Across app-store reviews, itch.io comment threads, and first-person essays published between 2021 and 2026, six complaints about AI roleplay show up independently, on more than one product: the AI forgets what happened earlier in the session, it repeats itself, it breaks character to narrate as an assistant, content filters interrupt scenes that aren't explicit, cost and rate limits force players to pay again mid-story, and nothing a player does carries real, lasting consequences. None of these is unique to one app — the point of gathering them here is that they recur across different products, platforms, and years, which is what separates a pattern from a single bad review.
This isn't a takedown of any one platform. Most of the evidence below happens to name Character.AI or AI Dungeon, simply because they're the two largest, most-reviewed products in the category — but each complaint is presented as a pattern the category shares, not a verdict on either product specifically. Where a company has visibly acted on a complaint, that's noted too, because "this was a real problem and here's what changed" is more useful than pretending nothing has moved since 2021.
The six complaints, at a glance
| Complaint | What it looks like in practice | How well evidenced |
|---|---|---|
| Memory loss across a session | The AI forgets earlier events, relationships, or who said what | Strong — App Store reviews, a first-person essay, and the platform's own dedicated blog posts |
| Repetitive / non-reactive output | Same phrasing or scene description regardless of player input | Strongest in this set — 4 products, 4 platforms, 2021–2025 |
| Breaks character / narrates as an assistant | The AI rewrites a scene, swaps a persona, or steps outside the fiction | Moderate — one detailed first-person account plus real (if small) search demand |
| Content filters interrupt non-explicit scenes | Ordinary dramatic or romantic writing gets blocked or flagged | Strong, but split across two eras — 2021 and 2025-2026 are different incidents |
| Cost / rate limits / paying again | Credits run out mid-story; players who already paid are asked to pay again | Strong — the platform's own help center plus independent player reviews |
| No real stakes — nothing you do can be lost | Failure and death don't produce lasting consequences | Re-sourced — the strongest evidence is structural, not a review quote |
1. Memory loss across a session
The most basic ask of a roleplay session — that the AI remembers what happened ten messages ago — is also one of the most consistently reported failure points.
On the AI Dungeon App Store page, reviewer UberBolton wrote of running out of context space: "There are still some hiccups, especially if you're a free user." Reviewer nickname153717372 described the same limit on 2025-03-31: "I can't turn it down a notch, I can go in there and like delete a few things to make space, but then it fills itself back up." Both reviews describe the same underlying limit — memory is a finite, manageable resource the player has to actively work around, not something that just works in the background.
A first-person essay on Medium, published 2025-10-28 by writer SocialTipster, described the same problem in narrative terms: "the story that resets itself every few messages, as if the characters have short-term amnesia," adding that "after long conversations, it begins to lose track of who is speaking and why."
This complaint is well-documented enough that Character.AI has written two separate blog posts specifically about memory — blog.character.ai/memory and blog.character.ai/helping-characters-remember-what-matters-most — which is itself a signal: a company doesn't publish two dedicated posts about a feature unless the complaint volume behind it justifies the attention. As of 2026, this is a live, current complaint rather than a stale, already-fixed one — both the essay above and the company's own memory-focused posts are recent.
2. Repetitive / non-reactive output
This is the most consistently documented complaint in the set — not because any single report is dramatic, but because it recurs across four different products on four different platforms, spanning four years.
On itch.io, in a thread about the game Formamorph, player whitdragons2 wrote on 2025-06-09: "the ai are dumb and lot of repeat dialogue," and a second commenter, loblawbob72, corroborated in the same thread: "Haven't been able to stop the repeat dialogue though."
On a second itch.io title, Out of Ctrl, three separate commenters independently described the same pattern in their own words: "responses are the same no matter what you type" (BloodyAngel88), "it responds the same no matter the input" (hostetej33), and "it's the same response no matter what you say" (mae/bunniimae).
On the Character.AI App Store page, reviewer Saltyfox808 wrote on 2024-06-17 that repeating phrases was one of several recurring quality issues. And on the AI Dungeon App Store page, reviewer Spy149 wrote as far back as 2021-07-05 that "the AI started to describe the room I was already in… It was word for word" — the oldest dated instance in this set, showing the complaint isn't new.
