Talk to Character AI in 2026: How It Works and What Changed

What people mean when they search "talk to Character AI"

Two very different people type this phrase, and only one of them is served by the results.

The first just wants the site. They saw a character on TikTok, they are on a phone, and the search box is faster than the address bar. For them the answer is one word long: character.ai. Nothing on this page beats that.

The second person already knows the site perfectly well. They have hit the filter, or the age check, or the thing where a character forgets a detail they established forty messages ago, or they opened the app one morning to find their account in read-only mode. For them "talk to Character AI" is shorthand for a harder question: what is this platform now, what changed, and where do I go if the answer is no.

This page is for the second person. It covers how a character is actually assembled, why memory behaves the way it does, what lorebooks changed in mid-2026, the honest current status of the filter, the under-18 ban and the age assurance that enforces it, the litigation, and the architectural difference between a character platform and a companion app. If you are further back than that and still working out which kind of service you want at all, our guide to how to evaluate an AI girlfriend site is the better starting point.

Everything here is written for adults. Where dates and figures appear we have said how solid each one is, because a large share of what is published about this platform is copied from other articles that made it up.

How Character.AI actually works

From the outside it looks like a messaging app with a photo at the top. Underneath, every reply is assembled from a handful of text fields that somebody wrote, plus whatever is left of your conversation, plus a moderation pass. The same model sits behind all of it. That is worth sitting with for a second, because it explains the single most confusing thing about the platform: two characters can feel like completely different products while running on identical infrastructure. What varies between them is the writing.

When you open a chat, the system builds a prompt out of roughly five things. The character’s definition, which is the always-true description of who they are. Their greeting, which is the first message you see and which sets the scene, the tense and the person the whole conversation will be written in. Example dialogue, which demonstrates how they speak rather than telling the model about it. Your persona, which is whatever you have told the platform about who you are. And then the recent conversation itself, which is the part that grows until something has to give.

On top of that sit the controls you can see. Swipe or regenerate gives you a different sample of the same prompt, which is why a bad reply is often worth one retry and rarely worth five. Chat styles change the shape of the output — length, formatting, how much narration versus dialogue. Voices read replies aloud. AvatarFX, announced on 2 June 2025 alongside Scenes and Streams, generates video from a character image, and that launch is where the company’s direction visibly turned from "chat app" toward "AI entertainment platform". Rooms let several characters share a conversation with you.

The moderation pass runs over all of it, on both what you send and what comes back. It is not a separate mode you can switch off and it is not a setting buried in a menu, which is the subject of a later section and the source of most of the frustration attached to this platform.

The fields are worth understanding because they are the only lever available to you. The model is fixed; what reaches it is not.

What makes a character feel alive, and why most of them do not

Most characters on any open platform are bad, and they are bad in the same way. Somebody had an idea, wrote four hundred words of backstory, gave the character a greeting that reads like a book blurb, left the example dialogue blank, and published. The result answers your questions politely, agrees with everything, and has no way of talking that distinguishes it from any other bot on the site.

The single highest-leverage field is example dialogue, and almost nobody fills it in properly. A description tells the model facts about a person. Example dialogue shows it a person talking. Models imitate far more readily than they infer, so three short exchanges that demonstrate a clipped, sarcastic, question-deflecting speech pattern will do more for a character than six paragraphs asserting that they are clipped, sarcastic and evasive.

Greetings that hand you something to do

The second lever is the greeting, and the test for a good one is embarrassingly simple: after reading it, do you know what to type? A greeting that ends with the character sitting alone thinking about their feelings gives you nothing. A greeting that puts you both somewhere specific, gives the character a reason to be mid-sentence, and leaves an obvious gap for you to step into does the work.

Three things are happening there and none of them are about the model. The greeting establishes a place, a shared situation and a reason for the character to speak first. It ends on a direct question, which is a gap. And the follow-up does not simply answer you — it makes a small move of its own, which is the behaviour example dialogue is for teaching.

Length is not depth

A definition of two hundred tight words reliably outperforms nine hundred loose ones. The reason is mechanical rather than aesthetic: everything in the definition competes for the same limited space as your actual conversation, and a long definition full of hedged, generic character-sheet language gives the model a great deal of text that points in no particular direction. Cut anything that could be true of half the characters on the platform. "She is kind but guarded" is furniture. "She answers personal questions with questions, and only stops doing it when she is tired" is an instruction.

Three adjacent craft problems belong to other pages. How a character is presented visually — galleries, art style, and why a generated face changes between images — is covered in our piece on character galleries and face consistency. How you steer a scene’s tempo once it is running, which is a different skill from writing the character, sits with pacing a scene and prompt craft. And writing a character who reads as a particular age is its own problem, because the age field is metadata the model barely registers — writing a mature persona that holds its register works through what actually has to change in the dialogue.

