AI Girlfriend Sexting, Read as a Policy Question

Every page ranking for ai girlfriend sexting is selling something. Not one of them tells you what the operator's own terms say about the messages you send.

So that's the question this page answers, and here is the short version. Adults-only, throughout, and the useful distinctions are documentary rather than promotional. Of the six companion platforms whose policies I could load, exactly one publishes how its moderation system works, exactly one publishes numbers from an age-verification programme, and exactly two publish a retention period for message content. One of those two keeps conversation data for up to 60 days after you leave and financial records for a minimum of ten years. Several platforms publish nothing loadable at all. If you're going to type your least guarded material into one of these, the operator's terms and privacy policy are the only documents that describe what happens to it, and most of the category is counting on you not opening them.

Everything below comes from the operators' own published terms, privacy policies and moderation documents, all fetched on 29 July 2026. Where a document would not load, that is stated rather than filled in.

Six Operators, Side By Side

I looked for six things on each platform. Who the legal entity is, whether adult content is addressed in writing, how moderation is described, what the retention position is, whether any age check beyond a checkbox is documented, and whether the documents load at all.

Documents fetched 2026-07-29 from each operator's own domain.

The Anima row is not an oversight. That company publishes fifteen separate legal documents, including an underage policy, and every one of them is rendered client side, so the pages returned no policy text at all to my requests. The privacy policy URL declared on its own Google Play listing gave me nineteen characters, and those characters were the page title. I go through that in more detail in the Anima review.

Moderation Happens To Your Account, Not Your Sentence

Kindroid is the only operator in this set that publishes how its system actually works, and the design is worth understanding because it's the opposite of what most people assume.

Its Moderation Guidelines, last updated 26 May 2026, describe a deliberate refusal to filter at the message level. The document states plainly that no filter is applied to the model itself and that the underlying models remain unchanged. Instead the system targets accounts, using what the company calls a context-aware system, and it names exactly three categories it scans for. Imminent self-harm, imminent harm to others, and child sexual abuse material.

That third one is defined tightly and worth quoting rather than paraphrasing loosely. Minor means under 18, regardless of the user's jurisdiction, and the violation occurs when a character depicted as a minor is placed in a sexual or abusive context. The document works through what does and does not cross that line on both sides. Zero tolerance, in those words.

The scans cover your messages and your input fields, meaning backstory, key memories and avatar descriptions, for context. And there's a limit stated on the scope. Scans are of current context, and historical chats and media are not scanned.

Enforcement is warn-first. The document guarantees a warning before any account lock, which exists because automated detection makes mistakes and the company says so.

Set that against the more common design, a filter sitting between you and the model, refusing or rewriting output in real time. That polices every message from every user. Kindroid's approach lets almost everything through and goes after accounts instead. Both are defensible, they fail differently, and the second one means the moderation system is reading more of your inputs more of the time, which is a trade nobody puts on a pricing page.

Who Reads It, And Under What Conditions

The single most useful sentence I found in this entire research pass is in that same document, and it concerns human eyes.

Kindroid states that all scans are performed by its AI system and that no human reads any content during the automated detection and enforcement process. If your account gets locked and you appeal, the position changes. Proceeding with an appeal gives explicit consent for a trained member of the Trust and Safety team to decrypt and review the specific content that caused the lock. That review is limited to the exact context that triggered it, and the consent covers violations in the preceding two days.

Notice the word decrypt. That's a claim about the storage architecture, not just about staffing, and it's the only such claim I found in the set.

Replika describes something different, and it's in the privacy policy rather than a moderation document. The company says it collects and immediately anonymises user feedback along with small portions of message and content data to train its own safety algorithms, and states that the anonymised data is used only internally and not used to train third-party large language models. It also says that data sent to third-party model providers for generating replies is processed transiently and is not retained by those providers beyond the time needed to produce the response, with carve-outs for service integrity, security and legal compliance.

Nobody else among the six documents any of this as specifically. What the rest publish about your logs, or fail to, sits alongside the store-level disclosures in the companion app store audit.

How Long The Logs Live

Retention is the number that matters most and almost nobody prints. Two operators do.

Replika's privacy policy sets out different clocks for different data. Profile information, messages and content, interests and preferences are processed for up to 60 days after termination of the contract. Account information and financial records, meaning payments, transactions and rewards data, are held for a minimum period of ten years, on stated legal retention grounds. Data collected for usage-trend analysis is anonymised and deleted within one year of collection. Face and head movement data from the TrueDepth API is described as processed on device and immediately discarded, retained for no period of time at all.

Sixty days against ten years, in the same document, for the same account. The conversation goes reasonably fast and the payment record does not. That's normal accounting law rather than anything sinister, but the asymmetry is worth seeing written down, because a lot of people assume deleting the app deletes the relationship with the company.

Nomi's privacy policy, effective 27 April 2026, takes a shorter line. It states that except as otherwise set out in the policy, none of your personal information is retained upon deletion of your account, and that deletion happens within 28 or so days of confirmation. The exception is the one to read twice. Information in the company's training or communications archives is excluded, though the policy says that upon deletion any information in those training archives would no longer be attributable to you.

