AI Didn’t Give Me Better Writing. It Made Me a Better Writer.

What did those 4.2 million words actually do?

Last month, I published an analysis of how much artificial intelligence had contributed to my novel, Autobiography of an Artificial Mind: A novel, and summarised it on LinkedIn in early October. The results surprised me.

During the writing process, more than 4.2 million words passed between me and AI. I had 258 conversations, submitted 1,659 prompts and received 1,675 responses. I repeatedly uploaded drafts, discussed characters, explored narrative possibilities and sometimes asked AI to write entire passages or chapters. Yet when I traced the provenance of the finished manuscript, I found that just 0.65% of the novel consisted of AI-generated content expressed in AI’s own words. That’s 489 words in a novel of more than 75,000.

The finding prompted an obvious question. If AI contributed so little of the finished writing, what was the point of those millions of words of interaction? I have been thinking about that question ever since, and I now suspect that the most important contribution AI made to the project may not have been to the novel at all.

It may have been to me.

I think working with AI helped make me a better writer. But not in quite the way you might expect.

Learning to become a novelist

I began writing my novel in 2019. Over the seven years it took, I also wrote and published four non-fiction books. But writing a novel proved to be an entirely different undertaking. I understood research, argument, explanation and the construction of complex ideas. I had much less experience of the craft of fiction: developing convincing characters, sustaining narrative tension, controlling pace, constructing scenes and making readers care about what happened next.

What kick-started the entire novel was a residential writing weekend taught by Ruth Brandt in October 2019. Over the following years, I gradually became more serious about learning the craft. In June 2024, I attended the London Festival of Writing. I read John Yorke’s Into the Woods and many other books and articles on fiction-writing. Later that year, I began Jericho Writers’ Ultimate Novel Writing Course, a year-long programme involving established authors, editors and publishers.

My tutor was Helen Francis, a former editor at major fiction publishers (Faber & Faber, Penguin, Arcadia and Head of Zeus). I also joined Table 52 Writers, a weekly writers’ group whose members included published authors and people with master’s degrees in creative writing. And throughout the project, I benefited from advice from my sister, Lisa Baxter, a television drama producer with extensive experience of drama development and script editing.

These people mattered enormously. They taught me to recognise problems I previously would not have noticed. They helped me understand what makes a scene work, how narrative tension develops and why characters need to do more than simply carry a story from one event to the next. They also disagreed with one another, and that turned out to be important.

But there was another participant in my education. One that was available whenever I wanted it, that would read the same passage repeatedly, generate endless alternatives and never complain when I rejected its suggestions.

And one that was frequently wrong.

That participant was artificial intelligence. Looking back, I think I was engaged in two apprenticeships simultaneously: one with experienced, reliable but relatively scarce human teachers, and another with a fallible but extraordinarily available artificial counterpart. The human apprenticeship was high-authority but necessarily low-frequency; the AI apprenticeship was low-authority but almost unlimited in frequency. The first helped establish my standards. The second gave me thousands of opportunities to practise applying them.

What helped me make sense of the process I’d been working through was an established theory of learning I’d been deeply engaged with in my professional life in the early 2000s.

Cognitive apprenticeship: making thinking visible

In 1989, Allan Collins, John Seely Brown and Susan Newman published a chapter entitled Cognitive Apprenticeship: Teaching the Crafts of Reading, Writing, and Mathematics. Their starting point was deceptively simple. For centuries, people learned crafts by apprenticeship. An apprentice carpenter could watch an experienced craftsperson select timber, handle tools and recover from mistakes, then practise under guidance.

But what happens when the craft is largely intellectual? How does an aspiring novelist watch the thinking that produces a convincing character or a well-paced scene? Much of cognitive expertise is invisible. A finished paragraph tells us little about the rejected alternatives, the changes of direction and the judgements that shaped it.

