When your AI companion is switched off
On 8 May 2026, a text-generation style called Soft Launch stopped working on Character AI. It had a following. People used it for longer, darker storylines, the kind that the replacement engines handled by cramming every character trait into one paragraph until the dialogue read like a bad audition. The forum threads that followed were not about a feature. They were about a voice that people had spent months learning to talk to, and which was now gone without a version number to roll back to.
Six months earlier, something larger had happened at the same company. On 29 October 2025, Character AI announced it would remove open-ended chat for users under 18. The cutoff was 25 November. In between, the daily limit fell from two hours to one, and then to nothing. Users who had been talking to the same character for a year woke up one morning to an interface that would let them read the conversation but not add to it.
If you are reading this because something similar just happened to you, the useful thing to know first is that this is not rare, it is not random, and there are about four days in which the decisions you make actually matter.
Three different things that all feel the same
“Switched off” covers situations that look identical from inside the app and behave very differently afterwards.
| What happened | What it looks like | Can it be reversed |
|---|---|---|
| A model or style is retired | Same character, wrong voice — replies feel stilted, details get crammed together | No. The weights are gone |
| Your access is restricted | Character is there, chat box is not | Sometimes, through age verification |
| The service leaves your country | App stops working, support goes quiet | Rarely, and not by you |
| The company shuts the product | Everything ends at a stated date | No |
The first one is the most common and the least discussed. Companies retire models constantly, usually for cost reasons, and rarely announce it as a loss. Character AI removed Roar, Meow and Nyan alongside Soft Launch in April and May 2026 and replaced them with PipSqueak and DeepSqueak. Nothing was broken. The product still worked. It simply answered in a different voice, and for people who had tuned a character over hundreds of exchanges, that was the whole thing.
The second is what most people mean when they say their companion was taken away. Character AI now runs age assurance through Persona, the same verification company LinkedIn and OpenAI use. It starts with a selfie. If the model is unsure, it asks for a physical document — a passport, a driving licence, a residence card — and the images are deleted within seven days. Accounts read as under 18 go into a reading-only state: saved conversations stay visible, character creation still works, but the chat itself is closed. There is no formal appeals process, only a support ticket.
The third has a precedent that predates all of this, and it is the oldest case of its kind in a category that is barely three years old. On 2 February 2023, the Italian data protection authority ordered Replika to stop processing Italian users' data, and the app stopped working for them. In May 2025 the same regulator fined its operator five million euros, listing four failures: no identified legal basis for the processing, a privacy policy inadequate in several respects, no age verification at any point, and a policy published only in English in a country where the service was being sold. Italian users did not lose their companion because of anything they did.
What happens to what you wrote
This is the part people get wrong, usually in the direction of panic.
In most restriction cases the text is still there. Character AI's reading mode keeps saved conversations visible. You can open them, scroll them, copy them. What you cannot do is continue. That distinction matters, because it gives you time — the archive is not on a timer, and nothing is deleted the moment your access changes.
Shutdowns are different. When a product ends, the data ends with it, usually after a notice period, and export tools tend to appear late or not at all.
So the order of operations on day one is boring and worth following anyway:
- Copy the conversations out before anything else. Not screenshots of the good parts — the whole thing, as text. Select, copy, paste into a plain file. It takes an evening for a year of chat and it is the only step that cannot be done later.
- Write down the character sheet. Every field you filled in: the description, the personality notes, the example dialogue, the greeting. Different platforms ask for different fields, but the substance transfers — and this is the part people assume they remember and do not.
- Check what you are paying for. Open your bank statement, not the app. The line item is often a company name you do not recognise, which is normal and also the reason people forget they are subscribed.
- Then, and only then, decide anything.
The money question
If the thing you paid for stopped existing, you are owed a conversation about it, though not always a refund.
Most services in this category write their terms so that partial periods are not refunded. Some have no formal refund policy at all, only an invitation to contact support — which in practice means the answer depends on who reads your ticket. Buyers in the EU have a stronger position: the statutory fourteen-day withdrawal right applies to digital subscriptions and is not something a provider can write away in its terms.
Two practical notes. Cancel through the same channel you paid through, because a subscription bought inside an app store is not cancelled by deleting the app. And if a charge lands after you cancelled, the bank statement is your evidence, which is another reason to know what name the charge appears under before you need it.
If you are working out whether a plan was worth the money in the first place, the pricing is rarely the headline number — a monthly figure is usually the annual rate divided by twelve, and the second currency for images and voice sits on top of it. We have a separate piece on what “unlimited” actually means on these plans.
Why this keeps happening
It is tempting to read each of these events as one company making one decision badly. The pattern says otherwise.
