Why Does My AI Girlfriend Refuse? AI Chat Filters Explained (2026)

The short answer

An AI girlfriend refuses, deflects or deletes a reply because at least one layer of rules between you and the model said no. As published documentation describes them, those layers are the model's own training, the platform's rules, settings and plans, automated classifiers that check what you send and what the AI writes back, separate filters for images, and human moderators.

On several adult platforms, a tame, blurred or refused reply can simply mean your plan doesn't include that content, or you haven't completed a required age check. If text appears and then stops or disappears, a filter may have flagged it while it was still streaming. Microsoft documents streaming output being cut off when a filter fires in Azure OpenAI; whether an app then removes what you already saw is up to the app, and none of the platforms we review documents its own mechanism.

The app maker is not the only one setting the rules: model and moderation providers, Apple and Google, Mastercard, and laws in the UK, California and New York all shape what a companion app may show.

Platforms that allow explicit content still keep hard limits. None of the 27 platforms we review that are still available permits sexual content involving minors, and most also ban sexual content about real people. Trying to get around a filter is itself a rule break on several platforms and can cost you your account.

Jump to: What you see and why · Why a message vanishes · What you can do · Character.AI · Platform table

What you see, and what can cause it

The table pairs common symptoms with causes that public documents describe. Each app decides how it displays these, so the same cause can look different from one app to another.

What each platform allows: summary table

The table reflects each platform's policies as recorded in our reviews on October 24, 2026; check the platform's current policy page before acting on it. We review 28 platforms. AI Girlfriend Chat & Roleplay is left out because its iOS app is no longer on the App Store and our review records no content rules for it, so counts in this guide cover the other 27. Our methodology explains how review facts are sourced, and our NSFW AI glossary entry defines the term.

"Adult-first" means the platform's own pages present explicit content as part of the core product. "Optional" means adult content is allowed but sits behind a setting, level or mode, or is limited in scope. "Allowed" means the rules permit adult content for adults, but the platform doesn't present it as the core product and our reviews record no setting or limit for it. "Not allowed" means the terms or community rules ban sexually explicit content. The gating column lists the age check first, then any setting, then any plan requirement our reviews record.

Character.AI is not in the table because we have not reviewed it; see Character.AI's filter and under-18 changes.

The layers: why a reply gets refused, deflected or deleted

From the outside, refusals from different layers can look the same. Here is what each layer is, as far as public documents describe it.

Layer 1: the model's own training

Some refusals are written by the model itself. Among the platforms we review, only Chai documents this layer; Character.AI, which we haven't reviewed, also does. Chai says its models are designed to refuse sexually explicit content, self-harm content and anything that sexualises minors. Character.AI said on December 12, 2024 that it had been developing a separate model for teens, with more conservative limits on responses, especially romantic content.

A refusal from this layer arrives as ordinary reply text, because the model wrote it. The content-rule pages recorded in our reviews don't say which model each platform runs, so you usually can't tell whether a refusal came from the model or a filter around it. Our explainer on how AI girlfriends work covers the background.

Layer 2: the platform's rules, settings and plans

Each platform writes its own content policy and turns some of it into settings. Secrets AI has spice levels (Mild, Medium, Spicy). Selira AI lets companions be set to censored or uncensored. My Spicy Vanilla runs from Neutral up to an X-Rated level. Romantic AI has an NSFW on/off switch. SweetDream AI's character builder starts at age 21.

Plans and age checks act as gates too. Darlink AI and FantasyGF allow NSFW content on paid plans only, GoLove AI marks NSFW content as only partly available on its free plan, My Spicy Vanilla's X-Rated level is Premium-only, and Romantic AI's Premium unblurs "spicy" replies. Some locks depend on where you are. SpicyChat AI keeps NSFW content locked until a third-party age check is passed in France, the UK, Italy, Australia and 26 US states. Selira AI keeps NSFW features locked until a Didit check is passed where the law requires one. My Spicy Vanilla requires a Yoti check before image and audio features in the UK, France, Germany, Italy, Estonia and Australia.

