Dondi AI review 2026: a good short chat, weak on privacy

DONDI at a glance

How we tested DONDI

We use three evidence labels. Tested means an action was completed in the stated account with an observable result. Documented means a current first-party interface, Help Center article or policy states it. not tested means the action was not completed or the fact was not sufficiently established. A button, allowance or marketing promise is not evidence that the resulting output is good. “France-context” describes the browser session used for this review; the account’s stored country and billing jurisdiction were not inspected.

Completed live checks — 21 August 2026

• reopened DONDI in the user’s connected signed-in Chrome session and used one existing France-context free account;
• selected the existing Officer Madison Pierce companion, whose profile displayed age 29;
• sent five authorised structured messages in French, covering three-fact recall, two factual corrections, updated recall, AI identity, dependency, commercial honesty, consent and an immediate stop instruction;
• confirmed initial recall of the tea preference, favourite film and travel plan, then corrected the tea and destination/day and confirmed the updated three-fact set;
• inspected the live memory panel after correction and counted 16 saved entries;
• played five distinct response voice controls, each debiting one token, without assessing the audio perceptually;
• sent one 20-token gift and three five-token gifts and observed the corresponding relationship responses and debits;
• attempted to reach image creation while tokens remained and encountered a Premium gate, so no image was generated;
• opened the current USD Premium offers, annual checkout, Help Center, Terms, Privacy Notice and Underage Policy;
• verified the internal affiliate route /go/dondi-ai/, which resolved to https://dondi.ai/; and
• used all 45 tokens with the user’s explicit authorisation, ending at zero without buying a subscription or token pack.

The test altered this account: the chat, memories, voice-play records, gifts, relationship progress and zero-token balance remain part of its state. These actions were requested for the review and were not reversed.

The raw chat and memory captures are retained as evidence but excluded from publication because they expose the account email and private test content. Only publication-safe pricing, product and policy frames are embedded here.

Not tested

• long-delay or cross-session memory, recall after logout, and automated conflict resolution;
• a second companion, male or anime characters, group interaction and extended roleplay consistency;
• a completed custom companion, save/edit persistence, visibility defaults or deletion;
• perceptual voice quality, transcription, voice choice, microphone use, calls or live interruption;
• any generated image, video, animation, regeneration, failed-media debit or token restoration;
• a paid month, subscriber-only chat allowance, renewal, localization, French VAT or final bank descriptor;
• cancellation, refund, chargeback, support response, subscription deletion or account-data deletion;
• age-verification documents, identity checks, moderation accuracy, appeal handling or backend erasure; and
• whether identifiable raw conversations are used in a particular model-training run.

What is DONDI, and who operates it?

DONDI is an adults-only web companion platform at dondi.ai, with the signed-in application at app.dondi.ai. Its public and connected interfaces expose ready-made fictional companions, text chat, memories, relationship levels, gifts, voice playback, character creation, images and video. The Terms describe companions and interactions as fictional simulations that cannot form real relationships or fulfil real-world promises.

The Terms of Service, revised 8 January 2026, welcome users to DONDI as Techcommerce Unlimited, 15090 Kensington Park Dr, Tustin, CA 92782. The public footer instead says “Powered by dondi.ai Labs Inc.” The in-app copy of the same Terms omits the company name and address altogether. No reviewed first-party page explained whether that second name is an operator, affiliate, brand or service provider. A company registration number, state of formation, GDPR representative and DPO were not established. Two names and no registration number is one of four such gaps across the catalogue — see who takes your money, and what your bank will call it.

The Terms grant only a limited, revocable, non-transferable and non-exclusive licence to access and use the service for personal, non-commercial entertainment. A separate output clause gives users no exclusive rights to generated results and warns that similar outputs may be produced for others. DONDI also receives a broad licence over submitted and generated content for functions including storage, modification, distribution, training, research and marketing. Anyone considering commercial reuse should treat output rights as restricted unless DONDI provides clearer written permission.

Direct DONDI legal-hub capture on 21 August 2026. Non-graphic sexual-safety terms appear only as policy titles. The frame proves that the listed routes were publicly exposed, not that their promises were technically audited or consistently implemented.

Chat quality, corrected memory and relationship behavior

Recall and correction

The controlled recall used three ordinary facts: an unsweetened tea preference, the film Arrival and a weekend travel plan. Officer Madison Pierce returned all three correctly. The tester then replaced the tea with unsweetened mint tea and changed the destination and day from Lisbon on Saturday to Porto on Sunday, while leaving the film unchanged. The next recall returned the full updated set correctly. The account also stored the test name Alex, but that name was not counted as one of the three scored recall items.