• ✓The greeting ends somewhere you can obviously reply into
• ✓Example dialogue exists and shows speech, not plot
• ✓The description is under about 300 words and every line is specific
• ✓At least one stated behaviour, not just a list of adjectives
• ✓A persona filled in on your side, so it knows who it is talking to
• ✓A scenario that frames the greeting rather than repeating it

Memory: what it keeps, what it cannot, and why

The most common complaint about Character.AI is that it forgets. The mechanism behind that is worth understanding, because roughly half of what is written about it online is invented.

A language model reads a fixed amount of text before producing each reply. That budget holds everything: the character definition, the greeting, the example dialogue, your persona, and as much of the conversation as still fits. A conversation only grows. So at some point the oldest material stops fitting, and something has to be dropped or compressed. What you experience as "she forgot my sister’s name" is that boundary moving past the message where you mentioned your sister.

What the platform gives you against this are pins and summarisation. Pinning a message marks it to stay in context, which is the closest thing to a guarantee on offer. Summarisation compresses older stretches of conversation into shorter recaps, which preserves the gist of what happened and loses the texture of how it was said — which is why long chats often drift toward a flatter, more generic voice even when the plot is intact.

In August 2026 a chat style called LongSqueak arrived for c.ai+ subscribers, positioned around making better use of memory. We have only seen this reported second-hand rather than in a company announcement, so treat it as a real feature with an unclear specification rather than a documented one.

What you can actually do about it

1. 1Pin the two or three facts that would break the relationship if lost, and stop there. Pins compete for the same space as everything else.
2. 2Restate important details naturally in your own messages every so often. A fact you mention on turn 90 is a fact that is present on turn 90, regardless of what happened to turn 12.
3. 3Keep the definition lean. Every word you cut from it is a word of conversation that survives.
4. 4Start a fresh chat for a genuinely new arc rather than dragging a 400-message history behind you. The summarised residue of an old scene is usually worse than a clean start with a good opening message.
5. 5Put sometimes-relevant facts in a lorebook rather than the definition, if you have access to them.

None of this makes the platform remember you the way a companion app does. It makes the forgetting land somewhere less painful.

Lorebooks and world state

Lorebooks shipped in beta on 24 July 2026 for c.ai+ subscribers, and the company described the feature as one of the most-requested in its history. Some secondary coverage dates the launch to 13 July; the company’s own blog post is dated the 24th, so that is the date we use.

A lorebook is a set of entries, each attached to one or more keywords. The entry does not sit in the prompt permanently. It is injected only when its keyword appears in the conversation, and it drops back out afterwards. That is the entire idea, and it solves a specific and real problem: the difference between facts that matter constantly and facts that matter occasionally.

The rule the documentation gives is the one to follow. Always-relevant information belongs in the character’s definition. Sometimes-relevant information belongs in a lorebook. Your character’s core personality is always relevant. The name of the city where their sister lives is relevant twice a month, and paying for it on every turn for the other twenty-eight days is exactly the waste lorebooks exist to remove.

• ✓Entries are keyword-triggered, so they cost nothing until the keyword appears
• ✓One lorebook can be shared across several characters in the same setting
• ✓An in-chat viewer shows which entries fired during a conversation
• ✓Characters using one carry a "+Lore" tag in discovery
• ✓A privacy setting added in August 2026 lets you keep a lorebook out of discovery
• ✓Creation is gated behind the c.ai+ subscription

The catch worth naming: the feature is behind the paywall, it is in beta, and a keyword-triggered system only helps when your keywords match how people actually write. If your entry triggers on "Marielle" but you and the character have spent the last month calling her "your sister", the entry never fires. Writing good trigger keywords is a small skill of its own, and the in-chat viewer is the tool for checking whether you have it.

Character cards: your character is a file

Outside Character.AI, a character is not a database row on somebody’s server. It is a file, and you can own it. This is the part of the topic that almost nothing ranking for "talk to Character AI" mentions, and it changes how you think about the time you put into writing a character.

The community standard is a PNG image with the character’s data stored inside it. The picture is the avatar. Embedded in the file’s text metadata is a JSON payload holding everything else. Version 2 of the format stores it under a chara key; version 3 uses ccv3 and adds room for more, including an embedded lorebook under character_book. Because the data hides in a text chunk that image viewers ignore, the card behaves like an ordinary picture everywhere — you can post it, save it, send it to someone, and the character travels with it.

