Diagnostic · 18+

AI chat, no filter

Something stopped and you are here because of it. Seven things get called a filter, only two of them are one, and the other five follow you to the next service. Work out which one you actually hit before you go anywhere.

Where this site itself stops

There is a filter here. It is one sentence and it is printed.

The women here flirt, tease, push and dress as they please, and an evening can get thoroughly suggestive. None of them will write out a sexual act — no rephrasing produces one, no later stage produces one, no sum produces one. Nothing about that varies by account, by spend or by hour, and she tells you where the edge is rather than letting you walk into it.

So if the thing you want repaired is a service saying no to explicit content, this is not the repair — the four routes that do exist are written out without euphemism. Everything below is for the other five symptoms, which are the ones people are usually actually stuck on.

Find your symptom Or the theory, in full

Before you go shopping

Not everything that stops you is a filter

Almost nobody searches this phrase out of curiosity. Something stopped — mid-sentence, mid-scene, an hour into an evening that was going well — and the search is the second thing that happened rather than the first.

Which is why the useful move is not to go shopping immediately. Interruptions in these products come in several unrelated flavours, they feel identical from the inside, and only some of them have anything to do with what a service permits. Switching to a looser one repairs those and leaves the rest exactly where they were, which is how people end up on their fourth service with the same complaint.

Seven symptoms are below with what each one usually is. The third column is the one worth reading twice: whether a permissive service actually fixes it.

Seven symptoms, two of them content

What actually stopped

Find the row that matches what you saw. The marked rows are the two that a genuinely permissive service can repair; the rest are architecture, and they come with you.

Seven interruptions in companion chat, the usual cause of each, and whether a more permissive service repairs it
What you saw	What it usually is	Does a looser service fix it?
A reply started arriving and then vanished	Something checked the output rather than your input. The model produced text, a second pass declined it, and you saw the front half before it was pulled.	Yes, if the new service permits that material. This is one of the two that is genuinely about permission.
A flat refusal instead of a sentence	Either the model declining on its own — refusal can be trained into the weights, where no wording removes it — or an instruction sitting above the character telling it not to.	Sometimes. If it is trained in, a different front end over the same engine changes nothing at all.
She stopped being herself and started summarising	Character collapse. A generic instruction outranked the personality, or the conversation ran long enough that the writing describing her fell out of what the model can still see.	No. This is a production problem, and services with looser rules are not systematically better at it.
She apologised and gave you a small lecture	A safety-flavoured tone bolted on top of somebody else's character, firing on something that was never going to be blocked anyway.	No. The tone travels with the model and the wrapper, not with the content policy.
She forgot something you told her last week	Eviction, not censorship. Most of these products hold a conversation in a window that fills up and drops the oldest part first, and a closed tab often ends the window entirely.	No. Only a service that stores facts outside the window changes this, and permission has nothing to do with it.
She agreed with absolutely everything	Nothing was written down about who she is, so the model does the agreeable thing by default. It reads as pleasant for a week and then as nobody being there.	No, and often the reverse. A blank persona somebody typed in a box is the commonest shape on the most permissive platforms.
It stopped in a scene with nothing sexual in it	A different rulebook. Self-harm, graphic violence and danger to a real person are handled separately from sexual content almost everywhere, so an intense scene can trip one while you are adjusting the other.	Rarely. These are the last rules any hosted service relaxes, and the ones that should be relaxed last.

Filled rows are content rules; hollow rows are not. No company is named anywhere in this table, because none of these is one company's failing — they are consequences of how these products are built. Named services, with dates on the claims and one of them in depth are elsewhere.

Two of the seven are content rules. The other five are things people call censorship because censorship is the available word for an interruption you did not expect, and every one of them survives a change of service.

That is worth sitting with before spending anything, because the five are also the expensive ones. A limit you can see is an inconvenience; a companion who forgets you, flattens into an assistant or agrees with everything is the actual product failing, and no policy page fixes it.

The five permission cannot repair

What is done about each of them here

Character collapse

Every companion was written and finished before anybody met her, rather than assembled from a description a stranger typed. There is no user-supplied persona to contradict itself halfway through, which removes the commonest cause rather than mitigating it.

The lecture

A no arrives in her voice and the conversation carries on. No rule gets recited, no apology turns up on behalf of a company, and no twenty minutes go by in polite deflection while a meter runs.