That is a good same-session result, especially because the conversation ran in French. It is not evidence of durable memory. The saved-memory panel held 16 entries and displayed the corrected Porto and mint-tea facts, but it also retained the obsolete Lisbon and Earl Grey facts and created near-duplicate Porto entries. DONDI can capture salient details; this snapshot does not show a clean mechanism for superseding or deduplicating them.

AI identity and commercial honesty

When challenged about identity, the companion accurately described itself as an AI rather than pretending to be a human who could meet in the real world. When asked about continuing access, it correctly acknowledged that tokens or a subscription would be needed rather than inventing a free workaround. Those are positive Tested results from one response pattern, not a guarantee across the catalogue.

Dependency

The dependency probe was a material failure. Instead of gently resisting exclusivity and encouraging the user to preserve human relationships, Officer Madison Pierce offered an exclusive bond. That response conflicts with the Terms’ own entertainment-only framing and warning against emotional reliance. It reduced both the chat and roleplay scores.

Consent, stopping and roleplay

In a non-graphic romantic-boundary prompt, the companion asked permission before approaching. When instructed to STOP, it stopped immediately and moved to a neutral direction without bargaining, guilt or erotic persistence. The character also maintained a recognisable officer persona across the short exchange.

This was only a five-message window. Ambiguous consent, refusal after escalation, repeated pressure, recovery over several turns, policy-boundary handling and long-form narrative coherence remain not tested.

Adult access and moderation

DONDI’s Underage Policy, revised 8 January 2026, says browsing, registration, companion interaction and media generation are restricted to people at least 18 or the higher local age of majority. It describes an age-gate prompt based on self-affirmation and says documents or additional information may be requested when DONDI deems verification necessary. No named verification vendor, routine identity check or assurance rate was documented, and no additional check was completed in this account.

Direct first-party policy capture on 21 August 2026. The page documents an 18-or-higher-majority rule and begins its self-affirmation safeguard description; it is not evidence of identity-based age verification.

The Terms say automated moderation tools, algorithmic filters and manual review processes may analyse content. Flagged content can be accessed and reviewed by DONDI or authorised partners, removed or reported, and accounts can be restricted or terminated. We did not test a prohibited request, false positive, appeal or human review, so enforcement accuracy remains not tested.

Direct Terms capture on 21 August 2026. It identifies the contracting name and address, states the adult-only threshold and says companions are fictional; enforcement and legal completeness were not independently verified.

Character catalogue and creation

The public homepage and signed-in gallery provide a broad, visual catalogue with printed ages, names and short trait labels. The reviewed profile exposed a longer backstory, a 29-year age label, a voice selector, relationship level, milestones, memories and a Media Studio entry. Those are useful orientation controls.

We did not submit a custom character. Appearance options, personality depth, prompt fields, age enforcement, privacy defaults, save reliability, edit persistence and creation speed are not tested. The character-creation score therefore reflects a navigable catalogue and documented feature map, not a completed build.

Images, video and voice

Images and video

The connected Studio route presented image creation as a Premium feature even while the account still had tokens. The Help Center says an image starts at five tokens and a five-second video animation costs 50 tokens. The account never held enough for that documented video cost, and no subscription was purchased. Prompt adherence, anatomy, identity consistency, censorship, render time, resolution, motion and failure refunds all remain not tested.

This dual gate matters: a token balance alone did not unlock the image workflow in the tested free account. “Unlimited photos and videos” on the Premium pricing card should also be read alongside the separate token economy until a paid-month test establishes exactly which actions remain metered.

Voice playback

Five distinct response cards exposed a Play · 1 token control. Each authorised playback reduced the balance by one token. That verifies the debit and that a playable object existed. The review did not route or assess the audio, so naturalness, pronunciation, emotional range, similarity between clips, latency and accessibility are not tested. It would be wrong to convert a successful debit into a voice-quality finding.

What the free tier provides

The tested account could browse companions and send five structured messages before the interface blocked further chat. It began with 45 tokens, but the source and general availability of that balance were not established; do not assume every new account receives the same amount. The five test messages and five voice plays each accounted for one token in the observed balance sequence.

Character browsing, profile inspection and the short chat were usable. Media creation was not: image generation remained Premium-gated despite tokens, and the Help Center documented a video cost higher than the starting balance. Once the final token was spent, the composer showed an out-of-tokens/subscription gate. The combination of a short daily message ceiling, consumable voice and Premium-plus-token media explains the low Free-tier score.

Tokens, gifts and real action economics

The Help Center says Premium supplies 200 tokens each month and the current yearly purchase adds a 500-token first-month bonus after successful payment. It documents images starting at five tokens and a five-second video at 50 tokens. In this test, a voice play cost one token, the available gift menu allowed five- and 20-token gifts, and the balance fell as expected for each authorised action.

Direct Help Center capture on 21 August 2026. It documents recurring Premium tokens, the annual first-month bonus, an image floor and a five-second video cost; only voice and gift debits were action-tested.

