Before
the Prompt

Introduction

I had been fairly quiet on LinkedIn for about a week and wanted to publish another Clarity Workflows Perspective. That is the unglamorous starting point.

One topic kept showing up in my feed and in conversations about generative AI: everything is starting to sound the same. People recognise certain phrases, certain structures, the polished cadence, the generic confidence. Once you notice it, you begin noticing it everywhere.

There is evidence behind the wider concern. Doshi and Hauser found that access to generative-AI ideas improved the judged creativity of individual short stories, especially for participants who began with lower creative scores. At the same time, the stories across the group became more similar to one another.^[1]

That result interested me because it refuses a comfortable answer. AI can help somebody produce something better while the overall field becomes narrower. Both things can happen at once.

My first instinct was to write about how people could collaborate with AI without ending up in that polished middle. Then a more awkward thought arrived. How much of my own material is AI-written?

Quite a lot.

I bring the subject, argue with the responses, change direction, ask for research, reject sections and edit what survives. The prose itself often starts with the model. If I am going to criticise AI-assisted sameness, I have to include my own work in the inspection.

Curiosity, Not Commands

The conversation that produced this Perspective did not begin with a clever prompt. It began with me saying, in effect, that I thought there was something interesting here and I could not yet see its shape.

That matters more than it sounds.

I could have typed: “Give me ten original ideas for a LinkedIn article about AI.” The model would have obliged. I would probably have received several competent suggestions. One of them might even have been useful.

Instead, I brought an irritation. People were criticising generic AI content, yet many of the criticisms themselves seemed to be converging. I did not think that made the criticism wrong. I wondered what it said about us.

The conversation moved from there. At one point the AI suggested that originality belonged to the human and could not really be supplied by the model. I pushed back. That felt too comforting. AI can generate unexpected associations. It can offer an idea I had never considered. Calling all of that uncreative simply because a machine produced it seemed like a convenient way of protecting human pride.

The argument changed.

Then I asked a nastier question: how did I know the revised argument was any more original? Perhaps other people had asked almost the same thing. Perhaps the model had given similar answers. Perhaps we were congratulating ourselves for escaping convergence while standing comfortably inside it.

The argument changed again.

That is closer to what I mean by collaboration. I do not need the model to hand me a finished opinion. I need somewhere to push an unfinished thought until I discover what survives the pushing.

Point of View?

An LLM can argue both sides of a question with unnerving ease. Give it a position, ask for the strongest case, then ask it to dismantle that case. It will usually do both without the existential discomfort a human being might feel when changing sides in public.

That is useful. It also exposes the part we still have to do ourselves.

At some stage I have to choose what I am prepared to publish under my own name. I have to recognise the point where fluent language stops being a possibility and becomes my position.

That choice involves more than factual checking. I have to ask whether I actually believe the conclusion. I have to notice when a response pleases me because it confirms something I wanted to believe. I have to decide whether I would still stand behind the sentence if somebody intelligent challenged it tomorrow.

The model has no reputation to protect in that exchange. I do.

The Human Condition

A friend and I had recently been talking about the way people can attach their value to outcomes and to other people's opinions of those outcomes. The conversation had nothing to do with AI at first. It came back to me while we were developing this piece.

If I post on LinkedIn, I can see what happens. People react or they do not. Some subjects travel further than others. Certain forms of confidence look professional. Certain kinds of uncertainty look risky. None of this needs a sinister explanation. Humans have always taken cues from the group.

Social psychology has studied that pull for decades. Sometimes we follow other people because we think they know something we do not. Sometimes we adjust what we say because belonging carries value and exclusion carries a cost. Public agreement and private acceptance can diverge.^[2]

I am hardly immune to that. If I notice that a particular conversation about AI is everywhere, I know there is an audience for it before I type the first sentence. An odd thought that occurred to me while making coffee comes with no such reassurance.

That is a peculiar feature of originality. The unfamiliar idea arrives without applause attached.

It may be rubbish. That possibility deserves more attention than romantic accounts of creativity usually give it. Plenty of original thoughts are simply bad thoughts. Still, the absence of social proof makes a new idea harder to judge, and humans generally dislike uncertainty when reputation is involved.

