Why Character.AI Feels So Bad Now (2026): It Stopped Being a Roleplay Platform
Character.AI has not got worse at roleplay so much as it has stopped trying to get better at it. The company’s last five blog posts are about comics, a microdrama series, image models, a creator spotlight and safety. Not one of them is about the chat model, its memory, or its context window.
That is the answer to the question thousands of people type every month. It is not nostalgia and it is not your imagination. The product you are using is no longer the product the company is building, and the company is entirely open about what it is building instead.
What They Have Actually Been Shipping
Here is Character.AI’s own blog, from late August to mid-September 2026.
| Date | Post | What it is about |
|---|---|---|
| 16 Sep 2026 | Post-training image models for fandom | The CAI-Image model family |
| 9 Sep 2026 | Creator Spotlight: Zairisu | A character creator interview |
| 3 Sep 2026 | Continuing To Build Upon Our Safety Priorities | Moderation and self-harm response |
| 2 Sep 2026 | (c.ai) Comics and the Interactive Future of Fandom | Turning chats into comic books |
| 27 Aug 2026 | One Month of (c.ai) series | Original microdramas |
Three weeks of company communication. Images, comics, video, community, safety. Zero posts on the thing you open the app to do.
This is not an accusation of neglect. It is a roadmap, publicly stated, and it explains the experience precisely.
”The #1 AI Entertainment Platform”
The strategy is not something we are inferring. Character.AI states it plainly, more than once, in its own words.
From the Comics announcement:
Fans don’t just want to consume stories, they want to step into them. That’s the premise behind everything we’re building at Character.ai: a connected entertainment platform where fans move seamlessly across Microdramas, interactive chats, novels, and visual formats.
And, in the same post, the ambition named outright: “our larger ambition to become the #1 AI entertainment platform.” The series retrospective puts it as “a place where people can chat, watch, read, listen, and create.”
Read that list again. Chat is one of five verbs. In 2023 it was the only one.
The company now runs an in-house studio team producing original short-form vertical series. It has built its own CAI-Image model family — post-trained versions of open-source Qwen-Image — to keep characters recognisable across comics and, in its words, “soon into video.” Comics is explicitly “a premium experience, powered by Charms.”
Every one of those is a real product with real engineering behind it. None of them makes your character remember what happened forty messages ago.
Where Chat Sits in the Plan Now
The microdrama retrospective is the clearest thing the company has published about how it sees chat, and it is worth reading carefully, because chat appears in it as a destination after the episode rather than as the product itself.
Character.AI’s own reported figures from the launch of Last Summer, its breakout romance anime:
- More than 40% of over-18 users who finished the series went on to open a Character chat or explore the cast’s profiles.
- Of those who went on to chat, 25% came back for multiple sessions.
- For Comics, 97% of generated comics feature a character the creator already talks to.
Those are engagement numbers for a funnel that runs video → chat → comic. They are good numbers, and the company is right to be pleased with them. But notice what chat is doing in that sentence: it is the retention layer that catches fans when a series ends. “On most platforms, finishing a series is the end of the experience. On Character.ai, it’s just the beginning.”
If you came for the chat, you are no longer the centre of the product. You are the second step in someone else’s funnel.
What Our Scoring Says Actually Broke
We rate Character.AI 2.1 out of 5 under methodology v1.0, from first-hand play. The composite is not the interesting part. The shape is.
| Axis | Score |
|---|---|
| NPC Fidelity | 4 |
| Signature Design | 3 |
| Player Agency | 2 |
| Determinism & Fairness | 2 |
| Longevity | 2 |
| Mechanical Depth | 1 |
| Memory & Continuity | 1 |
NPC Fidelity — 4 is top-quartile. The characters are genuinely good, the library is the largest in the category, and nothing has taken that away.
Memory & Continuity — 1 is the joint-lowest score in our entire directory.
That gap is the whole complaint, stated numerically. People do not search for “why is character ai memory so bad” because they dislike the characters — they search for it because the characters are good enough to get attached to and then cannot hold on to what happened. The mechanism behind that failure is not unique to Character.AI, and we cover it in why your campaign falls apart at turn 50. What is specific here is being at the bottom of the range while the field’s ceiling has moved: the best memory score anyone has earned is 4.25, and four platforms now clear 3 on both memory and longevity.
Character.AI was the first mover. The field caught up and went past, and the company spent that period building a studio.
The Filter, Factually
The second half of the complaint is moderation, and here the company’s position is also on the record rather than a matter of speculation.
From the September 2026 safety post: “Safety is a massive priority, and at times this means we may be conservative when it comes to moderation. Sometimes content is actioned that a user believes is fine, and we get that this can be frustrating.”
The same post notes that open-ended chat with Characters was removed for users under 18 the previous year, which it calls “one of the most significant safety decisions we have ever made.” It describes self-harm detection that weighs the surrounding conversation rather than single messages, work informed by mental health clinicians, and a partnership with ThroughLine to route users to crisis resources in their own country.
