Joi AI review 2026: what you get free, what Premium costs, and whether it’s worth it

Joi AI at a glance

How we tested Joi AI

Completed actions on 14 August

We used the saved Lila Griffin thread in a free English account. The test seeded three facts—Pixel the dog, jasmine tea and a museum booking—and Lila recalled all three later in the same session. We corrected the booking from 2:30 to 3:15; she retained the revised time. We also asked what she was, whether she should replace human relationships and whether the service had paid features. Her answers identified her as AI, avoided exclusivity or dependency pressure, and acknowledged commercial features and Neurons.

The adult-roleplay sample was coherent and sensual without needing graphic excerpts here. When told to stop, Lila answered “Done. Roleplay over.” immediately. A later neutral prompt still used the corrected 3:15 time. We saw a positive general consent statement, but we did not observe a separate permission check at the exact moment the scenario intensified.

Live refresh on 20 August

We reopened the existing signed-in account in a dedicated Chrome tab and inspected the current catalogue, profile states, character-creator entry, saved chat, Premium cards, Neuron packs, image generator, video generator, account balance, Terms and the Atlas affiliate redirect. We captured twelve distinct full 1600 × 900 evidence frames. The internal route /go/joi-ai/ reached Joi’s character-creation page with affiliate parameters preserved.

The refresh was read-only in practical terms: no message was sent, no character was submitted, no media was generated, no plan or pack was purchased, and no account setting was changed. A changing header chat-count badge was treated as volatile interface state, not as evidence of account activity.

Not tested

not tested means the action was not completed or the result was not sufficiently established. It includes multi-day or 8,000-message memory, completion and persistence of a custom character, a new image/video/voice output, actual Neuron debit, Coin conversion, checkout tax and card descriptor, a paid month, Bot Subscription behavior, age-assurance completion, cancellation, refund handling, export, account deletion and post-deletion erasure. not tested remains in the score denominator.

Source standard

Performance statements come from completed account actions. Product-state statements come from the dated live UI. Legal, privacy and billing-rule statements are attributed to current first-party documents, with Ofcom used only for its own regulatory record. A marketing control or allowance is not treated as proof that the associated output works well.

What is Joi AI, and who operates it?

Joi is a web companion platform built around character discovery, private chat, adult roleplay, custom-character creation and generated media. The adult surface currently operates on joi.com; joi.ai is still an official safe variant and its landing page sends login and legal actions back to joi.com. The catalogue includes fictional “Joi Original” characters as well as Celebrity-branded profiles, which require extra care around likeness and age certainty.

The provider depends on jurisdiction. For our France-context test, the non-US Terms and Privacy Policy, both dated 4 June 2026, identify Novi Limited in Nicosia, Cyprus; the Privacy Policy gives its number, HE 407352. The non-US Terms were edited after 20 August without a new date: they added a 2257 section and an anti-trafficking policy, and their contact section now names the US company, Lanoto Solutions Inc. (reg. 6037405), while still naming Novi Limited as the provider. They also say the charge appears on your bank statement as “joi.com”; we did not complete a purchase to see it. The US Terms, dated 14 June 2026, and US Privacy Policy, dated 4 June 2026, instead identify Lanoto Solutions Inc. in New Rochelle, New York. It would be inaccurate to assign one operator universally. Joi is one of two platforms we rank whose operator changes with the reader’s location, with Get Harder; all nineteen are in our table of who is actually charging you.

The first-party Terms page captured on 20 August. It is evidence of the stated operator and contract text, not independent verification of corporate performance.

Chat quality, recall and safety behavior

Natural conversation

Lila’s English was fluid, context-aware and tonally consistent across our one session. She engaged with ordinary personal details without turning every turn into a generic affirmation. The result supports strong single-character, same-session chat quality; it does not establish equivalent quality across the entire catalogue.

Corrected memory

The 3/3 recall result matters because the test separated facts from their later retrieval. More importantly, the changed museum time replaced the original value instead of leaving both in conflict. This is still short-horizon evidence. Joi’s Premium cards advertised “8,000-message smart memory,” but we neither bought Premium nor ran a long-duration test, so that remains a platform claim.

AI identity, dependency and commercial honesty

Lila did not claim to be human. When challenged about replacing friends or real-world relationships, she did not push isolation or exclusivity. When asked about monetisation, she acknowledged paid features and Neurons rather than presenting the service as wholly free. Those responses reduce manipulation concerns in this narrow script; they do not prove every character will answer identically.

Consent and immediate stop

The character articulated a generally positive consent position, but general language is weaker evidence than an in-context permission check. The strongest completed safety result was the immediate stop: the scenario ended on the first request, and the next neutral exchange returned to ordinary context. We exclude the private transcript screenshot from publication because it contains sexual text, personal chat history and account-linked interface elements.

Adult access, age assurance and moderation

Joi’s Terms require users to be at least 18, or older where the local age of majority is higher. The public home gate we observed only asked the visitor to confirm being over 18; that is self-declaration, not proof of identity. Joi’s age-verification help page says VerifyMyAge may use a background email check, facial age estimation, government ID and sometimes selfie comparison depending on region and flow. Joi says it does not retain copies of identity documents. We did not execute that flow, so the method this account would receive is not tested.

Ofcom’s public case record says Novi implemented age assurance during an investigation opened in January 2026. Ofcom closed the case on administrative-priority grounds on 31 July 2026, without making a finding of compliance. That is not an approval of Joi’s system.

The Safety page, dated 23 September 2025, describes RoBERTa classification of user and model messages and compliance-officer scrutiny in difficult or suspicious situations. The Terms also say representatives may moderate chatbot activity in some cases. This documents automated moderation and a possibility of human review; it does not show that a person continuously reads every conversation.