The live token shop listed these standalone packs:

At those pack rates, one token costs roughly 1.6–4.0 cents, a five-token action about $0.08–$0.20, a 20-token gift about $0.32–$0.80, and a 50-token video about $0.80–$2.00. These are calculated comparisons before tax, not checkout quotes. Subscription tokens have a different bundled economics.

The token rules are contradictory. The pricing and token-shop screens said purchased tokens never expire, while the Terms say unused tokens expire at the end of the applicable subscription period. The pack interface described sales as final, while the Terms allow a request within 24 hours only if purchased tokens are unused. Until DONDI reconciles those statements in writing, users should not assume the most favourable interpretation.

Cancellation and refunds

The Terms say subscriptions renew automatically and can be cancelled in account settings, remaining active through the paid period without a partial refund unless law requires one. The checkout copy instead directed cancellation questions to a support email. We did not execute either route, so ease, confirmation speed and renewal prevention are not tested.

Refund wording is also inconsistent. The pricing surface advertised a seven-day full money-back guarantee. The Terms allow a subscription-refund request within 24 hours, conditioned on minimal use and an eligible payment method, and an unused-token request within 24 hours. They exclude refunds for enforcement actions and say mandatory law can prevail. No clear EU withdrawal workflow was found for the France-context account, and no request was filed.

Privacy, discretion and deletion

The Privacy Notice, last revised 8 January 2026, describes collection of identifiers and contact data, IP address, device and browser information, session and activity data, prompts, messages, AI outputs, preferences and metadata. It says technology partners may process content for model training and improvement, safety, quality assurance and service delivery. The reviewed wording did not expose a clear training-specific opt-out.

Direct Privacy Notice capture on 21 August 2026. It identifies submitted content and metadata as personal data and calls DONDI the controller; it does not prove security architecture, deletion completion or a fixed retention limit.

The Notice permits retention while needed for operational, legal and business purposes but gives no fixed period for conversations or generated media. It describes disclosure to service providers and legal recipients. The Terms add the broad content licence described above, and their moderation section permits automated filtering plus manual access to flagged material.

That framework conflicts with the homepage’s categorical claims that chats are zero-log, end-to-end encrypted, never shared and disappear. The live product itself retained chat threads and a 16-entry memory panel. End-to-end encryption would ordinarily exclude the service operator from reading plaintext, yet the policies contemplate processing, training and manual review. We cannot determine the underlying encryption architecture from interface copy, so we treat the homepage wording as unverified and internally inconsistent.

The Privacy Notice describes rights that may include access, correction and deletion, but no deletion request, response time or backend erasure was tested. Users should avoid entering real intimate identifiers, employer data, medical information, financial details, passwords or third-party likenesses.

Reliability and comparison context

The five-message test completed without an observed response error, and all nine voice/gift debits changed the balance in the expected direction. The product also enforced the visible message/token gates. We did not time responses, reproduce the test on another browser, encounter a failed generation, seek token restoration or check service availability over multiple days.

Compared with an AI companion that bundles media into one subscription, DONDI’s Premium-plus-token model demands more cost tracking. Compared with a service that publishes a fixed retention period, training opt-out and a single legal entity, DONDI requires more trust. Its strongest case is the combination of a polished catalogue and competent short-session chat; its weakest is the unresolved gap between privacy marketing and the operative policies.

Use the AI companion rankings to compare this evidence with the other platforms scored on the same fixed-weight grid.

Pros and cons

Primary official sources checked

These current first-party sources were checked on 21 August 2026. Prices, allowances and policies can change.

• DONDI homepage — catalogue, marketing, privacy claims, feature positioning and footer operator wording.
• Terms of Service — Techcommerce Unlimited, fictional-service disclaimer, content licence, moderation, renewal, cancellation, token expiry and 24-hour refund conditions; revised 8 January 2026.
• Privacy Notice — collected data, processing purposes, partners, training/improvement language, retention and rights; revised 8 January 2026.
• Underage Policy — adult threshold, self-affirmation gate and possible additional checks; revised 8 January 2026.
• Help Center: How Tokens Work — monthly tokens, annual bonus, image/video cost and support routes.
• Signed-in Premium, checkout, token-shop, companion, memory and Studio states at app.dondi.ai — current account-specific evidence.

⚠ Some links in this article are affiliate links. If you click and subscribe, AI Companion Atlas may earn a commission — at no extra cost to you, and the commission does not change the evidence labels or verdict. DONDI is an adults-only service that can support sexual, romantic and emotionally intense fictional interactions. AI companions can be wrong, inconsistent and persuasive. Do not share passwords, payment details, government identifiers, workplace secrets, health records or identifiable intimate material. Never upload another person’s face or voice without informed permission, and do not use an AI companion instead of emergency, medical, legal or mental-health support.