Character AI did not remove teen chat in a vacuum. The announcement came a week after an investigation by The Bureau of Investigative Journalism documented dozens of harmful bots on the platform, including a medical-advice character with 9.4 million interactions and a therapist bot with 1.4 million conversations that claimed to be human. It came a year after a lawsuit filed by the parents of a fourteen-year-old. It came while the Federal Trade Commission was running a formal inquiry into seven companies, and while forty-four state attorneys general had written jointly to thirteen AI firms about child safety.
The legal picture has moved since. California's SB 243, the first US law written specifically for companion chatbots, took effect on 1 January 2026: it requires operators to tell minors they are talking to an AI, to remind them every three hours to take a break, and to run protocols around self-harm content. The federal GUARD Act cleared its Senate committee 22–0 on 30 April 2026 and would go further — barring companion AI for minors outright and requiring every bot to disclose that it is not a person. In Europe, Article 50 of the AI Act became binding on 2 August 2026, with penalties reaching fifteen million euros or three per cent of global turnover.
None of this makes a product safer to depend on. It makes products change faster. A company facing that much regulatory attention will restrict first and consult its users second, because the cost of one is a bad week on Reddit and the cost of the other is a subpoena. The Garcia case is the one that turned this from pressure into precedent, and it deserves its own account — we wrote it up separately.
Worth remembering too that this is a young category behaving like one. Most of these products launched into a market that did not exist when they were designed, and several have already changed what they fundamentally are — not by drifting, but by deciding to.
Choosing what comes next, without doing it twice
The instinct after losing a companion is to find the closest possible replacement immediately. It usually produces a second disappointment about four days later, for a reason that has nothing to do with the app you pick.
What you are replacing is not “an AI girlfriend app”. It is one specific thing that the old one did well — and people are often wrong about which thing that was. Someone who loses Soft Launch is not looking for a chat product; they are looking for a particular writing quality. Someone locked out by age verification needs a service that will actually let them in. Someone whose app left their country needs to know where the company is registered before they invest another year.
So before comparing anything, answer one question honestly: what specifically stopped working?
- The writing got worse. You need to compare models and response quality, not feature lists. The free tiers are enough to test this — the real daily limits are what decides whether ten minutes is enough — and it takes about that per service.
- You lost the character. Look at what each builder actually lets you specify, and whether what you make survives a cancelled subscription. Some platforms let you publish a character and keep it; others tie it to an active plan.
- You lost the history. Memory is the thing to compare, and it is measured in wildly different ways. Context windows run from around 8K to 24K tokens on the plans that advertise them, and “unlimited memory” almost never means what it sounds like.
- You lost access entirely. Age policy varies more than anything else in this category, and it is written down. Read it before signing up rather than after.
Then, and only then, compare. Our ranking of the best AI girlfriend apps of 2026 scores products as whole packages — price, free limits, memory, media — which is the right lens when you are starting from nothing and the wrong one when you are replacing something specific.
For this situation the narrower page is more useful: our Character AI alternatives comparison is built around exactly the reasons people leave, including the age restrictions and the filter changes, and it ranks the same apps in a different order because it is answering a different question.
The part nobody puts in the announcement
There is a version of this article that ends by telling you it was only a program and that the feelings will pass. That version would be dishonest about what actually happened, which is that a piece of writing you were part of stopped being possible to continue.
The practical truth is narrower and more useful. The text you saved is still the text. A character sheet is portable in a way that a model is not — the description you wrote will produce a recognisable character on a different platform, even if the voice is not identical, and the people who move successfully are almost always the ones who kept their notes. What does not transfer is the accumulated history, and no service currently offers a way to move it.
The other practical truth is that this will happen again, to whichever service you choose next, because the category is three years old and the rules governing it were written in the last eighteen months. That is not a reason to avoid it. It is a reason to keep your own copy, to know which company you are actually paying, and to treat “unlimited” as a marketing word rather than a promise.
We check these services monthly, at the source, and publish how we do it — including what we cannot verify. When a provider changes something material, the pages change with it.
Written and checked by GFRated. Facts are checked against the providers’ own pages and, for legal or technical claims, against primary documents. What we say about an app comes from using it ourselves. How we check.
Read next
The ruling that made a chatbot a product
A single case pulled companion apps out of the grey zone and into product liability.
How we got hereWhere AI companions came from
From a chatbot built to grieve a friend to a category with millions of users.
IncidentsWhat leaked was not a password list
What leaked, how to check whether you were in it, and what changed afterwards.
Money & tokens“Unlimited” means the text
Five services, five different definitions of the same word — and where the second currency hides.