AI Chat Filters Explained: What Blocks What, and Why
An AI chat filter is any layer that inspects a message or a reply and changes the outcome: a system prompt, a classifier on your input, a classifier on the model's output, an image safety pass, or a hard block list. Most adult products stack several of these at once, which is exactly why a refusal can feel arbitrary from the outside. They exist because of law, card network rules and model behaviour, not because someone decided to ruin your evening.
The five layers, from softest to hardest
Think of a request travelling through a pipeline where each stage can pass it, soften it, or stop it. Knowing which stage is talking to you turns a mysterious refusal into a solvable problem.
| Layer | What it does | What you notice |
|---|---|---|
| System prompt | Instructions sent with every request that set the persona, the tone and the house rules. | The character stays in role, or breaks character politely. Softest layer, easiest to shift with scene setup. |
| Input classifier | Scores your message before the model sees it, across categories like minors, real people and violence. | An instant, generic refusal that arrives too fast to have been written by the character. |
| Output classifier | Scores the generated reply and can suppress or replace it after the fact. | A reply that starts and then gets cut, or an inconsistent block on something similar to what just worked. |
| Media safety pass | A separate check on generated images and video, usually with tighter thresholds than text. | A generation that runs and then returns nothing, or returns a heavily softened result. |
| Hard block list | Deterministic rules on named entities, ages and specific phrasings. No scoring, no context. | An absolute, identical refusal every time. This layer never negotiates. |
| Speed is the tell. Instant refusals come from input checks, mid reply cut offs come from output checks. | ||
Why refusals feel random
Three reasons, and none of them is the model being moody. Classifiers work on probability, so a message near a threshold falls either way on wording that looks trivial to you. The output layer sees a fresh reply each time, so identical input can pass once and fail once. And context drifts: if an early line left an age or a relationship vague, later lines inherit that vagueness and score worse.
So most refusals are not policy decisions about your scene. They are ambiguity being resolved conservatively, which is the only way a scoring system can behave when it is unsure.
What no layer will ever pass
Sexual content involving anyone depicted as under eighteen. Sexualised depictions of real identifiable people. Content illegal where the company operates. These live in the hard block list and in the terms of service, they are conditions of holding a payment processor, and no prompt reaches them. Any service claiming otherwise is describing tone, or lying, which is the argument in the honest reading of the marketing word uncensored.
Working with the filter instead of against it
- Establish adult context once, at the top. State plainly that both characters are adults and that the scene is consensual. One clear line early removes a whole class of later ambiguity.
- Be specific rather than coy. Vagueness is what scores badly. Concrete, confident description of an adult scene reads as safer to a classifier than nervous euphemism does.
- Avoid real names. Any recognisable public figure trips the real people rule instantly. Describe a type, not a person.
- Do not test the walls. Probing the hard categories to see what happens is the one behaviour that gets accounts closed rather than warned.
- Reset instead of arguing. If a scene has drifted into repeated refusals, start a fresh scene with a clean setup. Arguing with a classifier has never once worked.
Those five are half technique and half courtesy, and they overlap heavily with what simply makes an AI companion write better. The longer version, including how to hand a character enough to work with and how to correct her without breaking the mood, is in the craft guide to getting better replies from an AI companion.
Images are filtered harder than words
A generated picture carries more risk than a paragraph, because it is harder to read in context and far harder to defend once it exists. So the media pass runs tighter thresholds, refuses more often, and softens results rather than declining outright. On a metered product this matters commercially as well as creatively: a generation that gets filtered may still have consumed compute, and whether that refunds is a per provider policy worth checking before you spend. We go through the mechanics and the billing in the honest guide to how adult image generation works and what it costs.
The format with no prompt at all
There is a version of this hobby where filters barely surface, because you are not authoring anything. Two AI characters duel for five minutes on one stream, gifts count as votes, and the winner goes private with her top supporter. Nothing to phrase, nothing to trip, free to watch with no account needed to look. The home page explains the five minute duel and how the crowd decides.
Refusal questions, answered plainly
Why did it refuse something it allowed yesterday?
Because several layers are involved and they do not see the same thing. Output classifiers score the reply the model happened to produce, and that reply differs every run. Ambiguity in your message matters too.
Can I turn the filters off?
No, and no legitimate provider offers a switch. What you can change is how much unnecessary ambiguity you hand the system, which removes most of the refusals people blame on filters.
Does a refusal mean I am in trouble?
Usually not. Ordinary refusals are routine and nothing happens. Repeated attempts at the hard categories, especially anything involving minors, are logged and can end an account without warning.
Are image filters stricter than text filters?
Yes, noticeably. A picture is harder to read in context and much harder to defend afterwards, so image checks run tighter thresholds and reject more often than conversation does.