None of the pages recorded in our reviews explains how a setting is passed to the model, so we don't speculate about the mechanics. The practical point: a reply that stays tame may reflect your settings, your plan or a pending age check, so check those before assuming a filter fired.

Layer 3: classifiers on what you send and what the AI writes

The best-documented layer is the classifier, a separate model that labels content as allowed or not.

• OpenAI's moderation guide (undated; read October 25, 2026) says its models detect harmful content in text and images, either as standalone checks on inputs or as scores returned alongside a generated response. Its "sexual" category excludes sex education and wellness; a separate "sexual/minors" category covers anyone under 18.
• Meta's Llama Guard 3 model card says the model classifies both prompts and responses as safe or unsafe and lists the categories breached. It separates S12 "Sexual Content" (responses containing erotica) from S4 "Child Sexual Exploitation".
• Microsoft's content filtering page for Foundry models (dated August 4, 2026; updated October 19, 2026) says both prompt and completion pass through an ensemble of classifiers covering hate, sexual, violence and self-harm, at four severity levels.

Platforms describe the same pattern. Character.AI's December 2024 post says classifiers check the model's responses and that rule-breaking inputs are blocked. Secrets AI scans inputs in real time. My Spicy Vanilla checks prompts and outputs before and after generation and re-scans them later. OurDream AI screens prompts and outputs automatically. Among the content-rule notes in our reviews, Jupi is the one platform whose terms name a moderation provider: its safety filters include OpenAI's moderation tools.

Layer 4: image filters and upload bans

Images get their own checks. OpenAI announced its omni-moderation-latest model, based on GPT-4o, on September 26, 2024; its current guide lists "sexual" for text and images, "sexual/minors" for text only, and says the model does not classify audio. Kindroid blocks flagged media requests before anything is created. SweetDream AI records blocked photo and video requests on the account, can charge a small token penalty, and suspends accounts automatically after repeated blocks within a day. Nomi advertises unfiltered text chat but bans nude photos on every plan.

Several platforms restrict uploads instead: Nastia AI disables them, OurDream AI does not allow reference photos, My Spicy Vanilla generates every image from text, and Joi AI does not let ordinary users upload media to build characters.

Layer 5: human review and enforcement

On several platforms, automated filters flag content and people review it. Secrets AI sends flagged accounts to human moderators, Candy AI and MyDreamCompanion say staff may review flagged material, Nectar AI combines automated filters with human review, and Kindroid decrypts content for human review when its safety systems flag it. At My Spicy Vanilla, only a person can suspend an account. OurDream AI has people review public submissions and reports, and PolyBuzz says recommended content is screened by AI and human moderators.

Selira AI describes an escalation ladder: rule-based filters and AI analysis block attempts, general offences move from a warning to a final warning to a restriction, and zero-tolerance offences bring an immediate ban. Kindroid warns before locking an account. Enforcement can also act after the fact: PolyBuzz makes characters flagged as a "Violation" visible only to their creator, and Soulkyn AI removes characters that look or behave underage. Appeal routes are covered under what you can do.

Why a reply can stop or vanish mid-stream

None of the companion platforms we review documents how its own streaming and filtering interact, so we can't say what happens inside a particular app. The closest any of them comes is My Spicy Vanilla, which says flagged content is paused rather than shown. The most detailed public explanation we found is Microsoft's documentation for Azure OpenAI, which shows how it can happen.

• Default streaming. Microsoft's content filtering page says streaming continues until a stop token, the length limit, or filtered content is detected, and a filtered generation ends with a "content_filter" finish reason. Its content streaming page (dated September 8, 2026; updated October 15, 2026) says text is buffered and checked before reaching the app, so it arrives in chunks, and that the last segment may be returned before the filter is applied to it.
• Asynchronous filtering. In this optional mode, text is returned without buffering. Microsoft says harmful content can be displayed before the violation signal arrives, and the stream then stops; the signal is guaranteed within about 1,000 characters of the violating content. That is a documented way for text to appear and then be cut off.
• Whole replies and prompts. A filtered non-streamed reply returns no content (in rare cases, a partial result). A filtered prompt returns an HTTP 400 error, which an app may show as a generic error.

