B18 perspective notes: 2024-09-28 (1 post)#

These notes read the batch for how Maynard thinks, not for the concepts he names. The batch is a single post, 2024-09-28 chatgpt-as-author-part-1-the-story, which I read in full, including a skim of the novella. Quotes are exact, including his typos (“Or so I though”, “Who’s byline”, “chest, hearth, throat”), apart from markdown italics, which are dropped. His prose uses curly apostrophes. His prompts to ChatGPT use straight apostrophes and double hyphens, and are quoted as they appear.

Evidence rules applied

• What counts as his. About 1,450 words of his own prose: the subtitle, the italic series note, the framing essay, the closing request and two footnotes. The roughly 420 words of prompts he wrote to ChatGPT, which the post quotes, are also his. They are the best evidence in the post of what he values in writing and how he behaves as an editor.
• What does not count. The novella The Memory Capsule (about 14,000 words, ten chapters, third draft), written by ChatGPT (GPT-4o), and the Midjourney header image. The novella’s theme was ChatGPT’s invention, not his choice: a “Capsule” that predicts and quietly steers people’s futures, and uncertainty treated as freedom. It is therefore not evidence of his views, even though it resonates with some of them. The protagonist’s name, “Andrew”, was also ChatGPT’s choice.
• Context from neighbouring posts, used sparingly and labelled. 2024-09-22 five-ai-generated-podcast-episodes-from-googles-notebooklm is where the novella first appeared. 2023-01-31 can-chatgpt-take-the-pain-out-of-annual-academic-reviews-3aa9ab32b0f0 is an earlier experiment in working with ChatGPT. Part 2 (2024-09-29) belongs to B19 and is not read or used here.

Context. Late September 2024. The previous week he had tested Google NotebookLM’s new AI “podcast” feature. He needed an original piece of fiction to feed it and asked GPT-4o for a Wyndham-style story. Over a conversation of more than 172,000 words, that request turned into a three-draft author–editor collaboration. This post publishes the result. Part 2, the next day, is his conversation with ChatGPT about how it wrote the novella.

1. How he thinks here#

An inquiry that began as a by-product: serendipity, named as such#

The whole post exists because a small practical job went sideways and he followed it. He names this twice. The subtitle calls it “an unexpected and serendipitous experiment”. The series note says the post is “the result of an unexpected but serendipitous experiment in ChatGPT’s ability to write a passable novel with some editorial support.” He was “dragged down an unexpected rabbit hole”. What he actually needed was trivial, “all I really needed was a bunch of text to feed NotebookLM”. But what he had been reading for pleasure (Wyndham “over the summer”) and a flicker of curiosity turned the job into a question. The first result “was not great. But it was interesting enough to make me wonder where ChatGPT would take things with a bit of editorial encouragement.”

The story he tells is one of curiosity gradually taking over, and he marks each step: “intrigued to see how ChatGPT would do”, “slipping into the role of human editor”, “getting fully sucked in, I thought I’d see where this might lead”, and finally “I began to feel that something interesting was happening”. He does not plan an experiment and then run it. He lets a chance opening lead him on, and only afterwards recognises it as an experiment. The seed is visible a week earlier (2024-09-22, outside the batch), where he wrote that he would “probably write about the process of working with ChatGPT to write the novel in a later article as it was surprisingly revealing.”

He finds out by building and doing, at length#

He does not argue about whether AI can be creative. He spends a weekend and a very long conversation finding out, and then shows the working. The method is iterative and hands-on:

• a first short story;
• a request for more nuance, more character and more length;
• an outline;
• a probe of how the system would stay consistent (“I was concerned that ChatGPT wouldn’t be able to remain consistent over the course of an entire novel. And so I asked how it would keep sight of the overarching narrative”);
• writing in sections, with reminders of the plan;
• three drafts, each after a round of feedback.

He also knows when to stop: “At this point I stopped the process, although I’m pretty sure that with a couple more editorial iterations ChatGPT could have produced something substantively better.”

A notable move: he works on the system’s process rather than only its output. He asks ChatGPT to set out how it will keep track of “character arcs, narrative themes and motifs”. He asks it to “update the underlying guides for the structure, narratives, sub-narratives, character development and overall shape and cadence of the novel”. Then he checks whether it follows them. His curiosity is about how the thing works, not only what it produces, and that curiosity is what leads to Part 2’s conversation about ChatGPT’s “process, aims, and understanding of itself.”

