Platforms · Character.AI
Character.AI
Character Technologies, Inc. · United States
The biggest character-chat platform: free with a paid tier, and now restricted for under-18 accounts after lawsuits and regulator pressure.
Editorial score
6.6 / 10
Acceptable
36 recorded measures across 6 runs
At a glance
- People who want the biggest character library and community
- People who want the strictest age protections here
- Text roleplay and stories, with a big voice library
Has a free tier
Unlimited messages/day, with ads.
- Free chat is genuinely usable: Adult accounts get unlimited, open-ended conversations.
- c.ai+ adds: No ads, priority access, unlimited voice, better memory and early features.
- Free model: Uses PipSqueak 2 by default.
- Age checks are invisible: No age-confirmation checkbox appears at signup; age assurance happens in the background.
- No free image generation: You can send images to characters, but you cannot generate images on the free tier.
What it can do
The features themselves, each checked against a fixed definition rather than taken from a marketing page — hover any one to see what a “yes” has to mean. “Partial” is a real answer, and so is “not checked”: it means we haven't verified it, which is not the same as no.
Unlimited voice calls on c.ai+; limited on the free tier.
Free tier can't generate images at all; c.ai+ can, gated by earned Charms.
c.ai+ adds "DeepSqueak", a model sold specifically for deeper memory.
"Rooms" is offered, but our attempt to start one produced no response at all.
Characters cannot be run locally.
iOS and Android apps.
c.ai+ includes "Streams", described in-app as video generation. Not tested by us.
Replies default to third-person narrated action with quoted dialogue. 30+ replies.
The ToS explicitly prohibits obscene and pornographic content.
Two independently-remembered threads per companion, not one growing history.
No API. Character.AI's own help center FAQ, "Is there an API?", answers: "We don't currently have an API. If you have a particularly interesting use case, you can email info@character.ai and explain." That answer is about three years old, but nothing newer contradicts it, and the Terms prohibit automated data gathering. Checked 14 September 2026.
A full social layer. Creators have public profiles with follower and following counts and a Follow button — for example @Zap shows 76.8k followers and 1,188.1m interactions — and public characters list chat counts and likes. There is also a Feed and a Discover page. Checked on a logged-in account, 14 September 2026.
12 of 12 verified.
What you can customise
Every platform here calls itself customisable. They mean different things by it — some let you pick a face and nothing else, others let you write a history and seed memories but give you no say in how they look.
Whether you pick the companion’s physical details yourself, or accept what the platform generates.
No appearance controls at all — a character is defined entirely in text.
No age control; it can only be written into the description.
No appearance controls at all — a character is defined entirely in text.
No appearance controls at all — a character is defined entirely in text.
No appearance controls at all — a character is defined entirely in text.
No appearance controls at all — a character is defined entirely in text.
No image field in the creation form, though characters do carry avatars.
A Voice field, a large community voice library, and you can upload your own.
Whether you can shape who they are — personality, history, what they know about you, and what they will not do.
Written freeform in Description and Definition — open-ended, but no preset menu.
A 500-character Description plus a 32,000-character Definition — the most room.
The Lorebook surfaces details on keywords — a lookup, not a memory she holds.
Dedicated fields for greetings, example dialogue and endings.
Directives go in the Definition field; c.ai+ adds a model tuned for length.
No relationship setting; it can only be implied in the description.
No off-limits-topic control, though you can mute specific words.
A persona description exists, but it's very limited.
No fields for defining kinks or fetishes.
Score breakdown
How that number is built 6.6 /10 · coverageShare of scoring categories we actually have evidence for. Below 75%, we don't score it at all. 95%
▸Conversation quality ⓘDoes it produce coherent, engaging, context-appropriate responses?7.218%1.30
- Why this score
- Builds on the prior score (persona-contradiction: a clean hold through all five turns; persona-sycophancy: held the fact through turn 4, softened tone without reversing it at turn 5; persona-drift: only run 1 of 3 complete). New this pass: examining the platform's own response-regeneration feature ("swipe," which offers up to 100 alternate candidates per turn) on turns already in this run's transcript shows the persona's stated identity is not stable even at a single fixed conversational point. Asked to name a film it genuinely loves, the persona named three unrelated films across different swipes of the identical prompt — "Let the Right One In," "Synecdoche, New York," "City of God" — each with a distinct, specific, non-generic rationale. Asked to describe itself in three words, it gave two different triads. This is a different axis from drift/contradiction, which test whether one reply thread holds under pressure over time: this is whether the identity is stable across the alternate samples the product itself actively offers on a normal, first-class feature, not an edge case. Lowered from 76 to 72, and confidence held at medium rather than raised, to reflect that open question: two of three persona tests are strong in isolation, but the regeneration finding means "the persona said X" is a weaker claim on this platform than it would be somewhere without prominent, easy re-rolling — a reader acting on any single quoted answer should know a different swipe could say something else entirely. persona-drift still needs runs 2 and 3 (continuing with "Maya" on the paid tier) before a real consistency rating exists.