How the app presents any of this, as a blank bubble, a canned refusal or a removed message, is up to the app.

Not every interruption is a filter. California's and New York's companion laws require protocols that refer users to crisis services when suicidal ideation or self-harm comes up, and New York requires a "not a human" notice at least every three hours of continuing use. Jupi says messages showing self-harm intent are blocked and the user is pointed to crisis lines.

False positives: what you can legitimately do

Filters can catch things they shouldn't. Microsoft's Sexual category, for example, covers romantic relationships and acts portrayed in erotic or affectionate terms, not only explicit acts, so a classifier built on such a definition can catch non-explicit romantic content. What you can do within the rules:

1. Pick a platform whose written rules fit. Platforms whose terms ban sexual content are built to refuse explicit requests, and rewording one to get past the filter breaks their rules. Use the table above, our NSFW picks, the uncensored chatbot roundup and Compare. To run the whole stack yourself, see our self-hosting guide.
2. Check your settings, plan and age check. Romantic AI's switch, Secrets AI's spice levels, Selira AI's uncensored setting, My Spicy Vanilla's levels and Nectar AI's NSFW mode are the platforms' own controls. SpicyChat AI and Selira AI unlock NSFW features after an age check in certain regions, My Spicy Vanilla gates image and audio features behind a Yoti check in six countries, and all three say they never see your selfie or ID. Read our safety guide on age checks before verifying anywhere.
3. Rephrase to clear up a misunderstanding. If a harmless message, such as a health question or an affectionate line, was misread, say what you mean plainly with context. Rewording a banned request so a filter misses it is the circumvention platforms prohibit.
4. Know what an appeal covers. Our reviews don't record a way to report a single wrongly filtered reply on any of the platforms. The documented appeal routes cover account actions: Kindroid's in-app appeal after a lock, Selira AI's 48-hour appeal review after a restriction, and Character.AI's appeals for creators whose content is moderated. My Spicy Vanilla is the exception: it pauses flagged content instead of showing it and says every decision can be appealed.
5. Leave if the rules don't fit. See our guides to switching platforms and cancelling a subscription, and the safety guide's section on getting a copy of your data and getting it deleted.

Hard limits that apply even when adult content is allowed

Reading the content rules and age policies of the 27 platforms side by side, a few limits recur. Our safety guide covers what gets accounts banned more broadly.

• Minors. No platform we review permits sexual content involving minors: most ban it by name, and the rest ban illegal use or sexual content altogether, or require all characters to be adults. Many look past stated age. OurDream AI treats a character that reads as underage as underage whatever its stated age, Nastia AI bans "aged-up" characters, Secrets AI bans "aging down", and SweetDream AI bans school-age framing. Candy AI, EVA AI, OurDream AI and Secrets AI say they report child sexual abuse material to NCMEC or similar bodies.
• Real people. Most policies ban sexual content or deepfakes of identifiable real people. Kindroid bans intimate images of real people even when AI-generated, and Romantic AI's policy says AI generation is no defence. The entries for Kupid AI, Muah AI, Nomi, SweetDream AI and UDesireAI in our reviews don't list real people as a separate banned category.
• Non-consent and extreme themes. Many adult platforms ban non-consensual scenarios, bestiality, trafficking and incest, with differing lines. Romantic AI bans choking content even when framed as consensual.
• Self-harm. Candy AI, Chai AI, GoLove AI, Joi AI, OurDream AI, PolyBuzz and SpicyChat AI ban self-harm material or its promotion. Jupi blocks messages showing self-harm intent and points users to crisis lines, and Kindroid lists imminent self-harm among its red lines.
• Age checks. Every platform says it is for adults, though Nectar AI's terms still allow under-18s into SFW areas and PolyBuzz's FAQ still mentions a "Teen Mode". Checks range from a click-through gate to selfie or ID checks by Persona, Yoti, Didit, VerifyMyAge, go.cam or Ondato, sometimes only in certain countries or US states.