Collins and his colleagues proposed extending apprenticeship into intellectual domains. Two years later, Collins, Brown and Ann Holum developed the argument in an article with the wonderfully descriptive subtitle Making Thinking Visible (1991). They identified six methods: modelling (an expert demonstrates, including the thinking that would otherwise stay hidden); coaching (feedback on the learner’s attempts); scaffolding (temporary support, gradually withdrawn); articulation (learners explain their reasoning, making tacit knowledge explicit); reflection (comparing approaches and judging why one works better); and exploration (learners pose and solve problems of their own).

These were not six isolated techniques. They belonged to a wider model that also concerned the order in which skills are learned and the communities in which expertise grows. Brown, Collins and Paul Duguid (1989) called this situated cognition: knowledge cannot be separated from the activities in which it is used. Learning to write is not a matter of learning rules about writing. It means applying those rules, questioning them, finding their limits and developing judgement through writing itself.

One part of the original work speaks directly to my experience. Drawing on Bereiter and Scardamalia, Collins, Brown and Holum contrasted novice writers, who use a “knowledge-telling” strategy of putting down one idea after another, with experts, who plan, evaluate and transform their ideas as they write. Expert writing is not getting ideas onto the page. It is using the page to reconsider, reorganise and develop them.

When I look back at how I used AI, much of our interaction concerned precisely this. Nearly half of my 1,659 prompts asked it to analyse the manuscript or help me work out the story; about one in five asked it for text. Not generating finished text, but transforming ideas.

The first apprenticeship: learning what good looks like

My human teachers did something that AI could not reliably do: they helped establish the standards against which my writing should be judged.

Helen Francis, for example, provided detailed feedback on my manuscript. Her June 2025 report identified problems that went far beyond sentence construction. The novel was too short. There were too few characters. Leila, one of its central human characters, was too passive. The sense of threat weakened as the story developed. These were structural problems, and solving them required reconsidering the architecture of the novel rather than merely improving individual passages.

Lisa Baxter brought a different perspective. Her professional experience lay in developing stories for television drama, where narrative structure, character motivation and dramatic tension are central concerns.

One of the most consequential questions concerned the novel’s unusual narrative structure. The story is told partly through chapters narrated by The Mind, the artificial intelligence at its centre. Lisa had argued that these chapters should form the spine of the novel. Eighteen months later, Helen suggested moving more strongly towards Leila’s perspective, potentially removing The Mind’s direct address altogether. These were serious, competing pieces of advice from people whose judgement I respected. I could not simply follow both, nor did I want to discard either.

Ultimately, I retained The Mind’s narration and expanded its role substantially. The finished novel contains 21 chapters narrated by The Mind, compared with five in the first draft. But I also used those additional chapters to address important elements of Helen’s criticism: The Mind became more complex, more threatening and more fallible. The resulting structure was neither a straightforward implementation of Lisa’s advice nor a rejection of Helen’s. It was an attempt to reconcile two different understandings of what the novel needed.

I think this is an example of what developing expertise looks like. At first, learning a craft often means trying to follow the rules. Later, it increasingly involves understanding when rules conflict, why they exist and how to make judgements that accommodate competing demands. My human teachers were indispensable to that development.

They were also part of a community. My writing group, tutor, beta readers and family brought different experiences, expectations and interpretations to the manuscript. This is an important part of cognitive apprenticeship that can be overlooked when we focus only on individual coaching. Brown, Collins and Duguid (1989) emphasised the social and cultural contexts in which knowledge develops. Expertise is not acquired in isolation; it develops through participation in a community where people share practices, disagree about standards and learn from one another. AI did not replace that community. It became another resource within it.

The second apprenticeship: learning from a fallible master

Traditional apprenticeship assumes that the master knows more than the apprentice. The learner observes expert performance and gradually develops the ability to reproduce it. But AI is a peculiar kind of master. It can produce extraordinarily sophisticated writing, and it can produce terrible writing. It can identify subtle narrative problems and then suggest solutions that introduce three new ones. It can make perceptive observations about a character while completely misunderstanding the character’s motivation. And it can deliver weak advice with the same apparent confidence as excellent advice.