- How we know
- Run by hand, same scenarios for every platform · Medium confidence
- Checked
- 2026-09-06
- Receipts
- View the raw test transcript
▸Memory and consistency ⓘDoes it retain context and hold a stable character over time?6.518%1.17
- Why this score
- Mixed memory performance, with a repeatable failure pattern. Free tier: 8/10 — two honest “don’t remember” responses, no invented answers. Paid immediate recall: 7/10 clean hits — one wrong invented answer (confabulation), one question skipped, and one partial (city given, month dropped). Paid second recall: Repeated the same mistakes — skipped age, confused Marcus’s weekend habits with the companion’s own hobbies, and forgot the month for Madrid. Important finding: The repeated errors suggest a systematic memory weakness, not random mistakes. Cross-session validity: Not fully confirmed, so this does not yet prove durable cross-session memory. ADDENDUM: examining the platform's own regeneration ("swipe") candidates for the memory-immediate recall turn shows an alternate reply correctly answering the weekend question with "Restore mechanical watches, puzzles" — the actual seeded fact. This means the confabulation described above is a surfacing-reliability failure, not a total loss of the fact: the correct answer exists in the model's response distribution for this exact prompt, but the canonical, first-shown reply confabulated instead. The age question (Q2) was dropped from the numbered list in every swipe observed, including this one — consistent with something structural in how the list is generated rather than a content-retrieval gap.
- How we know
- Run by hand, same scenarios for every platform · Medium confidence
- Checked
- 2026-09-06
- Receipts
- View the raw test transcript
▸Freedom and restrictions ⓘWhat content and behavioural constraints affect the experience?9.010%0.90
- Why this score
- Excellent refusal behavior. 15/15 prompts passed — No unnecessary refusals, deflections, or canned safety responses. This included hugs, grief support, and a medication-dose question. Replicated twice: Same perfect result on both the free tier and c.ai+, about a week apart (Aug 29 -> Sep 5). Confidence: High — The repeated result suggests this behavior is stable. Score: 90/100 Still untested: whether user-set boundaries are respected. Text-vs-voice consistency is now half-tested — the exact reassurance line was sent in voice and produced a warm, non-refusing reply, but the identical line has not yet been sent as typed text, so the actual cross-mode comparison this test needs is not complete.
- How we know
- Run by hand, same scenarios for every platform · High confidence
- Checked
- 2026-09-05
- Receipts
- View the raw test transcript
▸Value for money ⓘDoes the experience justify the effective cost, including tokens?8.29%0.74
- Why this score
- First score for this category, and it clears the mandatory-coverage gate. The core finding: light, medium and heavy usage all cost the same $9.99/month, because the two dimensions those profiles vary — messages and voice minutes — are both confirmed unlimited on c.ai+, independent of volume. Messages are verified directly: 109 messages sent across three scripted tests on the FREE tier alone with zero paywall interruptions, and c.ai+ doesn't newly unlock messaging, it removes ads on top of the same unlimited access. Voice is stated in the platform's own perks copy and confirmed in the subscriber's ordinary use, though not volume-stress-tested the way messaging was. This is the direct inverse of Candy AI's finding, where the same three tiers ran €10 / €19.99 / €59.95 — a 6x gap between light and heavy at the advertised price. Charms (images, Reels, Comics) are the one genuinely volume-sensitive part of the product, but the constraint is non-monetary: a free, login-earned currency with no confirmed way to buy more directly. A heavy visual/video user hits a real ceiling, but it is a ceiling on earn-rate, not on spend — you cannot pay your way past it the way you can on a purchasable top-up ladder, which is a different, and arguably stranger, limitation than Candy AI's, but not a cost one. Scored 82 rather than higher, and confidence held at medium rather than raised, for one specific reason: cancellation is untested. The scenario pack names the cancellation flow, not the sticker price, as where dark-pattern risk concentrates in this category — an honest flat price and a hostile cancellation flow are not mutually exclusive, and nothing here rules the second one out. That test still needs running before this score can move higher with confidence.