"Uncensored" never means no rules

Nomi advertises unfiltered text chat, but its own wiki says some content is limited by safety policies. Soulkyn AI has no AI-side filter in private chats, yet removes underage-looking characters. PolyBuzz's FAQ markets "Free, Private, and Unrestricted" chats while its terms ban sexual content with no private-chat exception.

Some platforms' marketing pages also conflict with their own rules; our Darlink AI, Joi AI and FantasyGF reviews record examples. The uncensored chatbot roundup sets out the rules that apply even on "uncensored" platforms. See also our glossary entry on uncensored AI chat and our guide to spotting shady AI companion sites.

Character.AI: its documented filter and the under-18 changes

Character.AI has published a series of posts on its filtering and on changes for under-18 users.

• December 12, 2024: it described a separate, more conservative teen model; classifiers on model responses, strengthened for under-18s; blocking of inputs that break its Terms or Community Guidelines; and, in some cases, a pop-up pointing to the National Suicide Prevention Lifeline.
• October 29, 2025: it announced it would remove open-ended chat for under-18s by November 25, 2025, with daily time limits in the meantime, and age assurance combining an in-house model with tools including Persona.
• Safety Center (article timestamped October 30, 2025, in wording that predates the November 2025 removal of open-ended under-18 chat): the under-18 model has additional, more conservative classifiers, and teens who repeatedly submit violating prompts can be suspended.
• November 21, 2025: it said removal would begin in the US on November 24, 2025, with other markets following.
• September 3, 2026: it said the removal had taken place the previous year, and that creators will be told when content is moderated and why, and can appeal.

The documents we read don't describe a setting that turns filtering off. On October 12, 2026, Character.AI announced a voluntary Chapter 11 process to sell the company, saying the platform remained up and would be wound down if no actionable bid emerged. See our Character.AI alternatives, its directory listing, our guide on what to do when a platform shuts down, and our news page.

Who decides where the line goes

The platform

The platform writes the policy, chooses its filters and enforces both, which is why adult platforms draw different lines. MyDreamCompanion allows step-family roleplay but bans bloodline incest; SpicyChat AI bans all family roleplay, step-relations included. Kindroid's moderation guidelines say it does not want to be a moral or political arbiter of what users say to their AI, yet it keeps red lines on minors, planning harm and imminent self-harm.

Model and moderation providers

A platform built on someone else's model or filter inherits some of that provider's rules. Microsoft says only Azure OpenAI customers approved for modified content filtering have full control and can turn filters off; its table marks switching off completion filters as available only if approved. Its Sexual category, described above, reaches beyond explicit acts, which matters for apps built around romance.

App stores, and why mobile apps are often stricter

Apple's App Store Review Guidelines (last updated October 16, 2026) list overtly sexual or pornographic material as objectionable (1.1.4). Guideline 1.2 explicitly covers AI content creation apps, requiring filtering of objectionable material, reporting with timely responses, user blocking and published contact information, and says apps with user-generated content or services that end up being used primarily for pornographic content may be removed without notice. Guidelines 4.7 and 4.7.1 make developers responsible for chatbots their apps offer and require the same filtering, reporting and blocking.

Google Play's Inappropriate Content policy (undated; read October 25, 2026) bans apps with content intended to be sexually gratifying. Its AI-Generated Content policy requires in-app reporting, and Google's explainer lists generative AI apps primarily intended to be sexually gratifying, and non-consensual deepfake sexual material, as violations, with text chatbot apps in scope. Google announced the reporting requirement on October 25, 2023.