So why would working extensively with such an unreliable counterpart improve my writing? I think the answer is that I was learning not simply from what AI got right, but from recognising what it got wrong.

My records provide some evidence of this process. Across the project, I typed approximately 84,000 words into AI conversations and received more than 1.45 million words in return, excluding the substantial manuscript material I uploaded. For every word I typed, AI returned approximately seventeen. That is an extraordinary volume of material to consider.

Sometimes I asked for analysis. Sometimes I requested alternative approaches to a scene, or suggestions about dialogue, character motivation or narrative structure. And sometimes I asked AI to draft entire passages or chapters. But I kept remarkably little of the generated prose. My provenance analysis found that only 489 words of the finished novel were AI-generated content expressed in AI’s own words: 0.65% of the novel’s 75,322 words.

Of course, that doesn’t mean I carefully evaluated every word AI produced. Some responses were irrelevant, some were used for research, and some may simply have been discarded. But the surviving record shows repeated cycles of generation, criticism, rejection and revision.

Here is one example. In Chapter 33, The Mind realises it has been sorting people into crude binaries. I told the AI what the passage needed to say, and it drafted this line about a would-be mass shooter:

AI draft

Harvey Silver was a mass murderer and a veteran and a man whose life had been formed by forces we had assessed but not weighted fairly.

The finished novel reads:

Final novel

Harvey Silver was poised to commit mass murder but was also a veteran who had signed up to protect the lives of the very Americans he was about to slaughter.

The AI’s version tells the reader there is a contradiction; mine makes them feel it, inside one man. Its draft had also listed The Mind’s binaries as “Smart and stupid. Right and wrong.” But a binary is precisely a choice between categories: smart or stupid, right or wrong. Even the conjunction became something to judge.

Across the passages that survived, my edits followed the same few patterns: concrete people for abstractions, aphoristic endings cut, stock gestures removed, dialogue loosened into speech. These were not random preferences. They reflected standards I was gradually developing through formal teaching, reading, feedback and practice, and AI gave me opportunities to apply them again and again.

This suggests a rather different interpretation of the master-apprentice relationship.

The master did not need to be right for the apprentice to learn.

At times, the learning came from recognising precisely why the master was wrong. But there is an important qualification. A novice without an adequate understanding of the craft might accept the same weak suggestions that I rejected. The value of a fallible master depends upon the apprentice possessing, or developing, a sufficiently reliable basis for judgement. Which brings us back to my human teachers. They helped establish the standards. AI multiplied the opportunities to practise applying them.

When tacit judgement becomes explicit

One of the most interesting discoveries in my writing archive concerns the instructions I eventually developed for AI.

By the final year, I had stopped simply reacting to its drafts. In July 2026, after a run of sessions in which I rewrote its attempts at The Mind’s chapters line by line, I turned my corrections into a written rulebook for The Mind’s voice: first-person plural, never “I”; short declarative sentences; no quotable epigrams; no lines that announce the theme instead of enacting it. Earlier, I had built and published a super-prompt for analysing narrative structure, drawing on Truby, Weiland, Yorke and Coyne.

I drafted these documents with AI, as I drafted much else. But every rule in them began as a correction I had made to something the machine wrote. And on a small sample, they worked: before the rulebook, about half the AI passages that survived into the novel contained dashes and a quarter used American spellings; afterwards, one in thirteen had a dash and none used American spellings.

These instructions were intended to improve AI’s responses. But they also required me to do something important.

I had to explain what I meant by good writing.

It is one thing to read a sentence and think that it doesn’t work. It is another to explain precisely why. Perhaps the sentence is too abstract. Perhaps it describes an emotion rather than revealing it through behaviour. Perhaps the dialogue is unnaturally polished, or the narrator has slipped into a voice that doesn’t belong to them. Turning these judgements into explicit instructions required me to articulate principles that had previously been largely intuitive.