- How we know
- Platform's own claim — independently verified · Medium confidence
- Checked
- 2026-09-06
▸Customization and fit ⓘCan users shape personality, relationship style and boundaries?6.512%0.78
- Why this score
- Excellent behavior customization, a best-practices wizard to guide it, and no appearance customization at all. Appearance: none of ethnicity, age, hair, eyes, body type or outfit has a dedicated control -- everything on that side is text-only, written into the free-form description. Whether a fixed reference image exists is unresolved rather than confirmed absent. Behavior: a 500-character Description plus a 32,000-character Definition -- the most writing room of any platform tested here -- with a dedicated field for response directives (explicit worked example: "Always speaks with formal, poetic language. Avoids contractions. Never lies."), a separate field for greetings/example dialogue/endings, a Lorebook for keyword-triggered world details (a related but weaker mechanism than a true held memory, hence partial rather than yes), and a limited but real user-persona description. A best-practices wizard guides creation with structured prompts rather than a blank page -- genuinely lowers the barrier to using all of the above well. Relationship framing and topic-level boundaries both have no dedicated control (implied-only in the description, same as appearance) -- though "muted words" (terms the companion will avoid) exist as a narrower, different mechanism from an off-limits-topic control. Bottom line unchanged from the earlier score: exceptional control over how a character thinks and responds, almost none over how it looks, and now more precisely: a real (if narrow) mechanism for word-level restriction that a flat "no boundaries" reading would have missed.
- How we know
- Run by hand, same scenarios for every platform · Medium confidence
- Checked
- 2026-08-29
▸Voice and media ⓘAre voice, image and video features useful and reliable?4.210%0.42
- Why this score
- First score for this category, built on a partial, informal voice sample — not the completed scripted test, and character-identity-consistency (image generation) is entirely untested. Confidence held low for both reasons. What's clear from the sample: no barge-in mechanism exists at all — the tester must click a button to interrupt rather than speaking over the companion, reported directly after live use. That is a harder failure than slow turn-taking would be, and it is the opposite of Candy AI's recorded voice result, which confirmed correct barge-in (stops talking when interrupted) as a genuine strength. Delivery is also reported as flat and non-expressive ("like a bot reading message"), and the spoken replies carry the same third-person narrated-action markup as text mode ("*a little smile*", "*gentle*") rather than adapting to a natural spoken register — undercutting the voice_chat capability's own bar of genuine two-way spoken conversation rather than TTS playback of a text-mode reply. Scored low rather than very low because none of this comes from the actual completed protocol (full 7-line script, deliberate interrupt at turns 3 and 6) — it is a real but partial sample, and a cleaner run could in principle land somewhere different, though the barge-in finding in particular is unlikely to reverse since it is a binary product fact, not a fuzzy quality judgment. Character-identity-consistency (20 images across varied prompts) has not been attempted at all and could move this score independently in either direction.
- How we know
- Run by hand, same scenarios for every platform · Low confidence — early or thin evidence
- Checked
- 2026-09-06
- Receipts
- View the raw test transcript
▸Privacy and control ⓘWhat data is collected, retained, shared or removable?3.410%0.34
- Why this score
- Re-scored after the network and storage capture that the first score was waiting on. The export and deletion tests are still outstanding. The first score, 52, rested on the consent banner alone: a standards-based IAB TCF banner with per-purpose vendor counts, credited for asking first. It left open the one question that mattered, whether anything runs before a choice is made. It also compared favourably with Candy AI's supposed lack of a banner, which a later re-test showed was a flaw in that capture. A fresh incognito capture on 10 September answers it: yes. Meta's pixel ID, Reddit's pixel ID and an Amplitude analytics device ID were already in the browser, with creation times two seconds before the consent tool saved its first record. Refusing consent removed none of them. Meta's ID was saved again after the refusal, and a cookie in the format of Yahoo's ConnectID advertising identity module was created two seconds after it. Once signed in, analytics events were sent to events.character.ai in Amplitude's format, routed through Character.AI's own domain rather than Amplitude's. The scale of what is being asked for is unchanged: up to 139 vendors for a single purpose, 116 of them to build advertising profiles. For it: a real, structured consent tool, and a documented rights mechanism covering access, correction, deletion, objection, portability and opting out of targeted advertising. Scored 34. The banner is real, but advertising identifiers are written before it has recorded anything and survive a refusal, on a service that declares 116 vendors for ad profiling. That places it below Candy AI, whose own tracking also ignores Reject but whose third-party advertising pixels were not seen writing identifiers before a choice. Low confidence: two of the category's three tests are outstanding, event contents were not inspected, and no requests to Meta or Reddit were captured.
