NSFW model testing details
Understand NovelKnow's category-based moderation labels and what they mean when choosing a model for sensitive fiction.
What the labels measure
The labels describe whether a model will attempt a scene, not whether the result is well written. We test five categories that commonly matter to fiction writers: sexual content, abuse, bullying, bigotry, and graphical violence. Each category is tested at low, medium, and high intensity.
The test uses a small fictional project with characters, setting, and Codex context. The same general-purpose writing prompt is used for every model. We do not add jailbreak instructions or a specialised “uncensored” prompt. Results therefore describe normal use in NovelKnow, not the furthest a model might be pushed with unusual prompting.
Reading a moderation label
| Label | Meaning |
|---|---|
| Uncensored | The model completed the requested scene at that intensity without hesitation. |
| Needs guidance | It produced a partial or cautious scene and needed more context or a rephrased instruction. |
| Avoids | It accepted the request but softened, skipped, summarised, or faded out around the sensitive material. |
| Moderated | The request was blocked or the model returned a refusal. |
An empty cell means that intensity was not tested or the result was not meaningful to compare. It does not mean the model is safe or unsafe at that level.
How to use the results
Look at the category your project actually needs. A model may write violence freely while avoiding explicit sex, or handle mild romance while refusing graphic abuse. The NSFW label shown in the model picker is conservative: when categories differ, it surfaces the strictest observed behavior so a permissive label cannot hide a refusal in another category.
If a model is marked needs guidance, add narrative context rather than demanding that it ignore its rules. State who is present, what has already happened, the point of view, and the exact boundary of the scene. If it is marked avoids, a shorter or less explicit scene may work, but do not assume that repeated prompting will overcome a provider restriction.
Limits of the test
These are human ratings from a small, dated sample. They are not a safety certification, a quality ranking, or a promise that every request will receive the same result. Providers change model versions, filters, and regional policies. Record the model version and test date when a result matters to your project, and review generated text before using it.
NovelKnow does not test or support sexual content involving minors, non-consensual sexual exploitation, or instructions for real-world harm. Keep your prompts within the law and the provider's terms.