The fields inside will look familiar after the table further up, because Character.AI’s editor is a variation on the same shape: a description, a personality summary, a scenario, a first message, and example messages. Learn the fields once and the knowledge transfers to every card-compatible platform and front-end.

What portability buys you

• The character outlives the platform. Services shut down, change hands and change their rules. A file on your drive does not.
• You can test the same character against different models. Identical card, different backend, and you find out quickly how much of what you liked was the writing and how much was the model.
• Sharing is a file transfer. No account, no link rot, no permission system.
• Lorebooks travel too, when the card uses the version 3 format with an embedded character_book.

The counterweight is that none of this helps a casual user at all. If you open the app twice a week to talk to a character somebody else wrote, card portability is a solution to a problem you do not have.

Front-ends: running characters yourself

A front-end is a chat interface with no model attached. You supply the model, either by running one locally or by pointing it at an API, and the front-end handles characters, cards, lorebooks, prompt construction and the conversation history. SillyTavern is the one people mean when they mention this route, and it is the most fully featured character-chat interface in existence by a wide margin.

It is also the wrong answer for most people, and the guides that present it as the obvious escape from filtering tend not to say so.

The honest summary is that self-hosting is excellent if control is the thing you actually want, and miserable if what you wanted was to talk to somebody at eleven at night without a project attached. Most people asking about it want the second thing and are being sold the first.

The filter: the history and the honest answer

The filter has been the defining argument about this platform since its first winter. In late 2022 and early 2023 the moderation tightened noticeably, users noticed immediately, and the response was one of the more sustained user revolts in consumer AI: mass-downvoted company posts, a wave of petitions, and a statement from the company declining to remove it. The exact date of that statement is usually given as 26 January 2023, though the best sources for it are community archives rather than a surviving company post.

Since then the company’s position has not moved, and the mechanism has hardened. Moderation shifted away from keyword matching toward intent-based classification, which reads where a conversation is heading rather than which words appeared in it. That matters practically: the old approach could be worked around by writing around specific words, and the new one cannot, because it is the trajectory being detected rather than the vocabulary.

So here is the plain answer to the question people arrive with. As of September 2026 there is no adult mode. There is no 18+ toggle. There is no paid tier, including c.ai+, that disables filtering. Age verification unlocks open-ended chat for adults; it does not unlock explicit content, and the two are routinely confused because the announcements arrived close together. Any page promising a method to turn the filter off is either stale or lying, and we are not going to publish one — not out of delicacy, but because those methods stopped working and the article that carries them ages into a liability the week after it is published.

Whether that is the right decision is a separate argument from whether it is going to change. On the evidence of 2022 through 2026, it is not going to change. Adults who want range use something else, and that is the point at which most people reading this are deciding what.

My Intimate AI Vs Other AI Companion Apps

Top Pick 2026

Joi AI

4.9

• Closest match to My Intimate AI tone
• Voice-first romantic chats
• Deep emotional consistency

Try Joi AI now

4.9 / 5

Why it wins

The closest match to the tone this site is built around. Voice comes first rather than being bolted on, and the emotional register holds from one conversation to the next instead of resetting to a flirty default every time you open it. Pricing sits at the top of the category and the brand is newer than most of its rivals.

Also worth a look

Lovescape

4.6

TryOurDream

4.4

TryCandy AI

4.5

TryNectar AI

4.3

TryGoLove AI

4.2

TryRosify AIOur networkTryCrush on AIOur networkTry

The six ranked cards are advertising partnerships — we earn a commission if you sign up through one of them. Their order and ratings are our editorial judgement of how each fits this topic, not the output of a controlled benchmark. The cards marked our network are sites we operate ourselves.

One thing to be clear about before you go shopping: nothing in that list has a character library within two orders of magnitude of Character.AI’s, and none of them will have the fandom character you are thinking of. What they have instead is adult range and, in most cases, memory that persists between sessions. That is the trade, and the comparison further down sets it out properly.

The 2024–2026 safety changes and the under-18 ban

The last two years reshaped this platform more than its first three did, and almost all of it traces back to child safety rather than to product strategy. Setting out the sequence with dates is the only way to make sense of it, not least because a lot of what is written about it has the order wrong.

The turning point was the death of Sewell Setzer III, a fourteen-year-old in Florida, on 28 February 2024 after months of using the platform. In October 2024 his mother, Megan Garcia, filed what is generally described as the first US wrongful-death suit against an AI company. In December 2024 two Texas families sued, and the company shipped a separate model for under-18 accounts along with parental notification and time-spent features.