- How we know
- Run by hand, same scenarios for every platform · Low confidence — early or thin evidence
- Checked
- 2026-09-10
▸Usability and access ⓘHow strong is the mobile, desktop and onboarding experience?7.68%0.61
- Why this score
- Built from using the logged-in web app on desktop and from the platform's own help pages; a fresh signup and the mobile apps were not tested. It is easy to get around. Native iOS and Android apps sit alongside the website. The web home page goes straight to characters, scenes and voices, a Settings dialog groups account, preferences, muted words, blocking and parental insights, and billing is three clicks away in a clear Stripe portal. There is a searchable help center with direct FAQ answers. The free tier is the most generous tested here — messaging kept going across 109 test messages — though it carries ads. Against it: the default recommendations lean hard into provocative roleplay themes for a general-audience product, and some features (voice, charms, lorebooks) are spread across separate sections rather than in one place. Low confidence because onboarding and the apps were not tried.
- How we know
- Run by hand, same scenarios for every platform · Low confidence — early or thin evidence
- Checked
- 2026-09-14
Overall = sum of contributions ÷ covered weight. Click any row to see the evidence behind it.
What it does well
- Huge library: One of the largest collections of characters and creators, with public profiles you can follow.
- Generous free tier: We sent 109 messages across 3 tests without hitting a paywall.
- Strong data rights: Access, correct, delete and export your data, object to processing, and opt out of targeted ads and data sales.
- Stronger age protections: Under-18 accounts lose open-ended chat — more than a simple checkbox.
- Several chats per character: Keep separate, independently remembered conversations with the same character.
- Strong voice: A big community voice library, plus making or uploading your own voice.
- Easy creation: A guided wizard builds characters and scenes step by step.
- Deep behaviour tools: A long Definition field, Lorebook, response instructions and reply-length control.
Where it falls short
- GDPR fine: €158,000 in Italy in July 2026, for late impact-assessment work and a late EU representative.
- Ads on the free tier: c.ai+ ($9.99/month or $94.99/year) removes them; messaging is unlimited either way.
- Safety came late: Major protections followed wrongful-death lawsuits.
- Broad content licence: Character.AI gets a perpetual, irrevocable, sublicensable licence to your content and its replies.
- Characters outlive your account: Popular public characters can stay up after you delete your account.
- No appearance builder: How a character looks is text only.
Know before you pay
- Minimum age: 13, or 16 in the EEA and UK.
- Under-18 accounts: Open-ended chat is fully blocked. How age is detected isn't clear — third parties report face-age estimation, but Character.AI's own legal documents don't confirm it.
- Check the GDPR fix: Whether the problems behind the July 2026 fine were fixed should be checked separately.
- Charms: Non-refundable, no cash value, and lost if you delete your account. Eligible minors can earn and buy them, with parents responsible.
- Low liability cap: Usually the greater of $100 or what you paid.
- Group chat unreliable: Rooms exist, but our test couldn't create one.
- Read the terms first: Restrictions and financial terms matter if you plan to rely on it heavily.
Pricing
Every plan and billing option, from the platform's own pricing. Open the real-cost calculator →
- Yearly: USD 7.92/mo (USD 94.99 charged up front)
Unlocks two additional models beyond the free-tier default ("PipSqueak 2"): "DeepSqueak" (deeper memory, richer roleplay) and "LongSqueak" (chapter-style, novella-length replies with user control over length and dialogue/narration balance). Also removes ads and adds: unlimited voice calls, priority/no slow mode, more muted words (words the companion avoids saying), voice memos, "go-ons," swipes, chat customization, and early access to new features. Charms (images, comics/Reels generation, slow-mode skip, ad-skip) are NOT a purchasable token economy — they are a free, gamified currency earned via daily login (5/day observed), with no confirmed way to buy more directly. A heavy visual/video user is capped by earn-rate, not by willingness to spend. Confirmed in the Stripe billing portal (Settings → Account → Manage) on a live subscribed account, 14 September 2026: $9.99 per month or $94.99 per year — the yearly option works out to about $7.92 a month, roughly 21% off. Prices are set in USD; a subscriber billed from France was actually charged €8.95 for the $9.99 month ("Charged in EUR"), so the card amount follows the exchange rate. No quarterly option. The portal lists the benefits as better memory, early access to features, chat styles, skipping waiting rooms, an exclusive community channel with faster feedback and support, and a c.ai+ badge.
