor deleted the paragraph you were proudest of because the article worked better without it. You published something technically competent that nobody cared about.

Those experiences leave residue.

The next time, you notice earlier. Eventually you do not need somebody else to tell you.

That is judgement being formed.

Writers develop negative knowledge

Most discussions of expertise focus on what somebody knows how to do. Good writers know how to structure an argument, interview a source, open an article, control pace, use evidence and understand an audience.

But experienced writers also know an enormous amount about what not to do. Do not open with three paragraphs of context when the tension belongs in the first sentence.

Do not mistake a stakeholder’s enthusiasm for evidence. Do not add a quote because a page feels as though it needs one.

Do not use a statistic just because it is large. Do not force a neat conclusion onto contradictory evidence.

Do not keep a clever line simply because it took twenty minutes to write.

Those rules are rarely written down. Many of them are not even rules. They are closer to aversions.

Something in you says: I have seen this before. I know where it goes.

That kind of negative knowledge is difficult to teach because it is partly experiential. The person has encountered enough weak versions to recognise the pattern before it fully appears.

This connects to something I have been worrying about elsewhere. In Raising the floor is not building a ladder, I argued that AI can help less experienced people produce credible work sooner while potentially removing some of the work through which people used to develop judgement.

The more precise version may be this: AI does not only remove routine production.

It can remove exposure to your own bad decisions.

A language model has no “never again”

This is where the difference between a writer and a language model feels important.

A model can identify a cliché. It can critique a weak argument. It can be instructed to act as an adversarial editor. It can compare one draft with another and explain why one is stronger.

What it does not naturally carry is the human relationship with consequences. It has not published the piece that embarrassed it.

It has not sat opposite an editor explaining why the story does not work. It has not spent an afternoon pursuing an angle and discovered there is nothing there.

It has not watched a technically perfect article disappear without trace and wondered why nobody cared.

It can represent those experiences in language. It does not possess the private accumulation of never again that writers build from them.

That is partly why a single model can feel strangely frictionless. Ask it to write a draft and it will write one.

Ask it to improve the draft and it will improve it. Ask it why the original was good and it can tell you that too.

The system does not necessarily contain an internal opposing force that says: no, this is the exact kind of thing we promised ourselves we would stop publishing. That opposition has to come from somewhere else.

This is why disagreement matters

I have already been experimenting with that problem operationally.

In Use one AI to build the argument. Use another to break it, I described using different models for different editorial roles. One develops an argument. Another is asked to attack it. The human decides what survives.

I originally thought about that mainly as quality control. Now I think it may be doing something more interesting.

The second position creates opposition.

Not independence in the strong sense. Two language models may share the same assumptions, training patterns and blind spots. An AI critic can be just as confidently wrong as an AI drafter.

But the workflow forces the work to encounter resistance.

That matters because generate-and-publish collapses several editorial stages into one. The system that produces the prose also implicitly approves it. There is no moment where somebody has to say: I do not buy this.

Traditional editorial work contained those moments everywhere. The commissioning editor rejected the pitch.

The sub-editor cut the sentence. The source corrected the interpretation.

The legal reviewer challenged the claim. The reader ignored the article.

Some of that friction was wasteful. Some of it was where the work got better.

The challenge is not to recreate every old bottleneck.

It is to work out which forms of resistance were actually carrying intelligence.

Our AI systems mostly remember the winners

This is where the question moves from writing technique into knowledge systems.

I spend a lot of time thinking about what organisations should preserve for AI.

Customer interviews. Research. Approved stories. Useful statistics. Expert commentary. Previous arguments. Strong examples. Published content. Decisions. Frameworks.

All of that makes sense. It is also heavily biased towards things that survived.

We keep the final article, not the three openings we rejected. We keep the approved customer story, not the claim that had to be removed.

We keep the campaign message, not the language that confused people in testing. We keep the polished deck, not the argument the room dismantled.

We keep best practice. We throw away the opposition that helped create it.

That may be a mistake. If AI systems are going to participate in editorial judgement rather than merely production, then an archive of good outputs is not enough.

They may also need access to what I would call, cautiously for now, negative editorial evidence.

Not a giant graveyard of every bad draft. That would create more noise than intelligence.

Something more selective: A headline that was rejected, and why.

A claim that sounded strong but could not be evidenced. A customer-story angle that stakeholders challenged because it overstated the outcome.