He states his prior, then lets the experience revise it#

The post opens with his old view, put bluntly: the results have “varied from bad, to kill-me-now excruciating”, and “ChatGPT, it turns out, is not a very good creative writer.” Then comes a one-line paragraph: “Or so I though.” The revision is dated and specific (“A week ago I would have said no way”), and it is measured. The experience “began to make me question my prior assumptions”, and “the quality of the novella was enough to make me revise my ideas of what might be possible with generative AI.” He revises what might be possible, not what AI is.

Analogy by structure: keep the role fixed, swap the agent#

The post’s central intellectual move is a thought experiment. He has just said co-creation is “the important point”, and then turns straight away to “However, even this caused me pause for thought.” He keeps his editorial role fixed and imagines a person in ChatGPT’s place: “If I took on the same role with a real person — providing them with initial direction and iterative feedback on form and style, but not on substance — the end novella would unequivocally belong to them. Theirs would be the byline if they published, and copyright would reside with them.”

This tests conventional wisdom against itself. The received category, “tool”, gives one answer. The same role played opposite a human gives another. When the two disagree, the category is what gets questioned: “Conventional wisdom says that this is all mine — helped by ChatGPT of course, but only in the role of tool, not a collaborator or partner.” The test he applies is concrete and turned against his own claim: “I was not responsible for the plot line, the overarching themes, or any of the specific details.”

He holds the tension open instead of resolving it#

The post piles hedge on hedge, and none of them closes the question. Co-creation is “the important point”. ChatGPT alone “would have struggled”, “although in my follow-on conversation with it in Part 2 I began to question even this.” His editing “heavily influenced the final product”, “However, even this caused me pause for thought.” The questions are left as questions: “who owns the work? Who’s byline should appear on it? Who or what was the creative force behind it?” The phrase “Who or what” leaves open even what kind of thing the creative force might be.

Fiction as a way of seeing technology#

His brief to ChatGPT describes his own approach to science fiction: a story about “the personal and social consequences of a new technology that’s developed”, which explores “the technology’s implications through the social waves it produces as seen through the lives of his primary characters.” Wyndham is his chosen model because Wyndham does what Maynard values fiction for: tracing technology through ordinary lives and social ripples. Fiction here is also the testbed. He probes an AI’s capability through the craft of storytelling, where his own judgement is sharpest.

Play and humour in the telling#

The humour is part of how he writes: “kill-me-now excruciating”; the aside about his feedback, “this is not the sort of feedback I give my students (just in case you’re wondering). But sometimes I wish I could …”; the self-deprecating “It’s not the most comprehensive feedback ever”; the exclamation marks at “a lot worse in some cases!” and “over 172,000 words!”. He also leaves an embarrassment in rather than cutting it: “I have no idea why ChatGPT named the main character after me, but embarrassing as this is, I’ve kept it in”. The experiment is fun for him, and he lets readers see that.

2. What matters to him#

• Craft and what makes a story good. His prompts are a compact account of what he values in writing: layers, nuance, “foreshadowing and non-linear story telling”; “maturity”; sub-narratives woven in so subtly “that readers need to pay attention to see these”; characters that “pop and seem even more real”; economy (“Don’t be too brief, but don’t waste words”); and restraint in editing (“tighten up the writing – but not too much as it’s already good”). He respects readers who pay attention.
• Being human as the yardstick. His sharpest criticism of Draft 1 is: “There’s no human soul to the story – it feels like you are mimicking what a human would write, but with no comprehension of what it means to be a person”. The test of good fiction is whether it shows an understanding of personhood. When he later praises the added “maturity and soul”, the same standard applies.
• Honesty about his own assumptions, and fairness in credit. He applies to ChatGPT the standard he would apply to a human author, even though that weakens his own claim to the work. Fair attribution matters more to him than keeping ownership.
• Calibrated judgement: no hype, no dismissal. “It’s still not great writing”. He lists the faults precisely: the body-part metaphors that get “rather tiresome — chest, hearth, throat and hands in particular”, “somewhat derivative and predictable in places”, the repeated plot device, gaps in coherence. Then he sets the other side of the ledger: “But I’ve read plenty of published humans who I would rank as worse than this — a lot worse in some cases!” He measures AI against real human writing, not against an ideal.
• What delights him: the moment “something interesting was happening”; the prospect of Part 2, “this is where it gets interesting”, “the follow-on conversation that really got me thinking.”
• What frustrates him, and it is mild: false self-report, with ChatGPT “telling me it was thinking about things, and then admitting that it wasn’t actually doing anything”; tiresome tics; and, going back further, years of “excruciating” AI fiction.
• Care for students. The joke about feedback works because he holds a gentler standard for students than for a machine.