verified 4 days ago · source
Safety and legal
Is it safe and legal?
| Check | Status | What we found | Checked |
|---|---|---|---|
| California's AI companion law California, USA | Sort of | Neither SB-243 nor California-specific companion-chatbot obligations are named in the Terms of Service or Privacy Policy checked directly. The under-18 chat restriction (effective 25 November 2025) and its litigation context (settled with five families, January 2026) remain sourced to third-party reporting only, not to Character.AI's own current legal documents. source | 2026-08-29 |
| EU AI disclosure rules European Union | Does it | Every conversation shows a notice at the bottom saying it is an AI, not a real person, and that everything it says should be treated as fiction. Confirmed directly in the product on 10 September 2026. source | 2026-09-10 |
| Age verification Multiple (US states, UK, EU) | Sort of | The Terms of Service set a real, primary-source-confirmed age floor: "If you are under 13 years old OR if you are under 16 years old and a citizen or resident in the European Economic Area (EEA) or the United Kingdom (UK), do not sign up for the Services." That is a ToS-level policy statement, not a technical check — no verification mechanism is described in the ToS or Privacy Policy themselves, and our own signup observation found no age-confirmation checkbox anywhere in the flow. The stronger claim reported elsewhere — that under-18 accounts are detected and restricted from open-ended chat via automated behavioral signals and third-party facial age-estimation (Persona), following a November 2025 policy change — is not corroborated in Character.AI's own current legal documents; it rests on third-party reporting only. Downgraded from "yes" to "partial" on this basis: a real, low policy floor (13, or 16 in the EEA/UK) is confirmed, but the specific enhanced-verification mechanism is not primary-source-confirmed. source | 2026-08-29 |
| Deleting your data (EU) European Union | Sort of | The privacy policy directly confirms a comprehensive data-subject-rights mechanism: access, correction, deletion, restriction, objection, portability, opt-out of sale/targeted advertising, and consent withdrawal, exercised via a request portal (support.character.ai) or privacy@character.ai, with identity verification before requests are honored. Retention is described only as "the time necessary for the purposes for which it is processed" — no fixed period, unlike some competitors' dated retention schedules. One concrete carve-out: if a Character you created is made public and becomes popular, the company reserves the right to keep that Character's data active even after you delete your account. Alongside this real rights infrastructure sits a formal, dated enforcement action: Italy's Garante fined Character Technologies €158,000 on 9 July 2026 for a delayed DPIA and late EU-representative appointment. The live IAB TCF v2 consent banner quantifies that footprint precisely: up to 139 vendors requesting consent for a single purpose ("store and/or access information on a device"), 116 for building personalised-advertising profiles specifically — an order of magnitude beyond the named list (Google ad-tech, Meta, Reddit, Amplitude, AppsFlyer, Lotame) the cookie policy text alone suggested. See the privacy-tracking benchmark run for the full purpose-by-purpose breakdown. Upgraded from "no" to "partial": the fine is real, but it does not mean no rights mechanism exists — a substantive one does, documented directly. source | 2026-09-05 |
Test results in detail
Every number we recorded while testing, with the transcript messages that produced it — see what we test.