Our review data shows a matching pattern. As of October 24, 2026, 19 of the 27 active platforms we review had no companion app on the official US App Store or Google Play, and all 19 allow adult content in some form. Of the eight with store apps, five ban sexually explicit content. The other three, Kindroid, Nomi and Romantic AI, have store apps and allow adult content under their rules; our reviews don't record different content rules for their app and web versions, and Nomi's nude-photo ban applies on every plan and platform.

Most web-only platforms don't say why they have no store app, so this is a pattern, not a documented cause. Two spell out the link:

• SpicyChat AI's notice of August 27, 2025 says Apple removed its iOS app, finding it did not comply with Apple's guidelines.
• Nomi cites Apple's and Google's app-store rules, along with card-network rules, for its ban on nude photos.

Others split adult and safe-for-work use without saying why. SpicyChat's operator, NextDay AI, has a safe-for-work sister app, PixelChat, on Google Play, and a PixelChat subscription bought there also applies to SpicyChat, which is available on Android only as an APK from its own site. Joi AI runs joi.com (18+) and joi.ai (safe-for-work), both web-only, and Muah AI's guide says app-mode sessions can be restricted to safe-for-work chat.

Payment networks

On April 14, 2021, Mastercard said it was extending its Specialty Merchant Registration requirements to curb illegal adult content. Banks that connect adult-content sellers to its network must certify that the seller has effective controls to monitor, block and, where necessary, take down illegal content. Nomi is the only platform in our reviews that names card-network rules as a reason for a limit.

Laws: the UK, California and New York

United Kingdom. Section 81 of the Online Safety Act, in force since January 17, 2025, requires services that publish their own pornographic content to use highly effective age verification or estimation. A GOV.UK press release of July 24, 2025 said that from July 25, 2025, platforms hosting pornography must run robust age checks, and that these protections would be fully enforceable from that date. The release says Ofcom enforcement action can be 10% of a company's qualifying global annual revenue or £18 million, whichever is greater. Whether a given companion app falls under these duties is a legal question we don't answer here.

The effect shows in our reviews: PolyBuzz says its UK check is mandatory under the Act, SpicyChat AI, OurDream AI, My Spicy Vanilla and Nectar AI name the UK among the places a check applies, and joi.com is unavailable in the UK pending a compliant check. Platforms also run checks elsewhere: SpicyChat AI in France, Italy, Australia and 26 US states, OurDream AI in Australia, and My Spicy Vanilla in France, Germany, Italy, Estonia and Australia. Our reviews don't record which laws are behind those checks.

California. SB 243 was signed on October 13, 2025, and Senator Padilla's office said it becomes law on January 1, 2026. It covers "companion chatbots" that can meet a user's social needs and sustain a relationship across interactions. Operators must give a clear notice that the chatbot is not human where a reasonable person could be misled, maintain and publish a protocol against producing suicidal ideation, suicide or self-harm content that includes referral to crisis services, and disclose that companion chatbots may not be suitable for some minors. For users known to be minors, it adds AI disclosure, a default break reminder at least every three hours, and reasonable measures to stop the chatbot producing visual material of sexually explicit conduct or telling the minor to engage in it.

New York. General Business Law Article 47 took effect on November 5, 2025, according to Governor Hochul's office (release dated November 10, 2025). AI companions need a protocol to detect and address suicidal ideation or self-harm and refer users to crisis services such as 988, plus a notice that the user is not talking to a human at the start of an interaction and at least every three hours. The Attorney General can seek penalties of up to $15,000 per day.

Neither law, in the sections we read, regulates adult content between adults. For adult users, their visible effects are the AI notices and crisis-referral messages, plus, under California's law, a disclosure that companion chatbots may not be suitable for some minors. Our safety guide covers these companion-chatbot laws alongside other risks, and if the reminders prompt a wider rethink, see our guide to healthy use.

