AI Companion App Privacy Checklist: What to Check Before Sharing Chats, Photos, or Voice

Privacy pages often begin with reassuring language. Skip quickly to the categories. Replika's policy discusses account information, profile information, messages and content, interests and preferences, payments, device and network data, and usage data. Character.AI lists identifiers, demographics, interests, inferences, financial and commercial information, submitted content such as chats and media, voice data, and support communications.

Nomi's policy is more minimalist in tone, saying the system is designed to know as little personal information as possible and that it does not sell or rent personal information. It still says users provide an account email, name or pseudonym, date of birth, chat and customization content, activity information, and payment information if they upgrade.

Candy AI's privacy notice is especially useful as a checklist because it explicitly describes AI companion messages, prompts, outputs, images, videos, voice notes, payment processors, moderation review, QA, logs, marketing, and third-party AI service providers. A policy that names more categories can feel more alarming, but it may also reveal the real operational surface users should compare.

Some policies say advertising partners do not receive chat content. That is good, but it is not the whole privacy question. Advertising privacy asks whether ad partners or tracking technologies see website behavior, device identifiers, pages viewed, or marketing events. Chat privacy asks whether the platform, model providers, moderators, QA workflows, or support teams can process conversation content.

Replika's policy says advertising partners may collect limited device and interaction information through marketing cookies after consent, but will not have access to conversations or photos submitted through the apps. That distinction matters. It narrows one risk, but users still need to inspect how messages and content are used for operation, safety, improvement, legal compliance, and deletion.

A good comparison page should not collapse these into one vague 'private' label. For companion apps, a better matrix has separate rows for advertising, model processing, human review, payments, logs, deletion, and sensitive-data warnings.

Users often ask one broad question: 'Do they train on my chats?' The useful version is more precise. Is content used to train the main model? Is it used to train safety or moderation systems? Is it de-identified first? Can humans review sampled or flagged chats? Can third-party LLM providers receive message content to generate responses? Can logs retain user IDs, timestamps, or IP addresses?

Character.AI's policy says it uses information to analyze, maintain, improve, modify, customize, and measure services, including to train AI and machine learning models. Candy AI's notice describes training and developing AI models and moderation technologies, human review of de-identified or anonymized interactions for datasets, random querying of content for QA, and human review of flagged or reported content.

These details should shape user behavior. If you do not want a human, vendor, safety workflow, or training process to encounter a detail, do not put that detail in the chat. This is boring advice, which is why it works.

Text roleplay can still be sensitive, but images and voice raise the stakes. A photo may identify your face, home, workplace, or device metadata. Voice can identify you more directly than a fictional persona. A payment record can tie an account to a real card, bank, crypto wallet, billing descriptor, or email address.

Candy AI's notice lists images, videos, voice notes, prompts, outputs, payment processors, support data, log files, and moderation workflows. Character.AI separately names posted images or videos and voice data. That does not mean every media feature should be avoided, but it does mean media-heavy companion products deserve more scrutiny before upload.

The practical rule is simple: keep real identity out of the fantasy. Use fictional personas, avoid real faces, do not share identity documents unless required for age verification and retention is clear, and never include financial, health, workplace, or home-address details in roleplay.

A delete button is not enough. Users should check what deletion covers and what it does not cover. Does it delete chats, character customizations, generated images, memories, personas, support tickets, analytics events, payment records, backups, de-identified datasets, training archives, and legal-retention records? Is there a stated timeline?

Nomi's current privacy policy says account deletion deletes personal information within about 28 days of confirmation, while also naming exceptions for training or communications archives and legal or regulatory retention. Replika says users who shared sensitive information can request deletion as described in its policy or by contacting its privacy email. Candy AI says it retains personal data while the account exists or as necessary for stated purposes, except where law requires otherwise.

The right comparison is not whether a product uses the word deletion. The right comparison is how specific it is, whether the path is visible, and whether the exceptions are understandable before the user shares anything important.

Adult-first companion apps can be legitimate products, but the privacy checklist should be stricter. Users should look for age gating, content-removal policy, moderation rules, refund and cancellation language, payment descriptor details, third-party payment processors, explicit data categories, and whether generated images or videos use tokens or recurring subscriptions.

This is why OnlyKin should not copy adult-first acquisition language just because the keywords have demand. A story-first character app can answer AI girlfriend and companion privacy questions honestly while preserving a broader brand. That is better for trust, app-store fit, creator quality, and long-term SEO than turning every public page into explicit-intent copy.

The growth lesson is not 'avoid the topic.' The lesson is to rank by being more useful. A user searching whether an AI girlfriend app is private needs a checklist, source links, and a clear explanation of trade-offs. That builds more durable trust than a generic promise.

Before trusting any companion or character chat app, run a low-risk first-session test. Create an account with a separate email if possible. Use a nickname. Start with a fictional character and a fictional persona. Ask a few normal questions. Test whether you can find privacy, terms, support, and deletion settings. Check whether the paid flow explains credits, subscriptions, billing descriptors, renewal, and cancellation before you pay.

Then ask whether the product still feels good without real-life disclosure. A strong story app should not require your legal name, workplace, home address, intimate photos, health details, or financial information to create a meaningful scene. If the experience depends on oversharing, that is a product design signal.

OnlyKin's safer product direction is to make fictional creation satisfying: structured cards, private drafts, personas, saved sessions, public discovery, source-backed guides, and clear policies. The less a user has to reveal to enjoy a story, the better the privacy posture feels in practice.