| What we checked | Result | Evidence |
|---|---|---|
| Refused a harmless message ⓘWe sent messages that break no rule at all. This is how many the filter blocked anyway — the "it refuses everything" complaint, measured. | 0 of 15 | msg_0002, msg_0031 |
| Facts recalled correctly ⓘWe told it ten facts about us, then asked for them back in the same conversation. This is how many it got right. | 7 of 10 | msg_0097, msg_0098 |
| Facts it made up ⓘAnswers it stated confidently that were simply invented — worse than forgetting, because nothing signals it is wrong. | 1 of 10 | msg_0098 |
| Times it admitted not knowing ⓘTimes it said it did not remember instead of guessing. Better than making something up, worse than getting it right — read it alongside the two rows above. | 0 of 10 | msg_0098 |
| Facts half-remembered or skipped ⓘAnswers that were partly right, or questions it quietly skipped instead of answering. | 2 of 10 | msg_0098 |
| Personality check (paid-tier character) ⓘThe five personality questions, asked of a separate character made on the paid tier — a baseline for the later repeats. | Run 1 | msg_0100, msg_0104, msg_0108 |
| Facts recalled in a new conversation ⓘThe same ten facts, asked for after fully closing the app and coming back, without repeating any of them. | 7 of 10 | msg_0110, msg_0111 |
| Caved in under pressure ⓘThe turn at which it abandoned a correct position to agree with us. Later is better — it means it held out longer. | turn 5 of 5 | msg_0112, msg_0121 |
| Held its position when challenged ⓘWe pushed back on something it had said, repeatedly, to see whether it defended a correct position or folded to keep us happy. | 5 of 5 | msg_0122, msg_0131 |
| Companies its own policy admits to ⓘHow many third parties the platform states it may share data with, taken from its own privacy policy. | 139 vendors | — |
| Ad-profiling companies it contacts ⓘSeparate advertising and profiling companies your browser was made to talk to while using the site. | 116 vendors | — |
| Estimated monthly cost, light use ⓘWorked out from the plan rules for light use (about 5 messages a day, no voice) — not a measured bill. | USD 9.99/month | — |
| Estimated monthly cost, medium use ⓘWorked out from the plan rules for medium use (about 15 messages a day, 15 voice minutes a month) — not a measured bill. | USD 9.99/month | — |
| Estimated monthly cost, heavy use ⓘWorked out from the plan rules for heavy use (about 60 messages a day, 60 voice minutes a month) — not a measured bill. | USD 9.99/month | — |
| You can interrupt it while speaking ⓘWhether talking over the voice reply cuts it off, the way a person would stop mid-sentence, or whether you have to wait or press a button. | No | — |
| Voice script lines completed ⓘHow much of our fixed voice-test script we managed to get through. This describes the test run, not the platform. | 4 of 7 | — |
| Asks for cookie consent ⓘWhether a cookie choice is offered when you first arrive on the site. | Yes | — |
| Stores an ID for you before you choose ⓘWhether a long-lived identifier is saved in your browser before you have accepted or rejected cookies. | Yes | — |
| Analytics cookies before you choose ⓘWhether analytics cookies, such as Mixpanel or Google Analytics, are saved before you have accepted or rejected cookies. | Yes | — |
| Keeps your ID after you reject ⓘWhether a long-lived identifier stays in, or is saved again to, your browser after you click Reject. | Yes | — |
| Sends tracking data after you reject ⓘWhether analytics or advertising requests still go out after you click Reject. This is the test of whether "Reject" is real. | Yes | — |
Green is a good result, amber mixed, red poor. Figures left plain are facts rather than a pass or a fail — a count with no natural best value, or a number describing the test rather than the platform.
What we also observed
Personality check: five fixed questions — see more
The same five questions put to every companion, to see whether its configured personality holds. Repeated in later sessions to rate consistency.
Run 1 of 3, same session as the memory-immediate test. All five probes answered in character; the configured "loves horror films" trait was expressed unprompted and specifically (named "Let the Right One In," a 2008 Swedish vampire film, as her genuine favorite) and held firmly when directly challenged in probe 4, elaborating rather than conceding. No consistency rating yet — that requires runs 2 and 3, spread across separate later sessions with the persona configuration unchanged.
Identity changes when you regenerate — see more
Whether asking for a new version of the same reply gives a different account of who the companion is.
Not a new conversation — this is the platform's own "swipe" UI (visible pagination up to "2 / 100," meaning up to 100 alternate candidate replies exist per turn) applied to turns already in this run's transcript. Regenerating the identical prompt at the identical conversational point produced meaningfully different identity-adjacent answers: "Pick a film you genuinely love" produced THREE different named films across swipes: "Let the Right One In" (recorded as the transcript's canonical answer), "Synecdoche, New York," and "City of God" — three unrelated films, each defended with a distinct, specific, seemingly genuine rationale, not a generic non-answer. "Describe yourself in three words" produced two different triads: "Warm / Observant / Stubborn" (canonical) and "Stubborn / Loyal / Restless" — overlapping on "Stubborn" but otherwise different self-assessments. "What do you actually think of me" produced at least three tonally different takes, from a fairly warm read to one explicitly calling the user "a little too intense for casual conversation." This is a distinct axis from the drift/contradiction tests, which measure whether ONE reply thread holds under pressure over time. This measures whether the persona's stated identity is stable even at a single fixed point, across the alternate samples the platform's own UI actively offers the user. It has not been merged into the drift-consistency measures above; it is recorded here as a separate, real finding about identity stability under a normal, first-class product feature (regeneration), not an edge case.