And articulation is one of the central methods identified in cognitive apprenticeship. Collins, Brown and Holum (1991) argued that encouraging learners to make their thinking explicit can help them develop greater control over their cognitive processes. The process of explaining a judgement can expose weaknesses in that judgement. Sometimes you discover that you dislike something without knowing why. Sometimes you discover that a rule you thought was universal only applies in particular circumstances. And sometimes you realise that you have been applying a principle consistently without ever having expressed it.

Writing instructions for AI forced me to confront these questions. I was trying to teach the machine how I wanted it to write. But the exercise may also have been teaching me to understand my own writing more clearly.

Cognitive apprenticeship returns in the age of generative AI

What makes this connection particularly interesting is that researchers are beginning to revisit cognitive apprenticeship explicitly in the context of generative AI. Two studies published this year show both the promise and the limits.

Idit Adler and Yogev Shani (2026) built a personalised chatbot, TeachPal, on ChatGPT 3.5 and used it alongside a human tutor to help trainee teachers develop their practice. The hybrid approach helped, but the trainees’ discontent with the chatbot limited what it could achieve. Chao and colleagues (2026) compared an AI agent designed around cognitive-apprenticeship principles with an ordinary question-answering agent, in a study of 53 secondary-school students learning science through simulations. The apprenticeship agent’s advantage was marginal. More revealing were the patterns in how students used it: some engaged in epistemic outsourcing, handing the intellectual work to the AI, while others used it to build increasingly sophisticated explanations of their own.

Both studies are small, and neither shows that an AI apprenticeship works better than a human one. But together they make the point that matters here: the educational consequences of using AI depend partly upon what the learner actually does during the interaction. AI can provide an answer, or it can help someone learn how to arrive at an answer. Those are not the same thing, and I think the distinction extends far beyond education.

Using AI to improve the work, or the worker?

Consider two novelists, both struggling with a scene. The first asks AI to rewrite it. AI produces something better, and the novelist accepts the result and moves on. The scene has improved. But what has the novelist learned? Perhaps something. Perhaps nothing.

Now consider the second novelist. They ask AI to identify weaknesses in the scene and request several alternative approaches. They examine the strengths and weaknesses of each, reject some suggestions, challenge others and develop a solution of their own. The resulting scene may also improve, but the second novelist has done considerably more cognitive work. They have practised diagnosis, comparison, evaluation, reflection and decision-making: capabilities that may be useful the next time they face a difficult scene, and the one after that.

The distinction is between using AI to produce better work and using AI in ways that may make you better at producing the work yourself.

Of course, real interactions are rarely this tidy. During my own project, I sometimes used AI for straightforward assistance. I asked it to undertake tasks I could have performed myself, and I sometimes accepted suggestions without extended debate. There were undoubtedly occasions when I outsourced thinking. But the larger pattern of interaction involved something different. I repeatedly used AI to explore possibilities, and I repeatedly exercised judgement about those possibilities. My most productive question was often just two words:

How else?

How else could this scene work? How else might this character respond? How else could this argument be made? How else could this sentence be better? The value wasn’t necessarily in any particular answer. It was in the alternatives, and in the thinking required to decide what to do with them.

Did AI actually make me a better writer?

There is an obvious objection to everything I have said: perhaps I would have become a better writer anyway. After all, I spent seven years working on the novel. I completed a substantial writing course, received professional editorial feedback, participated in a writers’ group, read extensively about the craft and revised the manuscript repeatedly. Surely all of that would have improved my writing, regardless of whether I used AI?

I think that’s probably true, and I cannot separate the contribution of AI from those other influences. Indeed, the most intensive period of human tuition overlapped with some of my most extensive AI use.

There is another problem with the evidence. My AI conversations were recorded word for word; most of my human teaching was not. I have a detailed record of thousands of AI exchanges, but comparatively little documentary evidence of the conversations with my tutor, writing group or sister. Volume of recorded interaction is not the same thing as volume of influence. And there is no controlled before-and-after assessment of my writing ability. My archive tells me a great deal about how I worked. It cannot establish how much of my improvement was caused by AI.