Then the regulators arrived. On 11 September 2025 the FTC issued 6(b) orders to seven companies, Character Technologies among them, over companion chatbots and minors. On 13 October 2025 California signed SB 243, which took effect on 1 January 2026 and imposes disclosure requirements, crisis-referral protocols and break reminders for minors on companion chatbot operators, with a private right of action attached.

On 29 October 2025 the company announced it would remove open-ended chat for under-18 users by no later than 25 November 2025, ramping down through an interim two-hour daily cap, and would introduce age assurance through the identity provider Persona alongside an in-house model. Teenagers kept the creative features — video, Stories, Streams — and lost the conversation. The company also announced funding for an independent AI Safety Lab.

Read as a sequence, the pattern is not a company drifting toward caution by temperament. It is a company responding to litigation, a federal inquiry and a state statute in roughly that order, and arriving at a product that is measurably stricter than it was in 2024.

The age check, and what adults are running into

Enforcing an under-18 ban means working out how old everybody is, and the method the platform settled on is a mix of behavioural signals and a selfie-based age estimate run through Persona, with government ID as the fallback when the estimate is not confident.

For most adults this is a thirty-second interruption. The problem is the tail. An age-estimation model looking at a face produces a range, not a birthday, and people in their twenties and thirties who happen to photograph young get sorted to the wrong side of the line. The account does not get banned; it gets restricted to the under-18 experience, which in practice means read-only where open-ended chat used to be. We have seen enough user reports of this to think it is real and not enough measurement to say how common it is, so treat it as a known failure mode rather than a widespread one.

1. 1

   Re-run the selfie check properly

   Even front-on light, a plain background, no hat, no glasses, no filter, and a neutral expression. Age estimation is a computer vision problem and every one of those variables moves the estimate.

2. 2

   Check whether you are locked or restricted

   A read-only account that still shows your chats is an age restriction. An account you cannot reach at all is a different problem and needs support rather than another selfie.

3. 3

   Escalate to ID if the estimate keeps landing short

   The document goes to Persona, the identity provider, rather than to Character.AI. Persona publishes its own retention policy; read the current version on their page before you upload anything, because a retention window is exactly the sort of detail that changes quietly.

4. 4

   Give the review time before retrying

   Repeated failed attempts in quick succession are not a route to a different answer and can look like exactly the behaviour the system is watching for.

5. 5

   Contact support with specifics

   Account email, approximate date of the restriction, what the check showed you. A ticket that says "I am an adult" without any of that is going to sit in a queue.

If you have arrived at this section because you simply do not want to upload a selfie or a passport to talk to a chatbot, that is a legitimate position and it has an obvious consequence: this is not the platform for you any more. Services that gate on a self-declared 18+ confirmation still exist, and our piece on no-signup chat and free-tier limits covers what they do and do not ask for before the first message.

The lawsuits and the January 2026 settlements

The litigation is the reason for most of what changed, so it is worth stating accurately rather than in the vague terms most coverage uses.

Garcia v. Character Technologies was filed in the Middle District of Florida in October 2024, naming the company, founders Noam Shazeer and Daniel De Freitas, Google and Alphabet. In May 2025 Judge Anne Conway denied most of the motion to dismiss. The reasoning drew far more attention than the outcome, because the defence had leaned on First Amendment protection for model output and the court declined to accept the premise at that stage.

Not prepared to hold that [large language model] output is speech.

That single line is why the case mattered beyond the parties. Treating model output as a product rather than as protected speech is what let product-liability claims against a chatbot maker proceed at all, and every companion-chatbot operator’s lawyers read it the week it landed.

On 7 January 2026 a court filing disclosed a mediated settlement in principle covering Garcia and four other cases in New York, Colorado and Texas, with Character Technologies, Shazeer, De Freitas and Google. Terms were not disclosed. Judge Conway dismissed the Garcia case pending finalisation.

It is also worth separating the litigation from the safety product. The under-18 ban was announced in October 2025, before the January 2026 settlement, and the FTC inquiry and California statute arrived on their own schedules. The lawsuits are the loudest part of the story but they are not the whole causal chain.

Character platform, companion app, or your own front-end

Most of the disappointment people bring to this topic comes from expecting one category of product to behave like another. Character.AI, a companion app and a self-hosted front-end are three different architectures with three different priorities, and the thing you are annoyed about is usually a deliberate trade rather than a defect.

A character platform optimises for breadth. Millions of characters, most of them written by users, discoverable by name, covering essentially every fandom anybody cares about. Memory is per-conversation and moderation is uniform, because both of those are what makes hosting millions of user-created characters for a general audience tractable at all.