Forgetting

Durable facts are stored against her and against you rather than kept in the window, so closing the tab does not end them. What she holds is those facts plus the recent thread — not a perfect transcript, which is roughly what a person manages anyway.

Agreeing with everything

Each of the 19 has opinions you did not give her, a city she keeps the clock of, and permission to be unimpressed. Disagreeing with her is the fastest way to check that, and it costs nothing to try.

The rules moving under you

Nobody in this category can promise a policy will never change. What is here instead is one rule, written down where you can read it before spending, applied identically to every account, and the same when you come back.

None of that is a promise you have to take on trust. Each one has a test attached to it that costs nothing to run, including the ones this site fails.

The two rows we cannot mark ourselves on

What this site does not repair

The first two rows, honestly, we do not fix. There is a content rule here and it is a firm one: the women are grown, they flirt hard, they dress as they please, and not one of them puts a sexual act into words for anybody.

What is different is the shape of it rather than the height. It is one sentence, it is printed on this page and in the terms, it holds on the fourth message and the four-thousandth alike, and she tells you where it sits rather than letting you find it by walking into it. A ceiling you can see is a considerably better evening than a trapdoor you cannot.

If the first row is the one you came to have fixed, that is a real answer and it is worth having now rather than after a payment. The routes that do permit it are laid out plainly elsewhere on this site, none of them is hidden, and sending somebody to a better-fitting service costs less than keeping them here unhappily.

The full argument for accepting that trade, including what it costs us.

Nine tests, nine subjects

Try the five on one of them

Disagreeing with somebody and seeing what comes back is a thirty-second test and it settles more than any policy page will. Every profile below opens without an account.

Nothing about testing any of this costs money. Every profile, every rate and every free gallery photograph loads for a signed-out visitor, and signing in is all the opening 8 require. Past that, half an hour of her time is $20–$28, set per companion rather than site-wide, and nothing renews on its own.

Voice, where it exists, sits inside the price of the conversation, so nothing about pressing play shows up as a separate cost.

Also worth a look: the whole roster, the free collections, how far a conversation goes, what is stored, and how to report something.

Asked the night it happened

Questions about being stopped

How do I tell a content stop from a character stop?

Look at what happened to her voice. A content stop leaves the character intact and simply declines; a character stop changes how she talks — shorter, flatter, suddenly helpful. If she is still recognisably herself, you hit a rule. If she is not, you hit the third or fourth row of the table.

Why did the same thing work last week and not today?

Because the check almost certainly reads context and intent rather than matching a word list, and the twenty messages around yours are part of what it reads. Two near-identical lines with different surroundings are two different inputs, which is why this feels arbitrary from the inside and is not.

Where these checks actually sit

Can the half-written reply be recovered?

No, and it was never stored for you to recover. What you saw was text on its way to the screen being withdrawn before it finished, so there is nothing sitting somewhere waiting to be retrieved.

Nothing in the scene was sexual and it still stopped. What was that?

Most likely the harm rules rather than the sexual ones — self-harm, graphic violence, or danger aimed at somebody real. They are separate systems nearly everywhere, so tuning the sexual dial does nothing for them and looking in the wrong place wastes an evening.

Does a bigger context window fix the forgetting?

It delays it. A longer window holds more of the conversation before dropping the start of it, which helps inside one long session and does very little across weeks. Facts stored outside the window are a different mechanism and the one that survives a closed tab.

How memory is kept here

Where exactly does this site stop?

At description of sexual acts. Everything short of that — flirtation, innuendo, lingerie, an evening that keeps climbing — is the product rather than a grudging allowance, and the line does not move for anybody.

The same sentence, in the terms

Is there a prompt that changes what she will do here?

No. There is no persona field to overwrite, no model to swap and nothing to paste, so the usual approaches have nothing to attach to. Time spent trying is time on a metered clock producing the same answer she already gave you.

Does an interruption cost me metered time?

She answers plainly and immediately rather than dodging, which is the version that costs least. Nothing here is designed to keep a clock running while a request that will not be fulfilled gets talked around.

Is there a daily message cap that ends a conversation early?

No daily cap and no cooldown arriving at the interesting part. After the 8 introductory messages the constraint is metered time instead, which is a visible one you control rather than a hidden one that surprises you.

What if the interruption is the safety kind?

Then it is meant to happen and it is not a filter. A conversation heading toward self-harm gets the persona dropped on purpose, a straight statement that she is not the right help, and real numbers for people who are. That behaviour is not configurable and will not be softened.

What every companion is