The anxiety in that sentence belongs to us. The language model does not lie awake wondering whether its peers will think it foolish. It does not fear losing a client. It does not feel the small sting of a post that disappears without response. It has no secret plan to flatten everyone's prose into one giant beige paragraph either.

It generates from patterns in language and from the context we supply. Broad context plus a conventional objective gives it every reason to return something broadly conventional. That is usually a sign that the system followed the brief.

From A to B — Fast!

I like efficiency. I have built a business around improving workflows, so I am hardly going to campaign for ceremonial inefficiency.

Humans have been shortening routes for as long as we have been making tools. Templates save time. Machines remove repetitive labour. A good process stops people doing the same irritating thing eleven times when twice would do.

Generative AI pushes that instinct into language and thinking. I can move from intention to output with ridiculous speed.

Imagine I have just met a prospective client. I need to send a follow-up email. I tell the model who they are, what their company does, what I want from the next conversation and the tone I need. Seconds later I have a professional email that will probably do the job.

Job done.

Except the meeting contained more than the job.

Perhaps they bought the coffee. Perhaps something they said made me laugh. Perhaps I noticed their bright red shoes because they were impossible not to notice. Perhaps there was a point where the meeting stopped feeling like two businesses circling each other and became a decent conversation between two people.

The model knows none of that unless I tell it.

So the email can hit the outcome while missing the relationship. It can sound perfectly professional and still feel as though it could have gone to almost anyone.

That is where the desire to get from A to B quickly becomes interesting. The destination has become so clear that we strip away the material gathered on the way there.

Ideas suffer from the same treatment.

People have thoughts in showers. They get halfway through a run and suddenly see a problem differently. An old failure becomes relevant to something that happened this morning. Somebody says an apparently throwaway sentence and a connection appears hours later. Sometimes the useful part is the waiting.

None of that looks efficient on a project plan.

It is still where a lot of human material comes from.

Put the Friction Back

We wandered badly while developing the first version of this article.

At one point the conversation had somehow collected Jerry Maguire, Galileo, William Wilberforce, Martin Luther King Jr., Jesus, Kodak and James Dyson. That is quite a guest list for a piece that started with LinkedIn posts sounding similar.

Most of it had to go.

The detour did something useful, though. It exposed a question that the cleaner route had missed: what happens when an idea carries a social cost?

That is why I have become wary of treating every bit of friction as a defect. Some friction is obviously pointless. Nobody becomes wiser because a form asks for the same address on four screens. Other friction is cognitive. It is the period where I have not decided what I think yet.

AI is very good at rescuing me from that feeling. A blank page can become five plausible paragraphs before I have finished the cup of tea beside me.

Sometimes I should let the tea go cold.

A useful collaboration can create resistance deliberately. I can ask the model what is weak in my argument and leave it there for a while. I can ask what evidence would make me change my mind. I can ask it to defend the conventional position I am tempted to dismiss. I can ask it where the idea sounds fashionable rather than earned.

Recent experimental work points in the same direction. A preregistered study presented at CHI 2026 compared different forms of human-AI co-creation. When the model took a more proactive role in refining ideas, quality improved while diversity and perceived ownership fell. A question-led mode kept the human generating more of the thinking and produced greater diversity while maintaining quality.^[3]

Related preregistered research found a similar distinction between using AI to generate the ideas themselves and using it to refine ideas that people had already produced. AI ideation compressed collective diversity; AI refinement preserved much more of it.^[4]

I would not build a universal law from two studies. I would take the practical hint seriously.

The way we invite AI into the process changes what comes out.

Something Is Better Than Nothing

This brings me back to the simplest part of the argument.

Bring something.

It can be embarrassingly incomplete. Mine often is. “This keeps bothering me and I cannot explain why” is enough to start. So is a memory, a disagreement, a strange comparison or a story that seems unrelated until it suddenly is not.

Then make the AI work on that material.

Ask it questions before you ask it for polished copy. Let it show you the hole in the argument. If it offers a sentence that sounds wonderful and says very little, delete it. If it misunderstands the point, work out whether the failure sits in the model or in your explanation. Follow an unexpected connection for a while. Come back if it leads nowhere.

That process can look untidy in the transcript. Good.