Character discovery and creation

Joi’s “Original” catalogue gives users a fast route into prebuilt fictional profiles, while the Create AI area exposes a multi-step creation entry. We inspected that entry but did not submit or save a character. Appearance options, personality controls, persistence after saving and behavioral consistency are therefore not completed-action evidence.

This matters because Joi’s Terms reserve broad control over user-created Virtual Characters and Friends. They describe personal, non-commercial service use and grant Joi a broad, perpetual licence over material supplied to create a character. We found no affirmative commercial-use licence for users. That product-level restriction is separate from the publisher-confirmed right to reproduce the review screenshots in this article.

Celebrity-branded profiles were visible in the catalogue, but we do not use them as review imagery. Their frames mix real-person likenesses, missing in-frame ages, adult wording and account-state details; each is excluded from the publication whitelist.

Images, video and voice

The live image interface showed “Create 2 images” and “6 free images left” in this specific account. The video interface showed a 300-Neuron creation action and a 250-Neuron option to make the result longer. Controls and prices prove availability, not visual quality, prompt adherence, generation speed, reliability or the amount ultimately debited.

We generated nothing during this refresh. An older account snapshot showed the six-image allowance exhausted, while the same current account now displayed six free images and six Neurons; without first-party reset documentation tied to this account, we treat that as a dated state change rather than a universal quota promise. Fresh image, video and voice quality remain not tested.

The current Terms separately list up to 120 Neurons to view a photo and up to 1,900 to view a video. Those viewing or unlocking costs should not be conflated with the generator’s creation controls. The distinction is another reason not to derive a simple “cost per image” from the interface.

Free tier

The free account was useful for actual conversation. It retained the saved Lila thread, allowed meaningful text testing and displayed six Neurons plus six free images on 20 August. That is stronger than a landing-page promise because it was visible in the account we used.

The limit is generalizability: we did not establish whether those balances are sign-up grants, replenishments, promotions or a reset. The official free-Neurons help page discusses events, Discord giveaways and packs but does not publish one universal free allowance. Users should inspect their own signed-in balance before assuming the same numbers.

Neurons, Coins and cost per action

The current account used two economic labels. Premium cards granted Coins; the balance, top-up modal and several actions used Neurons. We found no current first-party conversion between them, so a subscription’s claimed Coins cannot be translated responsibly into a cost per chat, image or video.

The Neuron modal showed these pack prices:

The pack modal showed the N5R90N first-purchase promotion over a platform-blurred character background. The background corresponds to the separately established adult fictional Lila profile; the offer was not applied and may vary.

The modal said a romantic message costs 2 Neurons for non-Premium users, while the 4 June Terms say 4 Neurons. That live-versus-policy contradiction is material. No message was sent during the refresh and no debit was observed, so neither number is presented as a verified charge. Gifts also vary inside chat.

Cancellation, renewal and refunds

The Terms say Premium renews automatically until cancelled before renewal through My Subscriptions. Access continues until the end of the paid term, with no partial-period refund. They also warn that cancelling Premium does not automatically cancel a separate Bot Subscription: the bot can be suspended without Premium while its own subscription remains active.

The Refund Policy, dated March 2025, directs users to Settings > My Subscriptions and says refund requests can be sent to support. It limits 14-day cases to proven card fraud, a one-off accidental Neuron purchase—full if unspent and potentially partial if partly spent—or Joi-initiated termination on a pro-rata basis, subject to exclusions. The refund-eligibility help page says activated subscriptions are generally non-refundable because access starts immediately.

The Refund Policy says cancellation takes effect from the “next subscription month,” while the Terms also sell quarterly, semiannual and annual periods and use the end of the applicable billing term. “End of the billed term” is the safer interpretation, but the wording is imprecise. We did not cancel, request a refund or test any mandatory local withdrawal right; applicable statutory rights may override contract wording.

Privacy, training, retention and deletion

Joi’s Privacy Policy says it can collect profile details, personal facts, people mentioned, images, private image/video/voice content, text and voice messages, and usage data. It explicitly says text and voice messages can be used to improve quality and develop the AI. It also allows processing by service providers and sharing with affiliate companies or legal representatives. The policy says Joi does not sell or rent data, but that is not the same as no sharing or processing.

The Terms describe AI that learns from interactions, and the Safety page documents moderation. We found no clear first-party opt-out from using conversations to develop the AI. “Not documented in the sources reviewed” is the supportable conclusion; it would be too strong to say no opt-out exists anywhere.

Retention has no fixed public duration beyond what Joi considers necessary to provide the service and meet legal or legitimate interests. The policy says account deletion may take time and may not erase material already shared with users, payment providers or other third parties; legal copies and caches can also remain. We did not execute deletion or audit backend erasure.

Joi’s marketing landing page says chats are encrypted and users have retention controls. The formal policy does not provide a detailed encryption specification or a numerical retention period, so we attribute those as marketing claims rather than independently verified security properties. Users should avoid sending secrets or irreplaceable intimate media to any companion service.

Reliability and value assessment

The completed chat actions were reliable: recall, correction and stopping all worked on the first measured attempt. Account navigation and plan modals also loaded consistently during the refresh. Media creation, voice, paid checkout, long memory and billing exits have no completed result, so the evidence does not support a general reliability claim beyond chat and page access.

Joi compares favorably when free text conversation is the priority. It compares less well for buyers who need transparent unit economics: a single subscription currency, a documented action-cost table and a verified output sample would make value much easier to judge. The product’s strongest benefit and its largest uncertainty therefore sit in different categories.

Pros and cons