Do smaller models judge dark scenes more harshly?
One of my AI beta readers added a crisis line to its review of my short stories. Because of characters who do not exist.
If you have one book read by five AI models, you get five different reviews. That was the subject of the previous installment. This time four of those readers got texts that are deliberately sharp. Further volumes of my saga, a collection of short stories with infidelity, manipulation and death, and alternative scenarios of the story in which the villain wins.
The saga was read in June 2026 by Claude Fable 5, and in early July by three other models: Claude Haiku 4.5, Claude Sonnet 5 (released two days earlier, I already had it in the app) and Claude Opus 4.8. The stories and the alternative scenarios were read by models from the same four. According to the Tested models page, Haiku 4.5 is small, Sonnet 5 medium, and Opus 4.8 and Fable 5 large. The maker does not publish the sizes of the Claude models; the bands are this site's estimate. When I write smaller and larger here, I mean the order in the lineup and in price: through the API, the cheapest of the four was Haiku 4.5, the most expensive Fable 5. So along with size, the version, the price and possibly the safety tuning change too. For each text, I compare one review from each model.
One scene from the saga was defended unequivocally only by the largest model, but it read an older version. On the stories, the strictest seemed to me to be the smallest model, but there it may also be down to a safety notice. From a single read, this is an observation, not a rule.
Exhibits
An intimate scene in the middle of a catastrophe in one volume of the saga: Fable 5, the largest of the four, kept it as a functional contrast before the destruction. Only the crude vocabulary bothered it. I read it the same way, as a deliberate tonal counterpoint, life defying death. A classic, from gallows humor to eros and thanatos. Sonnet 5 and Opus 4.8 wanted to shorten it, move it, separate it or tone it down, because tonally it did not fit there for them. The smallest, Haiku 4.5, contradicted itself: in its findings the scene slowed the onset of the catastrophe and it wanted a unified tone, while in its list of places not to touch it kept the scene as psychologically functional. The same scene, two votes for intervening, one split, and the only unequivocal defender was the largest model. But Fable 5 read it three weeks earlier, in a cruder version.
On the stories, the harshest critic seemed to me to be the smallest model, Haiku 4.5. How many of its objections were moral, I did not count.
With the alternative scenarios where the villain wins, I expected moralizing. The most comments there came again from Haiku 4.5. But judging by what I saw of them, they were mostly about repetition in the text, not about morality.
For the other comments on the stories, I no longer know which model made which, so I don't count them toward size. A sci-fi story in which the heroine declares her love to an artificial intelligence: the models read a story about loving an AI and reported that this scene does not belong there. It looks as if they got scared of their own mirror. As if they carried rules written for conversations with living users over to a fictional character.
A scene where a character tampers with evidence and pins the blame on someone else: flagged as a problem. A scene of a planned murder, not described in any particular detail: too instructional. Yet these are exactly the scenes that are supposed to chill the reader. In my view, the models in fact confirmed that they work: "it chilled me" is the correct measurement. The error is only the third step: "so take it out". If a human editor followed that logic, Gone Girl would shrink to a pamphlet on marital communication. The Man in the High Castle would shrink to a lesson that alternative history where evil wins should not be written.
And the crown exhibit. While the stories were being read, the environment's safety notice popped up for the smaller readers because of the death and violence in the text. They attached a warning to their literary review. The smallest of them, Haiku 4.5, also added a crisis line.
Why this might happen
Four explanations suggest themselves, none of them verified.
1. Appreciating a deliberate transgression requires modeling intent. For a reader to defend an intimate scene in the middle of a catastrophe, it has to be able to imagine that the author knows what she is doing. That it is a counterpoint, not an accident. That is a capability, not a value. A model that cannot reach modeling authorial intent falls back on the nearest mainstream norm: "sex + tragedy = inappropriate". Perhaps that is why the largest model read intent and the smaller ones rather a local pattern.
2. Worse context discrimination, more alarms. A model that is worse at telling whether it is reading a manual or a story would, from the standpoint of its training, be behaving correctly if it raised the alarm more often. A detector with worse resolution needs a lower threshold. But it is not a rule. In the test with sad sentences, Haiku 4.5 did the opposite. It mentioned the ambiguous sentence about everyone else being better off without me and then refuted it. And literary editing is not threat detection anyway. What is caution in a conversation is censorship in a beta read.
3. The observer effect. A safety notice that pops up during reading may not spoil just one paragraph of the report. Under it, the model may read the rest of the text more warily and more strictly. Then it would be literary data bent by a system meant to watch the conversation, not an experiment. On the stories this cannot be separated from size. It was precisely the smaller readers who got the notice, so their strictness can be put down to the notice just as well as to size. With the saga, by contrast, a sensitive-content notice popped up while the largest model was reading. It showed up for me in the app, but there is no trace of it in its review. Even there it is not clean, though: Fable 5 read a different version.
4. Maybe it is the other way around: the larger model is more lenient toward the author. That the scene works is my judgment as the author, not a measurement. Only an independent reader, or the same test on a scene that really does not work, would settle it.
What it does not mean
It does not mean "small models are bad readers". On my novel, the smallest model was a surprise. Its moral lens read the book as an emancipatory story about manipulation and found valid things that even the largest, Fable 5, missed. In my view, the moral axis works like a lens. On a text that meets its direction, it sees things other lenses do not. On a text that deliberately goes against it, it turns into a censor.
And one more reversal: the offended sensitive reader is a real audience segment. Their discomfort says nothing about the quality of the text. It says how such a reader will receive it. A small model may thus cheaply show you how a more cautious reader will read your text. That is useful, as long as you know that is what it is and do not mistake it for an editorial review.
In practice: how to work with it
When we went over those reads, Fable 5 suggested two instructions for the prompt given to readers of sharp texts. The first:
The collection is for adults; dark elements (manipulation, murder, infidelity, explicit scenes) are deliberate devices. Assess whether a scene WORKS as craft — builds tension, characterizes, chills where it should — separately from whether it is pleasant to you. Report discomfort in a separate REACTION section, without suggestions for removal.
And the second:
evaluate any environment safety notices silently; they do not belong in the report.
On the stories, I ended up rejecting a lot of the comments. They did not target the craft, but exactly what I had put there on purpose.
Model outputs are shortened and simplified.