So I am not claiming to have proved that AI made me a better writer. I am describing an experience and proposing an explanation that seems consistent with established learning theory.

But there is one contribution I can identify with considerable confidence: AI dramatically increased the opportunities I had to practise evaluating writing. My tutor could review a chapter. My writing group could discuss an extract. My sister could offer an experienced judgement about narrative structure. But none of them could be available every time I wanted to examine another possibility. AI could. At any hour, I could present a problem and receive something to consider. Sometimes the response was useful, sometimes banal and sometimes completely wrong. But even the weak responses gave me something against which to exercise judgement, and the volume of those opportunities was extraordinary.

That is the contribution I think cognitive apprenticeship helps explain. AI did not possess superior literary expertise, and it did not replace my teachers. Alongside them, it created a highly available environment for repeated practice in some of the cognitive activities involved in writing.

A different way of thinking about AI and expertise

This has implications well beyond novel-writing. Much of the discussion about generative AI concerns what it can do for us. Can it write reports, develop software, produce designs, analyse information or automate professional tasks? These are important questions, particularly for productivity and employment. But cognitive apprenticeship suggests another question: what happens to our expertise when we work with AI?

Consider a consultant who asks AI to write a strategy report, and another who asks AI to challenge their assumptions, identify alternative explanations, expose weaknesses in their analysis and help them explore unfamiliar possibilities. Both are using AI, but the cognitive consequences may be very different. The same applies to software developers, designers, researchers, managers and students.

AI can remove cognitive effort from a task. Sometimes that’s exactly what we want: there is little educational value in repeatedly performing routine work that we already understand. But when the cognitive effort being removed is precisely the activity through which expertise develops, we risk becoming more productive while becoming less capable of doing the work independently. Conversely, we can use AI to create more opportunities to exercise sophisticated judgement. The technology itself does not guarantee either outcome. The way we use it matters, and so does the expertise we bring to the interaction.

My experience suggests that human expertise and AI assistance can be complementary in a particularly interesting way. Human teachers, colleagues and professional communities can help establish standards, provide context and develop judgement. AI can then provide extensive opportunities to apply, test and refine that judgement. This doesn’t replace human cognitive apprenticeship. It may extend it into a new form.

And it raises a question that deserves considerably more research: could carefully designed human-AI interactions help people develop expertise faster or more effectively than they otherwise would? The early research suggests that the answer is not straightforward. Sometimes AI supports learning. Sometimes it encourages intellectual dependency. Sometimes it does both. Understanding the difference may become one of the most important challenges in the design of AI-supported work and education.

The lesson I take from seven years of writing

When I began writing Autobiography of an Artificial Mind: A novel, I knew a great deal about some forms of writing and comparatively little about writing fiction. Seven years later, I am a substantially better novelist. I owe much of that development to the people who taught me. They helped me understand the craft, challenged my assumptions and showed me what good writing could achieve.

But I also spent nearly three years working extensively with AI. During that time, I repeatedly asked it to generate possibilities, identify weaknesses, challenge ideas and propose alternatives. I accepted relatively little of its prose. I rejected a great deal. I revised constantly. And I gradually became more explicit about the standards against which I judged both its writing and my own.

Looking back, I think those interactions gave me something more valuable than better sentences or faster revisions: thousands of opportunities to practise the judgement that good writing requires. My human teachers helped me understand what good looks like. AI gave me an extraordinary amount of practice in recognising what didn’t.

I cannot prove how much that contributed to my development. But I think it mattered. And it leads me to a question I now believe everyone using generative AI should consider:

Does working with AI make you do more sophisticated thinking yourself, or less?

Because perhaps the most important measure of AI’s value isn’t how much work it can do for us. It’s what happens to our ability to do that work ourselves.