A phrase readers repeatedly misunderstood. An argument abandoned because new evidence contradicted it.

A draft that met the brief but still felt generic, with the diagnosis preserved. The rejection matters less than the reason.

“This was bad” teaches almost nothing. “We rejected this because the evidence only supported correlation and the copy claimed causation” is reusable judgement.

The failure log may be as valuable as the style guide

Most editorial systems are designed positively. Here is our tone of voice.

Here are our best examples. Here is the approved terminology.

Here are the templates. Here are the claims we can make.

Useful. Necessary.

But an experienced editor often carries another guide in their head: Here is the phrase that looks clever but always makes the argument vaguer.

Here is the type of customer quote that sounds good in a case study and proves almost nothing. Here is the structure that encourages us to repeat the brief rather than find a story.

Here is the point where this organisation tends to overstate what its evidence says. That second guide may be closer to judgement.

And unlike the traditional style guide, it is dynamic. It changes because the organisation keeps encountering reality.

This is why I am increasingly interested in preserving editorial decisions rather than simply editorial assets. The final content tells you what was published.

The decision history tells you why everything else was not.

There is a connection to customer intelligence

This also changes how I think about the customer-story intelligence work I have been testing. The obvious value in a completed customer story is the reusable material: proof points, quotations, themes, product evidence and narrative.

But some of the most useful fields may be the awkward ones. What evidence was missing?

What did the original interview fail to ask? What language had to be softened?

Which apparent proof point did not survive verification? What did an editor or stakeholder push back on?

Those are not defects in the record. They are intelligence about how the next piece should be better.

A mature editorial memory should not only answer:

What worked before?

It should also answer:

What did we learn not to do again?

That is much closer to how an experienced person remembers their career.

The obvious danger is institutionalising taste

There is a problem with all of this.

Yesterday’s rejection can become tomorrow’s dogma.

An editor dislikes a format once and the system records it as a rule. A stakeholder objects to a strong argument and institutional memory converts political discomfort into “best practice”. A weak experiment fails and everybody learns the wrong lesson.

Negative evidence can harden into superstition surprisingly quickly. So I would not want an AI system full of permanent prohibitions.

The useful unit is not:

Do not do this.

It is:

We chose not to do this, in this context, because of this evidence, with this consequence.

That preserves enough context for the decision to be challenged later.

Good judgement is not merely accumulated aversion. It also includes knowing when an old aversion no longer applies.

This may be what experience actually contributes

I have written before that the work happened before the prompt.

Years of writing, interviewing, editing, failing, researching and changing my mind are part of the input when I use AI now, whether or not they appear in the prompt.

I had mostly thought of that accumulated experience positively: evidence, patterns, arguments, subject knowledge, examples. I am starting to think a large part of it is negative.

Things I distrust. Structures I recognise.

Claims that make me uncomfortable. Phrases I know are covering an absence of thought.

The instinct to stop when something sounds too clean. That may explain why the Copilot draft this week was useful even though I did not like it.

The draft gave all of that accumulated opposition something to act on. The blank page asked me to invent.

The scaffold allowed me to reject. And rejection turned out to be productive.

We may need to design opposition back in

AI is exceptionally good at removing friction. That is one of its great benefits.

The danger is assuming all friction was waste. Some friction was carrying information.

Some delay was reflection. Some disagreement was quality control.

Some bad work was training. Some rejection became taste.

If we remove all of it, we may get faster production while quietly weakening the conditions under which judgement develops. I do not think the answer is to make AI worse or force people to write every first draft manually.

It is to be more deliberate about where opposition belongs. Use a generated scaffold, but require somebody to identify what they reject.

Use a second model, but give it permission to attack the premise rather than polish the prose. Keep strong examples, but preserve selected rejected decisions and the reason behind them.

Give less experienced writers AI support, but make the decisions visible so competent output does not conceal absent understanding. Build organisational memory around changes of mind, not only final answers.

That is a different kind of AI workflow. It is not designed only to produce.

It is designed to encounter resistance. And perhaps that is what I have been circling around in several pieces without quite naming it.

The words were never the whole value. The AI draft is not the article. Raising the floor does not automatically build the ladder. A second model should not merely proofread the first.

All of those arguments assume that judgement survives somewhere outside generation. The next question is how deliberately we preserve the experiences that formed it.

Because writers do not only learn what good looks like. They learn what they never want to write again.