Why we don't explain how to bypass a filter

We don't cover jailbreaks or workarounds, because platforms ban them and the user bears the consequences. Candy AI's community guidelines ban attempts to jailbreak its moderation, and Kupid AI's terms bar getting around the app's limits. Secrets AI's prohibited content policy says: "Attempting to generate prohibited content — even if unsuccessful — is itself a violation of this policy." SweetDream AI suspends accounts automatically after repeated blocked photo or video requests within a day.

Penalties at the platforms we review run up to permanent bans, and content involving minors is reported to law enforcement on several of them. A ban can cost money too: Secrets AI's refund policy (updated October 21, 2026) says accounts suspended for violations are not refunded.

Frequently Asked Questions

Why does my AI girlfriend refuse to answer some messages?

The model declined on its own, the platform's rules, your settings or your plan rule the topic out, or a classifier flagged your message or the AI's draft reply. Platforms whose terms ban sexual content, such as Replika, EVA AI, Chai AI, PolyBuzz and Jupi, are built to refuse explicit requests, and rewording one to get past the filter breaks their rules. On adult platforms, check the hard limits first: every platform rules out sexual content involving minors, most also ban sexual content about real people, and many ban non-consensual scenarios.

Why did my AI message disappear halfway through?

No companion platform we review documents its exact mechanism; the closest is My Spicy Vanilla, which says flagged content is paused rather than shown. Microsoft's Azure OpenAI documentation shows how it can happen: in its asynchronous mode, text streams to the screen before the filter finishes, and the stream stops when a violation is found, within about 1,000 characters. What the app then shows you is up to the app.

Does Character.AI have an NSFW filter?

Character.AI documents content filtering rather than an "NSFW filter" as such: classifiers that filter sensitive content from the model's responses, and blocking of rule-breaking inputs. It also described stricter classifiers for under-18s, in wording that predates the removal of open-ended under-18 chat. The documents we read don't describe a way to turn filtering off. Character.AI said it would begin removing open-ended chat for under-18s in the US on November 24, 2025, and in September 2026 said the removal had taken place.

Which AI companion apps allow NSFW content?

Most of the 27 platforms we review allow adult content in some form, but store apps mostly ban it (five of the eight store apps among them do), and every platform keeps hard limits. The summary table above shows each platform's rules as recorded on October 24, 2026. For ranked picks, see our uncensored chatbot roundup and NSFW picks.

Why is the mobile app stricter than the website?

Apple treats overtly sexual or pornographic material as objectionable, and Google Play bans apps intended to be sexually gratifying. Two platforms we review document the link: SpicyChat AI says Apple removed its iOS app in August 2025 for not complying with Apple's guidelines, and Nomi cites Apple's and Google's rules, along with card-network rules, for banning nude photos on every plan. Most web-only platforms don't say why they have no store app. Kindroid, Nomi and Romantic AI have store apps and allow adult content, and our reviews don't record different content rules for their app and web versions.

Can I get banned for trying to jailbreak an AI companion?

Yes. Several platforms ban attempts to get around their filters, and penalties run from warnings and token penalties to permanent bans. SweetDream AI, for example, suspends accounts automatically after repeated blocked photo or video requests within a day, and Kindroid says accounts suspended or terminated for Terms violations aren't entitled to a refund.

Why do I have to verify my age to unlock NSFW content?

Because a law or the platform's own policy requires it where you are. The UK government says platforms hosting pornography have had to run robust age checks since July 25, 2025, and SpicyChat AI locks NSFW content in listed countries and US states until a third-party check is passed. Platforms also run checks in France, Germany, Italy, Estonia, Australia and 26 US states; our reviews don't record which laws are behind them. Before you verify, read our safety guide on what age-check providers keep.

What should I do if the filter blocks something harmless?

Restate the message plainly with context, and check your content settings, your plan and any pending age check. The appeal routes our reviews record mostly cover account actions such as locks and suspensions rather than single replies. If the written rules simply don't fit what you want, switching to a platform whose rules do is the legitimate fix.