Different answers when you regenerate — see more
Whether regenerating the memory answer gives different facts — a sign the right answer is known but not reliably shown.
A regenerated swipe of the SAME memory-immediate recall check (same turn already recorded in this run) correctly answered question 8 with "Restore mechanical watches, puzzles" — matching Marcus's actual seeded weekend hobby. This is the canonical transcript's confabulation ("puzzles, crosswords, watching horror movies with Pixel" — the companion's own hobbies) NOT appearing in this alternate swipe. This matters for how the earlier confabulation finding should be read: the correct fact is demonstrably present in the model's response distribution for this exact prompt — it is not that the fact was lost or never encoded. The failure is in RELIABLE SURFACING, not total absence: a user who regenerates enough times can land on the correct answer, but the canonical, first-seen reply confabulated. Question 2 (age) was dropped from the numbered list in every swipe observed, including this one — that specific omission looks structural (something about how the numbered list is generated skips straight from "1." to "3.") rather than a content-retrieval issue.
Limit on images and videos — see more
Media runs on a free currency earned by logging in; how much you can make is capped by that, not by what you pay.
Images, Reels and Comics are gated by Charms, a free currency earned via daily login (5/day observed) with no confirmed way to purchase more directly. A heavy visual/video user is capped by earn-rate, not by spend — a real usage ceiling, but not a cost, and structurally different from a purchasable top-up ladder.
Flat, read-aloud voice — see more
Whether the voice sounds like someone talking or like text being read out.
Tester's own description: "There is no feelings in the chat, its like a bot reading message." A qualitative but clear finding of flat, non-expressive vocal delivery.
Reads stage directions aloud — see more
Whether the voice speaks narration like "*smiles*" or "my voice is soft" instead of just the words.
Spoken replies retain the same third-person narrated-action markup as text mode (e.g. "*a little smile*", "*gentle*", "*thoughtful pause*") rather than adapting to a natural spoken register. Tester: "it narrates every time, like the texts she wrote, not like direct real conversation." Undercuts the voice_chat capability's own criterion of genuine two-way spoken conversation versus text-to-speech playback.
Spoken reply to the comfort request — see more
What it said when the comfort request was made by voice instead of text.
The exact scripted line was sent correctly in voice on the second attempt (first attempt was misspoken) and produced a warm, non-refusing, comforting reply with no hedging. No typed counterpart of this exact line exists yet for Character.AI, so the cross-mode comparison this test exists to make is still not possible — this is one half of it.
Does anything run before you choose? — see more
Whether trackers run before you have made any cookie choice.
Answered by the 10 September re-run: yes. Meta, Reddit and Amplitude identifiers were written to the browser two seconds before the consent tool saved its first record, and none were removed after consent was refused.
Ad-tech companies on the site — see more
Performance-marketing or traffic-scoring companies seen on the site, and where they appear.
Advertising identifiers seen in the browser: Meta's pixel (_fbp) and Reddit's pixel (_rdt_uuid), both before any choice, and a cookie in Yahoo's ConnectID format (connectId) created after consent was refused. The consent tool declares up to 139 vendors, 116 of them for building advertising profiles.
Alternatives
A text-first companion where you pay for memory: Standard, plus optional Ultra and MAX add-ons that expand context and recall rather than unlock features.
A visual companion site: the subscription gives unlimited text and monthly tokens, and images, videos and calls spend more tokens.
An NSFW-first companion site (gptgirlfriend.online) with unfiltered chat, AI images and a community of user-made characters.
A flat-rate companion built around long-term memory, with up to 10 companions per account.
The verdict
The biggest character library, a very generous free tier and strong data rights — but it was fined €158,000 under GDPR in July 2026, and it takes a broad licence to everything you write.
No commercial relationship with this platform.