The transcript for this Perspective contains several moments where I told the AI, in substance, “No, that is not it.” One of those moments came after we had spent a long time building an argument about conformity. The argument was interesting. It had also drifted away from the article I originally wanted to write.

Saying no brought it back.

That may be the most ordinary form of human judgement in AI collaboration. No grand genius. No mystical spark. Just enough attention to notice when the work has stopped being yours.

Original, Not Groundbreaking

I also think we make originality harder than it needs to be.

If the standard is “nobody in history has ever thought this before”, almost all of us are in trouble. Human thought accumulates. We absorb things from books, films, conversations, mistakes, work, family and whatever else has stuck to us over the years. We recombine them because that is how minds work.

My Jerry Maguire association did not invent anything. The film is thirty years old. The questions about conformity are older still. The connection mattered because it arrived in this particular conversation and changed where I looked next.

That is enough for the claim I am making here.

I do not need this Perspective to be groundbreaking. I need the thinking behind it to have passed through me rather than around me.

The Things We Think and Do Not Say

While we were talking about the fear of saying something different, Jerry Maguire suddenly appeared in my head.

That sounds ridiculous written down, which is partly why I like it.

In the 1996 film, Jerry has his famous moment of clarity and writes a mission statement called The Things We Think and Do Not Say. He prints copies and puts them in everyone's pigeonhole. The idea challenges the culture of the sports agency around him. His colleagues' reaction is complicated; the organisation's response becomes much simpler. He loses his job.

The useful part for me comes before the Hollywood redemption.

Jerry can think the thought safely. The risk enters when he gives the thought a physical form and hands it to other people.

A real business story carries a related tension. In 1975, Kodak engineer Steven Sasson developed the first self-contained digital camera. Kodak already possessed the original thinking. The invention sat inside the organisation.^[5]

Digital photography also threatened assumptions that had made Kodak enormously successful. Film sat at the heart of the business. Years later, Kodak chose not to commercialise a self-contained digital SLR that Sasson and Robert Hills had developed, with protection of film sales forming part of that decision.^[5]

I do not read that story as “Kodak was stupid and the brave innovator was obviously right.” Business decisions rarely become more intelligent when compressed into morality plays. A company has to protect revenue, judge timing and decide which inventions deserve investment. Hindsight gets to cheat.

The story still shows why originality can become uncomfortable inside a system. A new idea can ask more of us than admiration. It may ask us to reconsider the arrangements that currently reward us.

LinkedIn is a comically small stage beside Kodak, but the behavioural mechanism scales down. I can see which subjects perform. I can use the language people already recognise. I can ask AI for a contribution to an existing conversation and receive one almost immediately.

Or I can bring the odd thought that I am still slightly nervous about and see whether it survives examination.

I know which route feels safer.

Who Wrote This?

AI drafted a large proportion of the prose you have just read.

That disclosure matters even more in 03.1 because this version was created specifically to sound less like the first AI-assisted version.

There is an obvious absurdity in asking an AI system to help remove the stylistic habits associated with AI systems. I am comfortable with the absurdity. It makes the experiment cleaner.

I set constraints. I wanted to get rid of the repeated “not this, but that” rhythm. I wanted far fewer em dashes. I wanted fewer symmetrical lists and less prose that sounded as though every paragraph had been polished into a quotation card. I asked for more of the actual route we took, including the moments where the route was messy.

The model still produced sentences. I still edited, questioned and chose them. That means 03.1 cannot serve as a “human version” beside an “AI version”. Both are AI-assisted.

The difference lies in the collaboration and the editorial intention.

That is exactly why I wanted to publish both.

Perspective 03 shows what happened when we aimed for a polished, evidence-led Perspective in the established Clarity Workflows style. Perspective 03.1 uses the same argument and evidence while deliberately resisting some of the habits that had crept into the prose.

I have a suspicion readers may prefer the first one.

If they do, that would be interesting too.

Perhaps the polished patterns people complain about survive partly because they work. Perhaps some signals we now label “AI writing” overlap with techniques editors and professional writers have used for years. Perhaps knowing that AI helped changes our judgement of prose we would otherwise have accepted without much thought.

I do not know yet.

That is a better place to finish than pretending I do.