Four products, four platforms, 2021 to 2025 — that spread is what makes this the best-evidenced complaint here, rather than one app's isolated bug.
3. The AI breaks character and narrates as an assistant
A subtler complaint than repetition: instead of staying inside the fiction, the AI steps outside it — narrating like a system message, rewriting a scene the player was already in, or taking over a character the player was controlling.
The clearest account of this comes from the same Medium essay cited above. SocialTipster wrote: "Writers report scenes being rewritten mid-conversation, villains turned into heroes, and entire personas swapped." The essay frames the shift starkly: "What used to be collaboration now feels like competition. The AI does not just respond anymore. It takes over." And specifically on tone control: "When a bot changes pronouns, rewrites scenes, or tones down conflict, it breaks the illusion of shared authorship."
This complaint also shows up in actual search behavior, not just review text: "character ai out of character" is a real, recorded search query (10 searches per month, difficulty 31, per Ahrefs) — a small number, but notable because it's the only complaint-shaped phrase in this category with enough competing content that Ahrefs can score its difficulty at all, meaning enough people are already searching and writing about it to register as a competitive term.
This is a moderate-strength complaint compared to memory loss or repetition — it rests on one detailed first-person account plus real keyword demand, rather than several independently corroborating reviewers.
4. Content filters interrupt non-explicit scenes
This complaint has two distinct eras, and conflating them would be dishonest — so they're kept separate here.
2021, a specific and since-superseded incident. In early 2021, AI Dungeon added content moderation after a discovery that the platform could generate sexual content involving minors. The resulting filter rollout was widely criticized at the time as both overbroad and privacy-invasive. Player @Sh1ptoast wrote on 2021-05-01: "Great idea, just a few problems 1 this has come with massive privacy violations 2 the censor is completely broken…" and player @emecho4 asked on 2021-04-28: "Does this mean you are reading unshared private stories?" That specific moderation rollout is a 2021 event on one product — not a description of how any current AI roleplay filter behaves, and it shouldn't be read as one.
2025-2026, a related but separate pattern. More recent reports describe a similar tension on a different platform: Robo Rhythms, in a piece published 2025-06-12 and updated 2026-07-31, wrote: "One moment you're immersed in a dark knight fantasy or crafting an emotional climax, and the next, your bot is frozen because you dared to write 'he pulls her close.'" and "You can write brutal violence, but suggest a romantic gesture, and suddenly the system locks up."
The throughline across both eras isn't one bug — it's a structural tension: a filter tuned to catch explicit content will also catch some fraction of non-explicit dramatic and romantic writing, and that tradeoff has recurred across more than one product and more than one era of the category. Character.AI has publicly named filter quality — a "finetuned filter" — among its stated priorities as of 2026, which suggests the company itself treats this as unresolved enough to keep working on rather than a solved problem.
5. Cost, rate limits, and paying again to keep playing
AI roleplay is compute-intensive — every response the AI generates costs the platform money — and that shows up downstream as credits, tokens, or tiers that players run out of mid-session.
AI Dungeon's own help center is the most direct evidence here, because it's the company explaining its own pricing decisions in response to player feedback. In its documentation on Shadow Tiers, AI Dungeon wrote: "Those players often gave feedback about wanting a more fluid and immersive experience, without worrying about their credits running out and having to buy more," and separately: "Many players also gave feedback about how the monthly credits offered in the packages wasn't enough for the entire month." The same document is candid that its higher tier isn't for most players: "Shadow Tiers are not the right option for the average player — not even the above average player."
Independent player reviews on a different product describe the same friction from the other side. On itch.io, in a comment thread for AI2U: With You 'Til The End, player grimjamzzz wrote: "Game asks for 10 USD then asks you to spend more so you can keep playing using tokens you have to buy." A second commenter, Wilfre, added in the same thread: "the token system is pretty scummy on a full priced game, a system like this should not be on a full priced product." A third, SirensLure, distinguished the two purchases explicitly: "I had 0 issues supporting the game and buying it once. What I do take issue is having to pay to play the game again."
