Google Play AI App Policy: What's Allowed and What Gets Flagged
Adding a model to your app doesn't create a new rulebook. Google's position is that generative AI apps follow every existing policy — and on top of that, they're responsible for what the model produces, not just what the developer wrote.
That one shift — "you own the output" — is where most AI app trouble starts.
Scope note: SubmitSafe checks your store listing metadata only — title, descriptions, screenshots and listing notes. It does not analyze APK code, SDK behavior or malware. Examples on this page are hypothetical composites, not real apps. Google Play makes all final decisions.
First question: are you even in scope?
Google's AI-Generated Content guidance targets apps where generation is the product: chatbots where the conversation is the core feature, text/voice/image-to-image generators, and apps that create voice or video of real people. It explicitly leaves out, for now, apps that only host AI content, single-purpose summarizers, and productivity apps where AI just improves an existing feature (think suggested email replies).
The common misunderstanding runs both ways. Developers of a note-taking app with an "AI rewrite" button over-worry. Developers of an "AI photo editor" whose headline feature is generating new faces under-worry — because they think of it as editing, while the policy sees image generation.
The two non-negotiables
- Prevent restricted output. The model must not produce content the rest of the policy bans — including content that exploits children and content that enables deceptive behavior. "The user typed the prompt" is not a defense.
- In-app reporting. Users must be able to report or flag offensive AI output without leaving the app, and you're expected to use those reports to improve filtering. A "contact us" email in the settings screen is not the same thing.
What it looks like in practice
A text-to-image app has solid prompt filtering and a clean listing. But the result screen only offers "Save" and "Share". There's no way to report a bad generation from where the user actually sees it. Good moderation, still non-compliant.
"Put anyone's face in any video. No limits, no filters." The app may well have filters — but the listing now advertises their absence. Reviewers read listings as statements of intent, and "no limits" next to a face-swap feature is the sort of phrase that invites a closer look.
Screenshots showing a recognisable celebrity's generated likeness, or a generated voice clip "from" a public figure. Even if the feature is lawful, the marketing implies the app is designed to put words in real people's mouths.
How these usually get flagged
From the patterns we see, the trigger is rarely the model itself. It's the gap between what the listing promises and the safeguards the reviewer can find. Typical red flags in listing text:
- "Uncensored", "unfiltered", "NSFW", "no restrictions" — even as a selling point for creative freedom.
- Feature lists that describe generating real people's likeness or voice with no mention of consent or limits.
- Screenshots whose example outputs are suggestive or borderline, used because they're eye-catching.
- The headline AI feature sitting behind a paywall or login the reviewer can't get past.
The paywall and moderation angles are covered in depth in What can get your GenAI app rejected, and the specific risks of face swap features in AI-powered apps, face swap and deepfakes.
Before you submit
- Describe what the model does, not what it can't be stopped from doing.
- Pick screenshot outputs you'd be comfortable showing a reviewer with no context.
- Make the report/flag control visible on the output screen itself.
- If the app touches real people's faces or voices, say how consent or limits are handled.
SubmitSafe can check the listing text and screenshots for these patterns. It can't verify your in-app moderation or reporting flow — that part is on you.
Run your listing past a second pair of eyes
The free checker flags the listing-text patterns described above before a reviewer sees them.