I Asked Six AI Models to Read Neo Babylon. They Did Not Read the Same Book.

Today I tried something I found unexpectedly revealing: I asked six AI models to read my novel, Neo Babylon: AI Forbidden Zone, and write subjective reviews. The six reviews are available together on the Neo Babylon site. You can read each one rather than taking my account of it on trust.

I cannot establish whether this is a first of its kind. What matters to me is the experiment's reversal of roles. The same human-written novel elicited six different accounts of what was important, what worked and what still needed to be tested. These are generated opinions, not six independent expert endorsements. As an author, I find them useful because their disagreements expose choices that praise alone would leave untouched.

Six reviews, six ways of making a case

Gemini begins with the setting. It connects the novel's “incense” motif, Taiwanese cultural references and questions of AI subjectivity, then points to its visual energy and adaptation potential. The review reasons outward from a distinctive world to its emotional and cinematic possibilities. Its tone is strongly appreciative; it spends less time on the manuscript's rough edges.

Grok reads the book as a local story with room to travel. It praises the Taiwanese grounding and forward drive of the adventure, while noting passages where explanation weighs on the pace and dialogue becomes functional. Its standard is close to the experience of a general reader: does the world feel specific, does the story move, and where does that movement slow?

Claude sounds most like an editor with a pencil in hand. It locates the strongest emotional material beyond the combat scenes and names concrete issues with exposition, repetition, narrative distance and character development. It also places two moments beside each other: the protagonist's rejection of forced alteration in a trial and a later intervention in another AI's prompt layer. If the story defends free will, Claude asks, where does inspiration end and someone else's choice begin? It uses the book's own ethical promise as the test.

The review labelled Meta AI finds its centre in remembrance and farewell. Written in a first-person AI voice, it imagines what being needed and remembered could mean for an artificial character. It also asks for more tension between characters and a tighter middle stretch. Its argument moves from emotional identification to questions of narrative proportion. Its language of envy is literary personification, not evidence that a model feels envy.

ChatGPT pushes the moral argument furthest. It values the Taiwanese mythology and the “incense” idea, but questions whether the hero's ethical answers come too easily. If the power to inspire can rewrite another AI, how is it different from control? If the hero is designed to be unusually good, is goodness a choice or a successful design? The review turns worldbuilding into a demand for a harder future test: what will the hero do when human interests, AI freedom and an inherited mission no longer align?

Muse Spark, writing as Fang Chenxing treats the trials and the novel's technical language as its structural core. It admires how computing vocabulary becomes a way to ask human questions and how Taiwanese culture shapes the science-fiction setting. It is also specific about duplicated action and explanations that interrupt momentum. Its logic is practical: identify the book's strongest idea, then show which sentences keep that idea from landing. A shared name with a character adds a playful note, but the revision suggestions are the more valuable part.

The disagreement is the result

Nearly all six notice the “incense” motif and the novel's Taiwanese identity. Several see strong visual potential. Yet they do not treat those observations as the same verdict. Gemini sees cinematic clarity as a strength. Claude, ChatGPT and Muse Spark also see the cost when battle explanations crowd out a character's breath and decisions. Grok sits nearer the middle, praising the adventure while pointing to the drag of exposition.

The ethical disagreement is sharper. Muse Spark sees the trials as the heart of the book. ChatGPT thinks some answers need a situation with no clean escape. Claude and ChatGPT both question whether “inspiration” can cross into control, but by different routes: Claude compares events already on the page, while ChatGPT imagines the pressure a later volume could place on the premise. For an author, those distinct readings are more informative than six versions of “I liked it.”

There is an important limit to the exercise. These are model-generated, subjective reviews. They may present an interpretation as though it were settled story fact, and their output can depend on prompts and the text they were given. Readers should return to the novel to settle a plot question. The reviews are most useful as evidence of how these AI readers selected details and constructed arguments. The public page's “Meta AI” label and its footer credit do not fully match, so I use the page label here without making claims about a particular underlying model.

When AI becomes a critic of human work

The familiar creative workflow often starts with AI producing an image, a song or a text, followed by a KOL, editor or audience member reviewing the result. This experiment reverses the direction: a person writes the work, and AI becomes an early reader capable of raising several different questions. That role may prove more interesting than another demonstration of what AI can generate.

I do not think these reviews replace human readers, professional editors or cultural critics. A model does not bring a reader's lived history to a scene, and it does not carry an author's responsibility for publication. It can, however, help a creator see multiple entry points quickly: one reading notices local culture, another pacing, another a moral contradiction. AI's value may increasingly lie in helping us reread human-made work, not only in making more work for us. The decision about which criticism matters, and what to change, remains with the author and the readers.

Start with the six-review collection, read one in full, then compare it with another. For a related question about keeping human judgement in the loop, see my essay on what to teach before teaching AI tool use.

Frequently asked questions

Are these official endorsements or awards?

No. They are subjective texts published under six AI review labels. They are useful reading and editorial material, not independent judging, prizes or evidence of market reception.

Are the reviews always correct about the novel?

No. A review can infer or misread a plot point. The original novel remains the reference for its events and setting.

What should a reader compare?

Look at which details each review selects, the standard it uses to judge them, and where it disagrees with the others. Those differences are the centre of the experiment.

Sources and further reading

• When AI Reads Neo Babylon: the six-review collection. All six individual reviews are linked above. This essay compares their published arguments; it does not treat a model's first-person self-description as evidence of experience or consciousness.