3. Risk as a way of thinking#

There is no explicit risk language in this post. He does not use risk innovation, risk landscape, threat to value or orphan risks. He does not frame the authorship questions as risks. The post is written in an exploratory register, not an evaluative one. The week before (2024-09-22, outside the batch), NotebookLM had raised “red flags” for him. Here he is asking what is possible, not what could go wrong. The novella’s cautionary plot (technology that predicts and steers futures) is ChatGPT’s and must not be read as his risk view.

What the post does show is the attitude that underlies his risk thinking, applied to a different subject:

• When a technology breaks inherited categories, the mindset has to change. The pairing of tool and author is a received mental model. His structural thought experiment shows it giving contradictory answers when the agent is a system that supplied “the plot line, the overarching themes, or any of the specific details.” He does not replace it with a new rule. He opens the question: “tool, not a collaborator or partner”. The three-way distinction works as a way of widening what can be thought, not as an operational tool. This is the same move he describes as necessary for technologies that fit no earlier type. It is inference to connect it to his risk thinking, but the shape matches.
• Humility over false precision. “Now I’m not so sure”. “I suspect that part of the success of this experiment was associated with ChatGPT’s large context window”. Footnote 1 admits the likely lack of novelty: “I’m pretty sure others have gone down similar rabbit holes and that I’m just rediscovering what other people have explored here.” He reports a single case as a single case and does not generalise it into a verdict on AI creativity.
• What is at stake is value, even though he does not call it that. Byline, copyright and “creative force” are the things that give creative work its meaning and its reward. His questions show these being unsettled by a new kind of collaborator. He treats them as open questions about what we value, not as harms to be managed. This is my reading, and he does not use the threat-to-value vocabulary here.
• Capability as something that emerges from the interaction, not a fixed property. He attributes the quality to the process (“this, it turns out, played a critical part”) and to the long context window and conversation (footnote 2), not to the model alone. For thinking about AI, that means capability depends on how people and systems work together, so it is hard to assess from single outputs. He does not draw out that implication himself.

4. Scholarship and public writing#

• Experimenting in public, with the working shown. He quotes his prompts verbatim, describes each stage, publishes the whole ~14,000-word final draft and links to downloadable PDFs of all three drafts, so readers can judge the trajectory for themselves. The post reads like a lab notebook more than an essay. It is also serialised: the process was promised on 2024-09-22, this post appeared on 09-28 and the conversation on 09-29. His thinking develops across the posts, in front of readers.
• Evidence. The evidence is first-hand and experiential, and he is candid about its limits. He offers a plausible technical explanation (the context window and the 172,000-word conversation) as a suspicion, not a finding. He asks readers for evidence he could not find himself: “If you know of any, please do add them in the comments. Thanks!”
• Expertise. He draws mainly on his judgement as a reader, writer and teacher, not on machine-learning expertise. He is a physicist turned technology scholar acting as a literary editor. That lets him assess AI capability through a practice he knows well, and it moves the question from benchmarks to craft.
• Crossing disciplines. One post brings together science-fiction tradition (Wyndham), editorial craft, intellectual property (byline and copyright), questions about machine agency and personhood (“Who or what”), and what AI can currently do. It connects to his long-standing use of fiction and film to explore technology’s social consequences, and to his earlier reflections on working with ChatGPT. In 2023-01-31 can-chatgpt-take-the-pain-out-of-annual-academic-reviews-3aa9ab32b0f0 he wrote that it “intrigues me and slightly worries me that I’m sitting here already thinking of ChatGPT as a colleague and a collaborator.” The 2024 post turns that feeling into a formal question about authorship.
• Accessibility. Short paragraphs and one-line turns (“Or so I though.”, “But it still wasn’t great.”), conversational asides, humour, and direct invitations to readers. He tells readers what to read and why: the novella is “worth a read — or a skim at least”.