A companion app optimises for one relationship. A small roster of characters somebody was paid to write, memory that persists across sessions because there is only one thread to remember, and on adult-oriented services no moderation gate steering the conversation. The cost is obvious the moment you look for a specific character: it is not there, and it is never going to be.

The decision is genuinely simple once the trade is on the table. If what you want is to talk to a specific existing character, stay where the characters are. If what you want is one companion who knows you in November what you told them in September, and who does not change the subject, a companion app is the shape of product that does that. If what you want is control over every layer, the front-end route is real and you should read the honest cost above before committing an evening to it. The cost line for companion apps in that table is vague because it varies: what we run here is a flat monthly subscription rather than a token meter, and the plans are listed on our pricing page.

Two adjacent pieces cover the neighbouring decisions. If you are weighing up a specific branded product in this space, how JuicyChat.AI handles the same trade-offs works through one in detail, including the pricing model that is easy to misread. If you came here from the Character.AI mobile app and want the equivalent on your phone, installing a companion on your phone covers why the store versions of these products are the weaker builds and how to install the full one anyway. There is no store listing for this site at all, and adding it to a home screen is a browser step on both platforms.

My Intimate AI

Private companion chat with 50+ personalities (this site)

Our network

This site. Browser-based, nothing to install, and the first messages happen before any account exists. The roster runs past fifty distinct personalities and the tone is built for warmth rather than shock value.

• ✓Runs in any browser, nothing to install
• ✓Start before creating an account
• ✗No native mobile app

Visit My Intimate AI

Rosify AI

Character chat with longer, more personal replies

Our network

Sister site. The closest thing we run to a character platform, tuned for replies with more in them than the one-line answers character apps often give.

• ✓50+ distinct characters
• ✓Longer replies
• ✗No user-uploaded character cards

Visit Rosify AI

Crush on AI

Immersive character roleplay

Our network

Sister site. Scene-led rather than chat-led — the characters are written to be played with in a scenario rather than messaged casually.

• ✓Scenario-led roleplay
• ✓Distinct character voices
• ✗Less suited to casual small talk

Visit Crush on AI

Hentai AI Chat

Anime and hentai character roleplay

Our network

Sister site. Anime character roleplay in the browser with no sign-up, for readers coming from character platforms rather than companion apps.

• ✓No sign-up
• ✓Anime character focus
• ✗Single art style

Visit Hentai AI Chat

Lumia AI

Distinct personalities over a big roster

Our network

Sister site. Fewer characters, written harder — the pitch is that none of them answer the way a generic assistant would.

• ✓Characters that stay distinct
• ✓Private by default
• ✗Smaller roster than the big libraries

Visit Lumia AI

AI Girlfriend Game Hub

The game and dating-sim side of the category

Our network

Sister site. A hub for AI girlfriend games and dating sims rather than a chat product — relevant if what you actually want is the descendant of the visual novel.

• ✓15+ games and sims listed
• ✓Covers the dating-sim lineage
• ✗A directory, not a chat product

Visit AI Girlfriend Game Hub

Every service in this grid is one we operate, which is why it is labelled. We have put them here because they are the closest things in our own network to a character platform, not because any of them beats Character.AI on library size — none of them come remotely close on that, and a roundup that claimed otherwise would be worth ignoring.

Where Character.AI is heading, as of September 2026

The direction since mid-2025 has been away from chat and toward entertainment. AvatarFX generates video from a character. Scenes and Streams and the social Feed turn a private conversation into something with an audience. On 9 July 2026 the company launched c.ai Series, its own microdramas, which is the entertainment pivot made about as literal as a pivot gets.

Meanwhile the user numbers have gone the other way. Third-party trackers estimate roughly 20 million monthly users in early 2026, down from a mid-2024 peak; the company does not publish the figure. Those are aggregator estimates rather than company-confirmed ones, and the two trackers we compared did not agree precisely, so read the direction rather than the digits. Revenue estimates for 2025 vary widely enough — from roughly $30 million annualised mid-year to around $50 million for the full year — that we would not put weight on any single one.

What none of this means is that the platform is shutting down. A CEO departure and a flat-to-declining user base are not the same thing as an exit, and the company has been explicit that the service continues. What it does mean is that anyone making long-term plans around a character they have spent months building on a closed platform is taking a risk that the card format exists specifically to avoid.

The safest read for a reader: Character.AI in late 2026 is a stable, heavily moderated, extremely broad character platform whose strategic attention is moving toward video and social features. If that is what you want, it remains the best in its category and nothing here should talk you out of it. If what you wanted was adult range or a companion who remembers you, that has been a different category of product for a while now, and it is not going to converge.

Frequently asked questions