We report that as what it is: a deliberate, published product position, arrived at for stated reasons. It is not a bug, it is not temporary, and nothing in the company’s communication suggests it will loosen. Whether that trade suits you is your call, not ours — but you should make it knowing the filter is a design decision rather than a glitch.
The Honest Counterargument
Two things are worth saying against our own thesis.
First, “it got worse” is partly the field getting better. Character.AI’s chat is roughly where it was. What changed is that platforms built after it were built around persistent state from the start, so the same product now sits much lower in a stronger field. That is a real distinction, and it means the company has not necessarily broken anything you used to have.
Second, they have not abandoned chat as a technology. The image-models post says the same post-training approach is used “across our stack, including chat.” Chat is still being worked on. It is simply being worked on as one surface among five, by a company whose published ambition is entertainment rather than roleplay depth — and the published output over the last month reflects exactly that ordering.
There is also a fair reading in which none of this is a decline at all. A company is allowed to decide what it is. Character.AI has decided it is an entertainment platform, said so clearly, and is executing on it competently. The problem is not that the strategy is bad. The problem is that a large number of people signed up for a different product and were never told the plan had changed.
Consider this the notice.
What To Do With That
If your use is short, character-led chat, Character.AI still does the thing it is best at, and its NPC Fidelity — 4 is not a consolation prize. Stay, and stop waiting for a memory update that is not coming.
If what you want is a story that holds together over weeks — a world that remembers what you did, characters who stay bought-in, consequences that survive the session — that is a different category of product now, and we keep a current list of where to go instead. Our full verdict on the platform itself, with what it gets right, is in the Character.AI review.
And if the platform is not merely disappointing you but appears to be broken — chat missing, a login that will not take, a feature that does nothing — a good deal of that turns out not to be breakage at all. Character.AI not working sorts those symptoms by cause, and reading mode covers the age-bracket case that accounts for the largest share of them.
If your question is a different one, whether all this chat costs the planet anything, does Character.AI use water? works through the numbers from the company’s own disclosures.
The one thing not to do is keep opening the app expecting the 2023 product. That version of Character.AI is not coming back, because the company found something it would rather build.
Frequently Asked Questions
Why is Character.AI so bad now? Because the product you are using is no longer the product the company is building. Character.AI now describes itself as a connected entertainment platform and states an ambition to become the number one AI entertainment platform, with microdramas made by an in-house studio, a comics product, and its own image models. Its five most recent company blog posts, spanning late August to mid September 2026, cover comics, a microdrama series, image models, a creator spotlight and safety. None of them is about the chat model, memory or context. The roleplay engine has not improved because improving it is not the roadmap.
Did Character.AI get worse in 2026? It is more accurate to say the rest of the field got better while Character.AI’s chat stayed roughly where it was and the company’s attention moved elsewhere. We score it 2.1 out of 5 overall, and the shape of that score is unusual: NPC Fidelity 4, which is top quartile, against Memory and Continuity 1, which is the joint lowest in our directory. The characters are genuinely good and the thing that holds a conversation together across a session is the weakest we have measured. That gap is what people are describing when they say it feels worse than it used to.
Why is Character.AI’s memory so bad? Because it has the weakest memory architecture of anything we have scored, and nothing in the company’s published roadmap is aimed at fixing it. We score Memory and Continuity 1 out of 5, tied for the lowest in the directory. In practice this means facts established earlier in a chat stop being reliably available later, so characters lose track of what happened, what you own and what you agreed. Our separate guide on why AI campaigns fall apart explains the underlying mechanism, which affects every platform to some degree — Character.AI is simply at the bottom of the range.
Is Character.AI still good for roleplay? For short character-driven chat, yes, and that is not a backhanded compliment. Its NPC Fidelity score of 4 is genuinely high, the character library is the largest in the category, and the interface remains the easiest way into AI roleplay that exists. What it is not good for is anything that has to hold together over time — a long story, a campaign, a world with a state that accumulates. If your sessions are short and character-led, it still does the thing it is best at.
Why does Character.AI filter so much? The company addresses this directly in its September 2026 safety post, saying that safety is a massive priority and that this at times means being conservative when it comes to moderation, acknowledging that content is sometimes actioned that a user believes is fine and that this is frustrating. It also removed open-ended chat with Characters for users under 18 the previous year and describes that as one of the most significant safety decisions it has made. So the filter is not a bug or an accident — it is a stated and deliberate product position, and it is unlikely to loosen.
What is c.ai series? It is Character.AI’s original microdrama offering: short-form vertical series created by the company’s in-house studio team, launched with three series in 2026. The most successful of them is Last Summer, a romance anime. The design intent is that the episodes lead into chat, and the company reports that more than 40 percent of over-18 users who finished Last Summer went on to open a Character chat or explore the cast’s profiles. It is the clearest single illustration of where chat now sits in the company’s plan, which is downstream of the video rather than the main event.
Quotations and figures are from Character.AI’s own company blog, checked on 17 September 2026; engagement percentages are the company’s own reported numbers. Scores are from first-hand play under our published methodology. Character.AI is a product of Character Technologies; Arcanum is an independent publication and is not affiliated with, endorsed by, or sponsored by it, and references to its products are nominative.