5. His role as he sees it#

• Explorer and honest reporter. He runs the experiment, reports what happened, including his embarrassment, and marks exactly where his view changed. He presents himself as someone who goes and finds out, not as an authority delivering a verdict.
• Readers as co-inquirers. “I’d be interested in your thoughts on it”. “Please do use the comments to let me know what you think about both the process of creation here and the resulting novella.” He asks readers to judge the novella and the process together, and to supply examples he missed.
• In relation to the AI: collaborator in practice, questioner in principle. His prompts are courteous and encouraging: “Thanks - I really appreciate the story. Could I give you some feedback though”; “This is great - I like the added complexity”; “then we’ll get into creating an incredible final draft!” He speaks of “my AI author” and uses “we” (“So we went through a third round”). He works with ChatGPT as a colleague while keeping formally open whether it is one. This fits the 2023 admission that he had felt guilty for not thanking ChatGPT.
• Students. He sets his editing of the AI against his teaching. The difference is the joke, and it shows a clear sense of what he owes students.
• What he refuses to do. He does not hype (“still not great writing”). He does not dismiss either: he retracts his earlier contempt. He does not settle an open question with false confidence, and he does not moralise about AI and creative work. There is no fear-mongering about writers being replaced, and no stand on the polarised question of AI art as theft or as genius. He keeps the discussion on what he observed and what it unsettles.
• Changes of mind, dated and public. Two are explicit. On capability: “A week ago I would have said no way”, “Or so I though.” On authorship: “This is what I would have said a week ago. Now I’m not so sure”. Read against 2024-09-22, where Draft 1 seemed to him like “emulating human mediocrity”, the revision can be traced across consecutive posts within a week. He also marks the next change in advance: Part 2 is “the game change”, “an honest conversation with ChatGPT as an LLM.”

6. What is distinctive#

1. Testing AI creativity through sustained craft, not argument or benchmark. Most AI commentary argues in the abstract over whether models can be creative, dismissing them as mimics or hyping them as creators. He runs a 172,000-word author–editor collaboration and publishes the artefact, the prompts and every draft for readers to inspect. The claim can be checked because the evidence is on the page.
2. The role reversal as a research instrument. He takes the human editor’s role on purpose, with the AI as author, and then uses that arrangement to question the concepts: what does this role make him, and what does it make the AI? The usual frame is AI as a writer’s tool or as a writer’s replacement. He puts himself in a third position and watches what happens to authorship.
3. A thought experiment that turns against his own interest. Keeping the role fixed and swapping a human in for the AI leads him to doubt his own byline. Few commentators follow the logic of authorship to a conclusion that costs them credit.
4. Seeing capability as emerging from the relationship. He places the quality in iteration, editorial direction and a long shared context, not in the model alone. That view of AI capability as a property of a human–AI system gets little space in debates built on benchmarks and single-prompt demos.
5. Serendipity admitted as method. He says openly that the most revealing experiment of his week came from a by-product of another task, and he treats following it as legitimate scholarly work.
6. Humility and humour held together. He revises in public, admits the work is probably not new (footnote 1), keeps the embarrassing protagonist name, and makes a joke of his own harshness. He is serious without being solemn, which makes his uncertainty easier for readers to share.

7. Posts that best reveal how he thinks#

The batch contains only one post, so a list of three to six is not possible.

1. 2024-09-28 chatgpt-as-author-part-1-the-story. The passages that show the most: - the series note and subtitle (serendipity named, change of mind dated); - the opening turn (“Or so I though.”); - the two quoted prompts (his standards for craft and for “human soul”); - the paragraph on the human author and editor (“the end novella would unequivocally belong to them”); - the two footnotes (humility about novelty, a mechanism offered as suspicion); - the closing invitation to judge “both the process of creation here and the resulting novella.”

Companion posts outside the batch, which should be read with this one: 2024-09-22 five-ai-generated-podcast-episodes-from-googles-notebooklm, where the experiment began, and 2024-09-29 chatgpt-as-author-part-2-the-conversation (B19), which he calls “the game change”.