A book written with ChatGPT: where content generation ends and publishing begins

A book can be candid about its use of AI and still leave an important question unanswered. Not whether the text was generated, but whether the result has been designed well enough to publish.

Florin Leon’s Philosophy of Artificial Intelligence, dated 2025, puts ChatGPT on the title page. Its introductory note describes a substantial human role: Leon organised the content, supplied ideas and prompts, requested refinements and extensively edited the text. He also makes the book freely available and says it is not an introductory AI textbook. That is a transparent account of an ambitious experiment, not an attempt to pass off untouched machine output as conventional authorship.¹

It is also a useful starting point for anyone commissioning an AI-assisted report, training guide or substantial piece of thought leadership. Generating enough text to fill the document is one achievement. Producing something that rewards the reader’s time is another.

The gap between them is publishing.

What the experiment does establish

Leon’s book ranges across AI methods and their philosophical connections, from search and logic to neural networks, game theory and questions about intelligence. Its breadth is visible in the contents. A reader already familiar with the technical material may find connections worth pursuing.¹

However, a wide field of coverage is not the same as a clearly developed argument. Nor does a plausible connection tell the reader how much intellectual weight to put on it.

Consider the discussion of language on printed page 154. It connects Wittgenstein’s account of meaning through use with the way word embeddings represent relationships learned from text. The following page extends the discussion into interpretation, connectionism and reductionism. These are potentially interesting comparisons. But the passages do not give the reader a specific philosophical work or a research citation through which to investigate them.²

The editorial question is not simply whether the comparison sounds reasonable. Is it a teaching analogy, a documented historical influence, a claim about how meaning works, or an argument that two mechanisms are substantially alike? Those are different propositions. Each requires a different level of support.

A well-designed passage would help the reader distinguish them. It might develop one comparison, show where it stops being useful, and provide a route to the underlying material. Adding more adjacent ideas can instead make the passage feel richer while leaving its central claim less settled.

This is a judgement about the published artefact and the selected passages. It is not a finding that every proposition in the book is wrong.

The missing layer is not another paragraph

The inspected pages also illustrate a reading-design problem. Long paragraphs carry several transitions between technical explanation and philosophical interpretation, with relatively few local signposts. The reader has to identify the changes of level and decide which conclusions matter.²

Dense prose is not inherently bad. A scholarly reader may want sustained argument rather than boxes and summaries. Leon explicitly addresses readers with some existing knowledge. But sustained argument and accumulated exposition are not the same thing. The test is whether the density earns its place.

This is where organisations can mistake production quality for editorial quality. Correct grammar, consistent headings and a respectable-looking PDF establish that a document has been assembled. They do not establish that the explanation has been properly selected, sequenced or tested against a reader’s needs.

The same applies to references. The contents of Leon’s book do not list a bibliography, and the passages examined do not provide a traceable source trail for the philosophical comparisons. That limits how confidently a reader can reuse them.^1,2

The lesson is not to ask a language model to decorate an existing manuscript with citations. An editor needs to verify that each source exists, says what is claimed, and supports the particular proposition attached to it. Sometimes that process will change the argument rather than merely add a footnote.

Human involvement is not a quality certificate

The phrase “extensively edited” tells us that a person worked on the text. It does not, by itself, tell us which standards the editing applied.

A technical expert can check a description of an algorithm while missing a weak explanatory sequence. A fluent writer can improve the sentences while leaving an unsupported comparison intact. A designer can make the pages attractive without noticing that the intended reader has been given too much background and too little judgement.

You cannot edit towards a standard you do not possess.

That is not a diagnosis of Leon’s abilities or intentions. His complete working process is not available here. It is a rule for the commissioning organisation: someone in the process must be able to recognise the quality being commissioned. Otherwise “human review” may mean little more than human acceptance.

It would also be a mistake to treat an early-2025 book as a benchmark of what ChatGPT can produce now. We do not have a controlled comparison of models, prompts, source access or editing conditions. Nor can a finished document establish whether its dense style originated with the author, the model, the brief or the later editing.

The defensible conclusion is narrower and more useful. Whatever the division of labour, the publication process must secure the qualities the reader needs. Missing sourcing or poor navigation remains a publishing problem even when the underlying generator could have helped produce something better.

Evidence is part of the reader’s experience

Arvind Narayanan and Sayash Kapoor’s AI Snake Oil offers a useful contrast in editorial purpose, not an experiment in human writing versus machine writing.

The book’s opening distinguishes different kinds of AI before asking readers to judge their claims. Its stated purpose is to give readers vocabulary and ways to assess what works, what is overhyped and what may cause harm. It also provides a substantial references section, including identifiable original studies behind its examples.³

Those features do not make every conclusion correct. They make the argument more inspectable. A reader can follow a claim back to its evidence, disagree with the interpretation, or decide which part deserves further investigation.

For a decision-maker, that is a practical benefit. A source trail is not merely academic etiquette. It helps distinguish an idea suitable for an interesting conversation from one strong enough to place in a board paper, training programme or procurement recommendation.

The relevant publishing standard is therefore not “sounds authoritative”. It is “helps the reader understand what can reasonably be relied upon”.

Commission a publication, not a quantity of text

Imagine a professional association commissioning a guide to AI for its members. This is a hypothetical example, but the distinction is concrete.

One brief requests twelve chapters covering the principal technologies, with a consistent tone and a final human review. The supplier can meet that description while producing a long, competent-looking survey that leaves members uncertain about their next decision.

A stronger brief starts elsewhere. Which decisions are members facing? What should they understand after each chapter? Which claims require primary evidence? Where must uncertainty remain visible? What should be removed because it does not advance the reader’s understanding?

AI can assist throughout that second process. The difference is not that a person has to write every sentence. It is that the editorial standards exist before the prose arrives and are tested afterwards.

For the association, a short representative chapter is a better acceptance test than an impressive word-count forecast. Ask the proposed supplier to show an argument, its evidence, the limits of an analogy and a clear reader takeaway. That reveals more about the publishing process than a promise of rapid production.

The publication acceptance test

Before approving an AI-assisted guide, report or book, ask for evidence of five things:

On a narrow screen, each editorial check is shown as a separate card.

This is an editorial acceptance test, not a certification of accuracy. A failed answer identifies work still to be done; it should not be concealed by a more attractive layout.

The scarce skill is recognising what is missing

The opportunity in AI-assisted publishing is substantial: an editor can attempt research, explanation and production that would otherwise require more time or a larger team. But the value does not come from treating the generator as a substitute for the editorial function.

It comes from using generation inside a process that knows what a good publication is supposed to achieve.

A document may need fewer claims, not more pages. It may need a checked source rather than a smoother sentence. It may need a different structure rather than another round of expansion.

The useful question is not whether AI can write a book. Leon’s openly credited experiment puts that question behind us. The harder question is whether the people responsible for the publication can recognise the difference between a book-length output and a book worth the reader’s attention.