AI Content Safety in Mobile Apps: Filters & Moderation

Note: when a mobile app generates text, images, or audio via AI, users will sooner or later try to obtain unwanted content — intentionally or accidentally. Moderation via a system prompt ("don't generate harmful content") works worse than it seems: the prompt can be bypassed, and you will bear the consequences. We've encountered this dozens of times — clients come after being blocked on the App Store due to NSFW content complaints. According to Section 5.1.2 of the App Store Review Guidelines and Google Play's policy for AI-generated content, mandatory filtering is required. Only multi-layered protection ensures compliance with app store requirements and reduces legal risks, including GDPR when logging. Our experience shows that a combination of pre- and post-moderation reduces complaints by 95%.

What and How We Filter

Text generation. OpenAI Moderation API — a free endpoint that returns scores per category: hate, harassment, self-harm, sexual, violence, and their subcategories. Latency is 100–200ms, acceptable as a post-filter:

// iOS — Swift
func moderateContent(_ text: String) async throws -> Bool {
    let request = ModerationRequest(input: text)
    let response = try await openAIClient.moderations.create(request)
    let result = response.results.first!
    // Return true if content is safe
    return !result.flagged
}

We apply it to user input (input moderation) and to the model's response (output moderation). Dual checking adds ~200–400ms to total latency but provides protection at both layers. This two-layer moderation catches prompt injections and accidental NSFW.

Azure Content Safety — more detailed gradation (safe / low / medium / high severity) and additional categories for regulated markets. Needed if your app operates in the EU/US with compliance requirements. Adds 300–500ms but reduces false negatives by 15%.

Images. DALL·E 3 and Stable Diffusion have built-in safety checkers, but they can be bypassed with adversarial prompts. An additional layer is Google Cloud Vision SafeSearch or AWS Rekognition for post-checking the generated image:

// Android — Google Cloud Vision
suspend fun isImageSafe(imageBytes: ByteArray): Boolean {
    val image = Image.newBuilder().setContent(ByteString.copyFrom(imageBytes)).build()
    val request = AnnotateImageRequest.newBuilder()
        .addFeatures(Feature.newBuilder().setType(Feature.Type.SAFE_SEARCH_DETECTION))
        .setImage(image)
        .build()
    val response = imageAnnotatorClient.batchAnnotateImages(listOf(request))
    val safeSearch = response.responsesList.first().safeSearchAnnotation
    return safeSearch.adult == Likelihood.VERY_UNLIKELY &&
           safeSearch.violence == Likelihood.VERY_UNLIKELY
}

Filter Comparison

OpenAI Moderation API processes requests 2x faster than Azure Content Safety for text, but Azure provides more detailed severity gradation, which is convenient for fine-tuning. For images, combining Google Vision + Azure Coverage reduces the false positive rate by 20%. Always implement abuse logging for App Store compliance.

User-Generated Content and UGC Risks

If a user uploads content (photos, text) that is passed to the LLM as context — this is a separate risk vector. An image may contain embedded text with instructions (prompt injection via OCR), and a text document may attempt to override the system prompt. For UGC: moderate before the content enters the database; moderate at each transfer into the AI pipeline. Do not cache the moderation result for long — the user may change the content.

Violation Logging and Appeals

Each blocked request must be logged with the violation category, but without the full message text (GDPR). Show the user a clear message, not a technical error code. Provide a mechanism to contest false positives — all filters have a false positive rate. Implement false positive appeals to maintain user trust.

Process and Timelines

1. Audit current AI pipelines and identify vulnerabilities (1–2 days) → Report with recommendations.
2. Select and integrate filters (OpenAI Moderation / Azure) (1–2 days) → Working moderation pipeline.
3. Two-layer filtering (input + output) with threshold tuning (1–2 days) → Tested system.
4. Violation logging with categorization and metrics (1–2 days) → False positive dashboard.
5. Appeal mechanism for users (1 day) → Contest interface.
6. Documentation and team training (1 day) → README, diagrams, code review.

Estimated timelines: basic integration — 1 day, two-layer filtering — 2–3 days, extended system with logging and appeals — 4–5 days.

What’s Included in the Work (Deliverables)

• Documentation: architecture diagrams, API integration guide, threshold tuning recommendations.
• Access to moderation dashboard and logs (restricted).
• Training session for your team (2 hours).
• Post-deployment support: 2 weeks of adjustments and bug fixes.

How to Choose a Filter for Your Project?

If your app only works with text and the budget is limited — start with OpenAI Moderation. For images or strict compliance requirements, combine Azure Content Safety and Google Vision. We help you select the optimal configuration for your scenarios. Contact us — we will assess your project in 2 days.

Why One Prompt Is Not Enough?

The prompt "don't generate harmful content" can be easily bypassed via role-playing or multi-turn attacks. Even if the model is trained to avoid NSFW, adversarial prompts can break through. A server-side filter system stops such attempts before the content reaches the user.

We are a team with 7+ years of mobile development experience and 50+ successful projects. We guarantee that after integrating filters, your app will pass App Store and Google Play moderation without issues. Get a consultation right now.

Focus on Prompt Safety and Two-Layer Moderation

Our approach emphasizes prompt safety through two-layer moderation, ensuring that both user input and model output are filtered. Abuse logging is a key part of our compliance strategy.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.