Rewritten August 2026. This page was a 2023 list of named platforms. Several no longer exist and others have changed substantially, which is normal for this category. This version explains how character AI platforms work, what to check before using one, and what the safety and privacy considerations actually are.
Character AI platforms let you create a persona — a name, a personality, a way of speaking — and then converse with it. People use them for interactive fiction, language practice, brainstorming, roleplay and writing.
“SFW” in this context means the platform enforces content limits rather than allowing anything. That distinction matters more than any feature comparison, and it is the first thing to establish about a service.
How these platforms work
Three parts, and knowing them explains most of the behaviour people find puzzling:
- The model — a general-purpose language model doing the actual work.
- The character definition — a description, example dialogue and instructions, supplied as hidden context on every message.
- The memory — a limited window of recent conversation. This is why characters forget things from earlier: it fell out of the window, not because anything went wrong.
Nothing persists between messages. The character exists as text passed to a model each time you write.
What to check before choosing one
1. The content policy, and whether it is stable
Platforms differ enormously — some enforce strict limits, some almost none. Two things to know:
- Policies change, sometimes abruptly. This category has repeatedly seen platforms tighten filters overnight, with users finding characters they had used for months behaving completely differently. Nothing you build is guaranteed to keep working.
- Filters are imperfect in both directions. They block ordinary conversations and miss things they should catch. Frustration with a filter is not evidence of a broken service.
2. What happens to your conversations
Assume everything you type is stored, may be reviewed by humans for moderation, and may be used to improve the service. Roleplay conversations are often more personal than people realise in retrospect. Read the privacy policy for: training use, third-party sharing, and whether deletion actually works. If it is vague, treat what you write as permanent.
3. The cost model
- Free tiers typically limit message volume, response speed or memory length.
- Subscriptions commonly buy priority access, longer memory and faster responses.
- Subscribe through the app store where possible, so cancelling is straightforward.
4. Whether it is appropriate for the user
Many of these platforms are used heavily by teenagers, and this is where the category has attracted the most scrutiny. Two things worth stating plainly:
- Age ratings are not a reliable guide. User-created characters mean the content depends on other users, not on the platform’s rating.
- These are not mental health tools. They have no duty of care and cannot recognise a crisis. If someone is struggling, a GP, a therapist or a crisis line is the answer — those are free in most countries and they are staffed by people.
Writing a character that works
Most disappointing results come from a thin definition rather than a weak model.
- Be specific. “Sarcastic but fundamentally kind; deflects compliments; talks about food constantly” produces far more than “friendly and funny”.
- Give example dialogue. This does more than description — the model imitates a voice much better than it follows an adjective.
- Describe how they speak, not only what they are like. Sentence length, formality, verbal habits.
- Include what they will not do. Constraints shape a character as much as traits.
- Keep it concise. The definition consumes the same context budget as the conversation, so an enormous description leaves less room for memory.
When a character drifts
Long conversations lose the thread as early messages fall out of the context window. Restate the important details, or start a new conversation with a summary. This is a limitation of how the technology works, not a fault in the character.
What these platforms are genuinely good for
- Interactive fiction and collaborative storytelling.
- Language practice, where a patient partner who never tires is genuinely valuable.
- Rehearsing a difficult conversation before having it.
- Brainstorming with a character who holds a particular perspective.
- Writing help — testing dialogue for a character you are developing.
And what they are not good for: factual information, which they will invent confidently; and anything confidential, which goes to their servers.
Running one privately
If the privacy questions are the sticking point, local models are the answer. Tools such as Ollama, LM Studio or SillyTavern with a local backend run entirely on your own machine — nothing is transmitted, no policy applies, and nothing changes overnight.
It needs a reasonable computer, and quality varies with the model you can run. For people who care about privacy or about a platform changing under them, it is the only real answer.
Deleting your account and data
- Cancel any subscription first, through the app store.
- Export or screenshot anything you want to keep — most platforms have no export.
- Delete the account in the platform’s settings, and keep the confirmation.
- Email a deletion request if there is no in-app option. In the EU and UK you have a legal right to erasure.
- Revoke linked sign-ins in your Google, Apple or Discord account settings.
Frequently asked questions
What does SFW mean here?
That the platform enforces content limits rather than permitting anything. How strict varies, and policies change.
Are these platforms free?
Most have a free tier with limits on messages, speed or memory. Subscriptions buy priority and longer memory.
Why does my character forget things?
Limited context. Only a window of recent conversation is available on each message; older content falls out. Restate what matters, or start fresh with a summary.
Are my conversations private?
Assume not. They are stored, may be reviewed for moderation, and may be used for training depending on the policy. Local models are the only genuinely private option.
Can I use these for a school project?
Check the rules where you are. They vary from prohibited to permitted with disclosure. Also remember these tools invent facts, so nothing from them should be cited as a source.
Are they suitable for children?
User-created characters mean the content depends on other users rather than on the platform’s rating. For a younger user, a general-purpose assistant with parental controls is a far more appropriate choice.