Two different products, one from the platform itself and one from independent players who'd already paid once — the same shape of complaint either way: the initial purchase doesn't cover as much play as expected.
6. No real stakes — nothing you do can be lost or failed
This is the complaint with the least first-person review evidence and the strongest structural evidence — and it's worth explaining why those two things can both be true.
Freeform AI roleplay tends to accommodate the player. If you fail, the story usually doesn't record the failure; if a character "dies," the AI can simply forget it happened later; whatever you try, the narrative tends to bend to make it work rather than let it break something permanently. That shape of complaint is easy to describe in the abstract but hard to pin to a single dated player quote — which is exactly why the strongest evidence for it isn't a review at all.
In April 2026, Latitude — the company behind AI Dungeon, the category's largest incumbent — launched a second, structurally different product called Voyage, built specifically around persistent consequences. Two independent outlets covered the launch. TechCrunch (2026-04-21) reported that Voyage's "World Engine… leverages multiple AI systems that can narrate actions, manage gameplay, track characters and objects, and remember backstories and relationships, ensuring continuity throughout the game," and quoted Latitude CEO Nick Walton describing it as a system with "deterministic systems, challenges, progression, and persistence." On NPC memory specifically, the piece notes: "if you betray a character's trust, they may choose to avoid you or become a rival in future encounters." Early testing scale, per the same report: over 160,000 unique AI-generated characters, with players averaging nearly 3,000 gameplay choices each.
GamesBeat (2026-04-23) described the same system in close to identical terms — strong independent corroboration that both outlets are accurately quoting the same underlying product language: the World Engine is "a deterministic system that governs world state, character memory, and rule enforcement" that "tracks health, inventory, currency, geography, relationships and long-term consequences across thousands of turns so the AI can't hallucinate them away." Walton is quoted saying the company is "moving from static narratives to simulated worlds," and that Voyage "represents five years of R&D and six prototype engines."
The company that owns the category's biggest existing product rebuilt from scratch, across five years and six prototypes, specifically to make consequences stick. That's a stronger signal than any single player review — the party with the most information about why players leave concluded the no-stakes problem was real enough to justify a new product.
What a rules-driven design changes
Read together, these six complaints share a root cause more often than not: a model generating freeform text, with nothing underneath it tracking state, has no mechanism to remember what happened, no reason to vary its output, no fixed boundary between narrator and character, and no way to make a failure or a death stick. Voyage's own framing — a deterministic engine underneath the generative layer, tracking health, inventory, relationships, and consequences turn over turn — is one direct response to that gap.
That's the same reasoning behind how we build at /adventure: rules and state that sit underneath the AI's writing, so a roll matters and what happened last session is still true this session. It doesn't fix every complaint here — filter tuning and pricing are product decisions, not structural ones — but it addresses the two that are structural: memory and stakes. For the wider map of what's playable in the category, the best AI games guide and the best AI RPG games roundup cover it in more depth.
Sources
- AI Dungeon — App Store reviews
- Character.AI — App Store reviews
- SocialTipster, "Character AI Roleplay Is Falling Apart", Medium, 2025-10-28
- Formamorph — itch.io comments
- Out of Ctrl — itch.io comments
- AI2U: With You 'Til The End — itch.io comments
- AI Dungeon Help Center — About Shadow Tiers
- Character.AI Blog — Memory
- Character.AI Blog — Helping characters remember what matters most
- TechSpot, "Machine learning text adventure AI Dungeon now censored," 2021
- Robo Rhythms, "Character AI Filters Are Ruining Everything", published 2025-06-12, updated 2026-07-31
- Lauren Forristal, "Voyage is an AI RPG platform for creating custom gaming worlds with AI-generated NPC interactions", TechCrunch, 2026-04-21
- Sam Smith, "Latitude launches AI adventure Voyage to kill the scripted RPG", GamesBeat, 2026-04-23