Kindroid's terms and Candy's terms both address account termination without publishing a retention window for message content. CrushOn's terms of use, last modified 22 July 2025, likewise cover a great deal of ground without one.

The Only Operator Publishing Age-Check Numbers

Here's the part of this research I did not expect to find, and it's the strongest single disclosure in the category.

Kindroid publishes an age verification policy, last updated 3 April 2026, describing a layered system rather than a checkbox. Paid users are treated as verified by the payment on file, whether through the App Store, Google Play or web billing, and stay verified even if the subscription lapses. Free accounts run through a behavioural age estimation system that operates passively and only surfaces when it detects signals consistent with a minor. A flagged user verifies either by subscribing or through Yoti, a third-party provider, using selfie-based age estimation or a government ID. Yoti processes that independently and deletes the images and biometric data afterwards, and Kindroid states it does not receive or store any images or identity documents.

Then it publishes the outcomes. Fewer than 2 percent of new free accounts have triggered the behavioural system since deployment. Of those flagged, approximately 55 percent did not complete verification.

That second figure is the interesting one and the company printed it voluntarily. Roughly half the accounts asked to prove their age walked away instead. You can read that generously, as friction driving off adults who resent the ask, or ungenerously, and the honest answer is that both are happening and nobody outside the company can split them. Either way it's a real measurement of a real programme, published with a number attached, which is more than the rest of this category has ever offered.

Everyone else relies on assertion. Nomi's terms say the company requires users to provide their age and intends to implement additional age-verification measures, reserving the right to do so at any time. That's an intention, honestly labelled. Replika's privacy policy says the services are not intended for under-18s and that discovered minor accounts will be blocked and deleted, which is a response rather than a check. Candy's terms note it may require further information to verify or re-verify that you are an adult, without describing what. CrushOn asks you to affirm you're over 18. Why none of that constitutes verification, and what the store layer does and doesn't add, is the subject of the age gate audit.

Weak age assurance is the fair criticism to make of this category, and it should be made plainly rather than nervously. A product built for adults that cannot tell whether it's talking to one has a design gap, and one company in six has spent money closing it.

The Clause About Real People

One prohibition in Replika's terms is the one most likely to catch an ordinary user by surprise, and I've not seen it flagged anywhere.

It bars creating, instructing or configuring any AI companion to represent, impersonate or simulate a specific real person, living or deceased, using their name, likeness, biographical details or other identifying information, without that person's explicit written consent. The clause names the obvious cases directly, including configuring a companion to roleplay as a current or former romantic partner, a family member, a colleague or a public figure.

Building a companion modelled on your ex is a thing a lot of people would do without thinking about it twice. On this platform it's a terms violation, in writing, and the same document bars soliciting personal information from anyone under 18. Whether any of that is enforced I genuinely don't know, and enforcement isn't described. The clause exists though, and it's the only one of its kind I found.

CrushOn's footer carries something in a similar family, a compliance statement referencing the US record-keeping requirements at 18 U.S.C. 2257. That statute applies to producers of sexually explicit material involving real performers, and an operator publishing one has decided it's in scope. That's a signal about what a platform thinks it is, and it costs nothing to look for.

Questions People Actually Ask

Do any of these apps have no filter? Several advertise the word uncensored. Kindroid is the only one that explains what that means operationally, and even there three hard limits apply and one of them is absolute. Treat the marketing word as marketing.

Are my conversations private? They're private from other users. They're not private from the operator by default, and only one of the six describes encryption in a way that implies its own staff can't casually read content. Read the retention section of the privacy policy before deciding how candid to be.

Can I export my conversations? Not on any of the six, as far as their documents describe. Deletion is offered. Portability isn't, which means every relationship in this category is locked to whoever hosts it.

What happens if the company shuts down? Nothing in these documents promises you anything, and the retention clauses cut both ways. This is the clearest argument for the local-model route, which is compared against hosted platforms in the roleplay chatbot guide.

Why do the ranking pages not cover any of this? Because most of them are published by companies selling a competing product, which is a pattern across the whole category and one that the web-only companion sites review documents in detail.

What I'd Read First

Open the privacy policy and search it for the word retention. Thirty seconds. If there's a number, the company has thought about this. If there's only a paragraph about retaining data as long as necessary, they haven't, or they'd rather not say.

Then look for a moderation document that is separate from the terms of service. One operator in six has one, it's specific, it names its limits, and it publishes what happens on appeal. That's the shape of a company that expects to be read.

Then check whether the age policy describes a mechanism or an intention. There is a large difference between a payment-based check plus a third-party estimation service with published rates, and a sentence saying additional measures may be implemented at any time. Both are honest. Only one is a control. The rest of the pricing and policy picture across the category sits in the main AI girlfriend guide, and the Replika document review takes one operator's paperwork apart line by line.